Computer chassis with magnetic transparent side panel and electromagnetic adsorption control method

By constructing a temperature-magnetic mapping set, and using support vector regression and random forest algorithms to calculate the predicted magnetic force and stress concentration areas, electromagnetic compensation control signals are generated. This solves the problem of uneven magnetic force attenuation in magnetically attached transparent side panels under high-temperature environments, achieving stress balance and improved safety of the glass side panels.

CN123113550APending Publication Date: 2026-11-17DONGGUAN ZHANYU TECH CO LTD
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202611403088.0
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2026-09-10
Publication Date
2026-11-17

AI Technical Summary

Technical Problem

The magnetic force of the magnetic transparent side panel weakens unevenly under high temperature, resulting in uneven stress on the glass side panel, increasing the risk of breakage, and lacking the ability to detect and intervene in advance.

Method used

By acquiring the real-time temperature sequence and initial magnetic force value of each adsorption location inside the device casing, a temperature-magnetic force mapping set is constructed. The nonlinear mapping relationship is processed by the support vector regression algorithm, the real-time magnetic force prediction value is calculated, a global magnetic force attenuation gradient map is generated, the stress concentration region is calculated by the random forest algorithm, and an electromagnetic compensation control signal is generated when the peak stress exceeds the safety threshold to adjust the output current of the electromagnetic coil to achieve balanced force.

Benefits of technology

It significantly improves the stress stability and safety of the glass side panel, avoids the risk of side panel falling off or breaking, and achieves accurate identification and dynamic control of magnetic force changes.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN123113550A_ABST
    Figure CN123113550A_ABST
Patent Text Reader

Abstract

This application provides a computer chassis with a magnetically attached transparent side panel and an electromagnetic adsorption control method, comprising: acquiring real-time temperature sequences of each adsorption position inside the device casing and initial magnetic force values ​​corresponding to each adsorption position, and constructing a temperature-magnetic force mapping set containing spatial coordinate information; if the maximum magnetic force difference value in the global magnetic force attenuation gradient map is greater than a preset difference threshold, then extracting the deformation feature vector of the side panel surface based on the global magnetic force attenuation gradient map; using the deformation feature vector, calculating the stress concentration region inside the side panel using a random forest algorithm, and determining a transparent panel stress distribution matrix containing multi-dimensional stress information; acquiring the transparent panel stress distribution matrix, extracting peak stress data from the transparent panel stress distribution matrix, and determining whether the peak stress data exceeds a preset safety threshold.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This invention relates to the field of information technology, and in particular to a computer chassis with a magnetically attached transparent side panel and an electromagnetic adsorption control method. Background Technology

[0002] To showcase the aesthetic appeal of internal hardware, high-end computer cases are increasingly adopting transparent side panel designs. Magnetic mounting, requiring no screws and offering easy installation and removal, has become the mainstream method for securing side panels in high-end transparent cases. Users can easily remove the side panel with a simple pull when upgrading graphics cards, cleaning dust, or replacing water cooling systems. This convenient experience significantly enhances product prestige and user satisfaction.

[0003] However, the magnetic transparent side panel structures currently on the market have revealed many hidden dangers in actual use. Most products are left unattended after leaving the factory, and neither the user nor the case itself has any way of knowing how much magnetic force is left, whether the side panel is truly attached, or whether the force on the various attachment points is consistent. Once a problem occurs, it is often only noticed after the side panel accidentally falls off or the glass shatters, lacking the ability to detect and intervene beforehand.

[0004] A deeper analysis reveals that the root of the problem lies in the fact that the magnet's performance changes with environment and time. Inside the computer case, due to the prolonged heat generated by the graphics card and processor, the operating temperature often remains between 40 and 60 degrees Celsius. Under such sustained high temperatures, permanent magnets undergo thermal demagnetization, gradually losing their magnetic force. This magnetic attenuation does not occur uniformly; the location of each attraction point varies in distance from the heat source, and there may be slight differences between batches of magnets used. This results in varying rates of magnetic force decline at different attraction points. The side panel, initially under even force, gradually becomes unbalanced, with some points tightly attracted and others loosely attracted. This creates uneven tensile stress within the glass, further increasing the risk of breakage. When the overall magnetic force weakens to a certain critical value, a single bump during transport or an accidental collision is enough to cause the side panel to detach.

[0005] Therefore, how to detect changes in magnetic attraction before it weakens to a dangerous level, and how to promptly pull the magnetic attraction points back to a uniform level when uneven force occurs, so as to prevent users from experiencing side panel detachment or glass breakage without their knowledge, have become key issues in improving the safety and long-term reliability of transparent chassis. Summary of the Invention

[0006] This invention provides an electromagnetic adsorption control method, mainly comprising:

[0007] The system acquires real-time temperature sequences and initial magnetic force values ​​at each adsorption location inside the device casing, constructing a temperature-magnetic force mapping set containing spatial coordinate information. Based on this set, a support vector regression algorithm is used to process the nonlinear mapping relationship between the real-time temperature sequences and the initial magnetic force values, calculating the predicted real-time magnetic force value for each adsorption location at the current moment. The predicted real-time magnetic force value for each adsorption location at the current moment is then acquired, and the magnetic force difference between adjacent adsorption locations is calculated to generate a global magnetic force attenuation gradient map reflecting the overall force state. If the maximum magnetic force difference value in the global magnetic force attenuation gradient map is greater than a preset difference threshold, the deformation feature vector of the side panel surface is extracted based on the global magnetic force attenuation gradient map. Using the deformation feature vector, a random forest algorithm is employed to calculate the stress concentration region inside the side panel, determining a transparent panel stress distribution matrix containing multi-dimensional stress information. The transparent panel stress distribution matrix is ​​obtained, and peak stress data is extracted from it. It is then determined whether the peak stress data exceeds a preset safety threshold. If the peak stress data exceeds the preset safety threshold, an independent electromagnetic compensation control signal is generated for each adsorption position based on the deviation between the real-time magnetic force prediction value and the preset standard magnetic force value. The electromagnetic coil at the corresponding adsorption position is driven according to the independent electromagnetic compensation control signal, and the output current of the electromagnetic coil is adjusted to obtain a balanced target magnetic force state with consistent force at each adsorption position. Furthermore, the step of acquiring the real-time temperature sequence of each adsorption location inside the device casing and the initial magnetic force value corresponding to each adsorption location, and constructing a temperature-magnetic mapping set containing spatial coordinate information, includes: acquiring at least one adsorption point on the wall of the device casing; collecting a first temperature value and a first magnetic force value corresponding to the adsorption point to obtain an initial data set; for the initial data set, determining that if the first temperature value is greater than a preset temperature threshold, then removing the corresponding first magnetic force value to obtain a filtered data set; acquiring a second temperature value and a second magnetic force value from the filtered data set, and processing them using a support vector machine algorithm in conjunction with the coordinate values ​​of the adsorption point to determine a spatial cluster set; extracting the sequence length and acquisition rate corresponding to the second temperature value based on the spatial cluster set to obtain a target sequence set; calculating the correlation degree between the second temperature value and the second magnetic force value using the target sequence set, and determining that if the correlation degree is greater than a preset correlation threshold, then obtaining an initial mapping set; and classifying the initial mapping set using a random forest algorithm to determine the target mapping set.Furthermore, the step of processing the nonlinear mapping relationship between the real-time temperature sequence and the initial magnetic force value using a support vector regression algorithm based on the temperature-magnetic force mapping set, and calculating the real-time magnetic force prediction value of each adsorption position at the current moment, includes: extracting the historical temperature sequence and initial magnetic force value corresponding to each spatial node from the temperature-magnetic force mapping set, and combining them to form a mapping sample; processing the mapping sample using a support vector regression algorithm to analyze the nonlinear characteristics between the historical temperature sequence and the initial magnetic force value, and establishing a regression model; obtaining the real-time temperature sequence of the current time sequence node, performing nonlinear mapping processing on the real-time temperature sequence through the regression model, and calculating the initial magnetic force prediction value of each spatial node; obtaining a preset error compensation coefficient, using the error compensation coefficient to weight the initial magnetic force prediction value, and if the weighted value is greater than a preset magnetic force threshold, then the weighted value is determined as the real-time magnetic force prediction value of each spatial node at the current moment. Furthermore, the step of obtaining the real-time magnetic force prediction value of each adsorption position at the current moment, calculating the magnetic force difference value between adjacent adsorption positions, and generating a global magnetic force attenuation gradient map reflecting the overall force state includes: obtaining the real-time magnetic force prediction value and corresponding spatial coordinates of each adsorption position at the current moment; determining the adjacent adsorption positions corresponding to each adsorption position based on the spatial coordinates; calculating the magnetic force difference value between each adsorption position and its corresponding adjacent adsorption position based on the real-time magnetic force prediction value, and calculating the Euclidean distance between them based on the spatial coordinates; performing a division operation based on the magnetic force difference value and the Euclidean distance to calculate the magnetic force attenuation gradient value corresponding to each adsorption position, and combining the magnetic force attenuation gradient values ​​to generate a global magnetic force attenuation matrix; mapping the global magnetic force attenuation matrix through a preset two-dimensional grid to obtain the attenuation value corresponding to each grid node; if the attenuation value is greater than a preset attenuation threshold, extracting the corresponding grid node, and generating a global magnetic force attenuation gradient map reflecting the overall force state based on the extracted grid node.Further, the step of extracting the deformation feature vector of the side panel surface based on the global magnetic attenuation gradient map includes: obtaining the global magnetic attenuation gradient map; extracting the maximum magnetic difference value contained in the global magnetic attenuation gradient map; determining whether the maximum magnetic difference value is greater than a preset difference limit; if it is greater than the difference limit, then determining the global magnetic attenuation gradient map as the target gradient map; extracting the initial contour line corresponding to the side panel surface using an edge detection algorithm based on the target gradient map; obtaining each pixel point contained in the initial contour line; calculating the curvature value corresponding to each pixel point; determining whether the curvature value is greater than a preset curvature threshold; if it is greater than the curvature threshold, then extracting the corresponding pixel point as a deformation pixel point; calculating the normal vector and deflection angle corresponding to each deformation pixel point based on the deformation pixel point; and performing dimensionality reduction processing on the normal vector and the deflection angle using a principal component analysis algorithm to obtain the deformation feature vector corresponding to the side panel surface. Furthermore, the step of calculating the stress concentration region inside the side panel using the deformation feature vector and employing the random forest algorithm to determine the transparent panel stress distribution matrix containing multi-dimensional stress information includes: obtaining the deformation feature vector; using the decision tree group included in the random forest algorithm to perform node splitting calculation on the deformation feature vector to obtain the Gini coefficient corresponding to each node; determining whether the Gini coefficient is greater than a preset coefficient threshold; if it is greater than the coefficient threshold, extracting the corresponding node as the target node; extracting the leaf node corresponding to the target node by depth-first traversal; determining the stress concentration region based on the region label contained in the leaf node; obtaining the region boundary contained in the stress concentration region; extracting the deformation feature vector corresponding to the region boundary; calculating the stress information based on the deformation feature vector; arranging the stress information using matrix mapping to determine the initial stress matrix; performing tensor decomposition on the initial stress matrix; extracting the multi-dimensional tensor corresponding to the initial stress matrix; expanding the dimensions of the initial stress matrix based on the multi-dimensional tensor to obtain the transparent panel stress distribution matrix containing multi-dimensional stress information.Further, the step of obtaining the transparent panel stress distribution matrix, extracting peak stress data from the transparent panel stress distribution matrix, and determining whether the peak stress data exceeds a preset safety threshold includes: obtaining the transparent panel stress distribution matrix, extracting peak stress from the transparent panel stress distribution matrix, and determining that the peak stress value is an outlier if the peak stress is greater than a preset threshold; obtaining the outlier, extracting the coordinate system corresponding to the outlier in the transparent panel stress distribution matrix, and constructing an initial gradient map based on the coordinate system; obtaining the initial gradient map, performing clustering calculations on the initial gradient map using a density clustering algorithm to obtain multiple connected components; obtaining the multiple connected components, extracting the contour lines surrounding the connected components, and calculating the area value of the enclosed region based on the contour lines; obtaining the area value, and if the area value is greater than a preset area threshold, extracting the connected component corresponding to the area value as a target connected component, and calculating the centroid of the target connected component; obtaining the centroid, calculating the distance between the centroid and a preset center reference point, and determining that the region where the target connected component is located is a risk area if the distance value is less than a preset distance threshold. Furthermore, the step of generating independent electromagnetic compensation control signals for each of the adsorption positions based on the deviation between the real-time magnetic force prediction value and the preset standard magnetic force value includes: acquiring peak stress data of the risk zone; if the peak stress data is greater than a preset safety threshold, acquiring the real-time magnetic force prediction value and the preset standard magnetic force value for each adsorption position on the panel adsorption device; calculating the difference between the real-time magnetic force prediction value and the standard magnetic force value to obtain magnetic force deviation data for each adsorption position; acquiring the magnetic force deviation data, performing a product operation based on the magnetic force deviation data and the coil impedance parameters corresponding to each adsorption position to obtain the compensation current value corresponding to each adsorption position; acquiring the compensation current value, encoding the compensation current value using pulse width modulation, and determining the independent electromagnetic compensation control signal for each adsorption position.

[0008] Computer cases with magnetic transparent side panels, including:

[0009] Chassis body;

[0010] A transparent side panel is detachably mounted on one side of the main chassis body;

[0011] Multiple electromagnetic adsorption components are respectively disposed on the main body of the chassis at each adsorption position corresponding to the transparent side panel, and each electromagnetic adsorption component includes an electromagnetic coil.

[0012] A controller is connected to each of the electromagnetic coils. The controller is used to execute an electromagnetic adsorption control method to adjust the output current of each of the electromagnetic coils so that the force at each adsorption position is balanced.

[0013] The technical solutions provided by the embodiments of the present invention may include the following beneficial effects:

[0014] The electromagnetic adsorption control method of this invention addresses the problem of accurately identifying and dynamically controlling the stress concentration and deformation risks caused by uneven force on the side plates during vacuum adsorption of glass enclosures. It collects real-time temperature sequences and initial magnetic force values ​​at each adsorption location inside the enclosure to construct a temperature-magnetic force mapping set containing spatial coordinate information. A support vector regression algorithm is used to process the nonlinear mapping relationship between temperature and magnetic force, calculating the real-time magnetic force prediction value for each adsorption location at the current moment. Furthermore, by calculating the magnetic force difference between adjacent adsorption locations, a global magnetic force attenuation gradient map is generated. When the maximum magnetic force difference exceeds a threshold, the deformation feature vector on the side plate surface is extracted. A random forest algorithm is then used to calculate the stress concentration region, determining the glass stress distribution matrix. When the peak stress exceeds a safety threshold, an independent electromagnetic compensation control signal is generated based on the deviation between the real-time magnetic force prediction value and the standard magnetic force value. This signal drives the corresponding electromagnetic coil to adjust the output current, achieving a balanced target magnetic force state with consistent force at each adsorption location, thereby significantly improving the stability and safety of the glass side plate under stress. Attached Figure Description

[0015] Figure 1 This is a flowchart of an electromagnetic adsorption control method according to the present invention;

[0016] Figure 2 This is a schematic diagram of the S102 frame in an electromagnetic adsorption control method of the present invention;

[0017] Figure 3 This is a schematic diagram of the S103 frame in an electromagnetic adsorption control method of the present invention;

[0018] Figure 4 This is a schematic diagram of the S107 frame in an electromagnetic adsorption control method of the present invention;

[0019] Figure 5 This is a three-dimensional structural diagram of the computer case with magnetic transparent side panel according to the present invention. Detailed Implementation

[0020] To further understand the content of this invention, a detailed description of the invention is provided in conjunction with the accompanying drawings and embodiments. The specific embodiments described herein are for illustrative purposes only and are not intended to limit the invention. It should also be noted that, for ease of description, only the parts relevant to the invention are shown in the accompanying drawings.

[0021] like Figure 5 As shown, the computer case with a magnetic transparent side panel includes:

[0022] Chassis body 1;

[0023] A transparent side panel 2 is detachably installed on one side of the main body of the chassis 1;

[0024] Multiple electromagnetic adsorption components 3 are respectively disposed on the main body 1 of the chassis at each adsorption position corresponding to the transparent side panel 2, and each electromagnetic adsorption component includes an electromagnetic coil.

[0025] Controller 4 is connected to each of the electromagnetic coils. Controller 4 is used to execute an electromagnetic adsorption control method to adjust the output current of each of the electromagnetic coils so that the force at each adsorption position is balanced.

[0026] like Figures 1-4 As shown, the electromagnetic adsorption control method in this embodiment may specifically include:

[0027] S101. Obtain the real-time temperature sequence of each adsorption position inside the chassis and the initial magnetic force value corresponding to each adsorption position, and construct a temperature-magnetic mapping set containing spatial coordinate information.

[0028] At least one adsorption point is obtained on the chassis wall. A first temperature value and a first magnetic force value are collected at each adsorption point to obtain an initial data set. For this initial data set, if the first temperature value is greater than a preset temperature threshold of 80 degrees Celsius, the corresponding first magnetic force value is discarded, resulting in a filtered data set. The temperature and magnetic force readings retained in the filtered data set are normalized to the 0-1 range using a maximum-minimum normalization method. The normalized temperature is recorded as the second temperature value, and the normalized magnetic force is recorded as the second magnetic force value. Combining the three-dimensional coordinates (x, y, z) of the adsorption point, a five-dimensional feature vector is formed from the second temperature value, the second magnetic force value, and the coordinates. This vector is input into a support vector machine with a radial basis function kernel. Spatial labels are assigned to three pre-divided regions on the chassis wall: top wall, side wall, and bottom wall. Adsorption points with the same label are grouped into the same cluster, resulting in a spatial cluster set. Based on the spatial cluster set, the sequences corresponding to the second temperature values ​​are sorted by timestamp of the adsorption points within each cluster. The sequence length L and sampling rate f are extracted, where L ranges from 128 to 512 sampling points, and f ranges from 10 to 50 Hz, thus obtaining the target sequence set. Using the target sequence set, the correlation degree r between the second temperature value sequence and the second magnetic force value sequence within the same cluster is calculated using the Pearson correlation coefficient. The value of r ranges from -1 to 1. If the absolute value of r is greater than a preset correlation threshold of 0.75, the temperature-magnetic force pair of that cluster is included in the initial mapping set. Using the initial mapping set as training samples and the cluster label as the category marker, a random forest algorithm is used for classification. The forest size is set to 100 decision trees, and the wall region category to which each temperature-magnetic force pair belongs is output, resulting in the target mapping set. This target mapping set is the temperature-magnetic force mapping set used in subsequent step S102.

[0029] The system automatically collects temperature data at each adsorption location using a temperature sensor array deployed inside the chassis, sampling ten times per second. For example, for adsorption nodes with spatial three-dimensional coordinates of 10, 20, and 15, the system acquires real-time temperature sequences of 35.2°C, 35.5°C, 36.1°C, 36.4°C, and 36.8°C within half a second. The system then uses a Kalman filter algorithm to smooth the time series to remove high-frequency thermal noise, calculating the effective reference temperature within that time window to be 36.0°C. Next, the system automatically accesses the underlying hardware configuration database, retrieves and extracts the initial magnetic force calibration value of the electromagnet corresponding to that coordinate location as 45.5 Newtons, and simultaneously reads the magnetic force temperature decay coefficient measured at a reference temperature of 25°C when the electromagnet was manufactured. To establish a dynamic correlation between temperature changes and magnetic force, the system abandons Curie's law for paramagnetic materials and instead uses a linear attenuation compensation model that conforms to the working principle of an energized electromagnet. The physical basis for this model is that the resistance of the copper winding increases with temperature rise, the excitation current decreases under constant voltage drive, and thus the attraction force decreases according to an approximately linear law. The expression is F=F0×[1-β×(T-T0)], where F is the real-time magnetic force, F0 is the initial calibration magnetic force of 45.5 Newtons, T is the effective reference temperature obtained by Kalman filtering, T0 is the factory calibration reference temperature of 25 degrees, and β is the magnetic force temperature attenuation coefficient at this coordinate point. Its value comes from the graded actual measurement of the temperature effect of the winding resistance of the same type of electromagnet on the production line and the comparison fitting with the attraction force. Under the normal working condition of constant voltage excitation, the value range is approximately 0.0020 to 0.0030 per degree. The calibration value of this coordinate point of the chassis is 0.00227 per degree. Substituting T=36.0, T0=25, and β=0.00227 into the model, we obtain a temperature difference of 11 degrees, an attenuation ratio of 2.5%, and the actual magnetic force is updated to 44.36 Newtons. The remaining 127 adsorption points are recalculated point by point according to their respective smoothing temperatures, calibrated magnetic forces, and calibration coefficients to obtain their respective real-time attenuation ratios and magnetic force values.

[0030] S102. Based on the temperature-magnetic mapping set, the support vector regression algorithm is used to process the nonlinear mapping relationship between the real-time temperature sequence and the initial magnetic force value, and to calculate the real-time magnetic force prediction value of each adsorption position at the current moment.

[0031] From the temperature-magnetic mapping set constructed in S101, the smoothed temperature sequence and the initial magnetic force value collected at the same time are extracted for each spatial node, and mapped samples are formed according to the node number to form a regression training sample. Support vector regression is used to process the mapped samples, with the smoothed temperature sequence as the independent variable and the initial magnetic force value as the dependent variable. A radial basis function kernel is used to fit the nonlinear characteristics between the two to establish a regression model. During training, the training set and validation set are divided in a 7:3 ratio. The real-time temperature sequence of the current time-series node is obtained, smoothed using the same moving average method as in S101, and then input into the regression model for nonlinear mapping processing, outputting the initial magnetic force prediction value M0 for each spatial node. The error compensation coefficient k is only used to correct the fitting residual of the regression model itself and does not overlap functionally with the aforementioned magnetic attenuation compensation algorithm. The latter compensates for magnetic attenuation over time, while the former compensates for the deviation between the model output and the true value at the same time. The value of k is calibrated based on the average relative error between the initial magnetic force prediction value and the measured value on the validation set, typically falling between 0.85 and 1.15; in this embodiment, it is set to 0.92. Weight allocation is performed in the form M1 = k * M0, where M1 is the allocated value. The magnetic force threshold Mt is set according to the effective detection lower limit of the sensor, generally between 0.05 and 0.2 millitalas; in this embodiment, it is set to 0.1 millitalas. If M1 is greater than Mt, then M1 is used as the real-time magnetic force prediction value for that spatial node at the current moment; if it is not greater than Mt, it is determined to be a noise interval, and the stable value from the previous moment is used to fill it.

[0032] In one implementation, after acquiring a set of temperature-magnetic mappings containing spatial coordinate information, the system further performs an operation to accurately assess the magnetic state under the current thermal environment.

[0033] The system pre-constructs a temperature-magnetic mapping set to store the correspondence between temperature and magnetic field strength values ​​collected by each monitoring node inside the chassis under different operating conditions. Each record includes a node number, timestamp, temperature reading, magnetic flux density reading, and corresponding operating load label. The sampling period is set to once every 500 milliseconds, and a single node continuously stores no less than 2000 historical records. Specifically, the system extracts data from this temperature-magnetic mapping set as basic input, processes it through normalization, and then feeds it into a machine learning model to analyze the complex physical relationship between temperature distribution and magnetic field changes, thereby achieving dynamic monitoring of the operating status of each node inside the chassis. The "temperature-magnetic mapping set" mentioned below refers to the mapping set defined here; both are the same concept.

[0034] It should be noted that the nonlinear mapping relationship mentioned here refers to the fact that the magnetic force attenuation of the electromagnet inside the chassis does not exhibit a simple direct or inverse proportional trend with temperature changes. In the actual chassis operating environment, the local heat accumulation at the adsorption location, the thermal conductivity of the surrounding metal components, and the resistive heating effect of the electromagnetic coil itself are all intertwined, resulting in a highly nonlinear characteristic in the influence of temperature fluctuations on the initial magnetic force.

[0035] For example, when the temperature rises slowly within a certain range, the magnetic force may only fluctuate slightly, but once the temperature exceeds a certain critical threshold, the magnetic force may experience a precipitous nonlinear decay.

[0036] For example, to accurately capture the aforementioned complex physical changes, the system employs a support vector regression algorithm to process this data. Support vector regression is a regression analysis method based on statistical learning theory. Its core idea is to find an optimal hyperplane that minimizes the total deviation of all sample points from that hyperplane. In the current business scenario, the system first constructs the input feature vector for the algorithm. The system traverses the temperature-magnetic mapping set, and for each specific spatial coordinate, it concatenates the temperature value, temperature change rate, and initial magnetic value at each time point in the corresponding real-time temperature sequence to form a high-dimensional feature input matrix. This matrix not only contains the absolute temperature state at the current moment but also implicitly includes dynamic trend information about heat accumulation. Furthermore, in the processing of the support vector regression algorithm, the selection and application of the kernel function is a key step in resolving nonlinear mapping relationships. Since the input temperature sequence and magnetic changes are difficult to fit with linear functions in the original feature space, the system introduces a radial basis function kernel. This kernel function can implicitly map the original low-dimensional temperature and magnetic feature data to an infinite-dimensional high-dimensional feature space. In this high-dimensional space, the originally nonlinear relationship between temperature fluctuations and magnetic force decay is transformed into a linearly separable problem. The system solves a dual optimization problem to determine a set of support vectors. These support vectors correspond to key adsorption locations inside the chassis that are on the verge of drastic temperature changes or at the critical point of magnetic force decay. By assigning higher weights to these key data points, the algorithm can accurately construct a regression model between the temperature sequence, initial magnetic force, and real-time magnetic force.

[0037] Understandably, based on the trained or configured regression model described above, the system performs inference calculations for each adsorption location within the chassis. The system inputs the feature vector collected and constructed at the current moment into the support vector regression model. The model uses a Gaussian radial basis kernel k=exp-g*‖x-x'‖^2, where x is the current feature vector, x' is the support vector, and g is the kernel width parameter. The final output is f=Σαi*k+b, which represents the real-time absolute magnetic force prediction value at the adsorption location at the current moment, in Newtons, not the change in magnetic force. In an optional embodiment, considering the significant differences in heat dissipation conditions in different areas inside the chassis, the system divides the chassis into three zones—high-temperature zone, transition zone, and low-temperature zone—based on the distance from the adsorption location to the main heat source and the measured temperature gradient. The classification thresholds are set as follows: temperature gradient greater than 5 degrees Celsius per centimeter, between 2 and 5 degrees Celsius per centimeter, and less than 2 degrees Celsius per centimeter, respectively. For high-temperature regions, where sample fluctuations are significant, the penalty parameter C is set to 80 to 120, and the kernel width g to 0.05 to 0.1, to enhance the fitting ability for drastic local changes. In the transition region, C is set to 30 to 60, and g to 0.1 to 0.2. In the low-temperature region, where data is stable, C is set to 5 to 15, and g to 0.2 to 0.4, to suppress overfitting. The system re-acquires the temperature gradient of each zone every 30 seconds and dynamically updates C and g within the aforementioned intervals using linear interpolation. For example, when the measured gradient in a transition region is 3.5 degrees Celsius per centimeter, C is set to 45, and g to 0.15, thus achieving adaptive configuration of the zone parameters.

[0038] For example, the adsorption location near the chassis cooling fan has a high frequency but small amplitude of temperature fluctuations. The system assigns a small penalty parameter to the coordinate nodes in this area to enhance the model's tolerance to high-frequency small fluctuations and prevent overfitting. However, for adsorption locations in chassis dead corners or poorly ventilated areas, the system increases the penalty parameter to ensure that the model has extremely high sensitivity and prediction accuracy to the significant decrease in magnetic force caused by sustained high temperatures.

[0039] Preferably, after outputting the real-time magnetic force prediction value, the system also performs a physical boundary constraint check on the prediction result. The upper boundary is set based on the low-temperature saturation magnetization intensity of the electromagnet core material. Referring to the measured curve of the saturation magnetic flux density of silicon steel sheet at 80 Kelvin, the peak attraction force obtained by multiplying it by the effective cross-sectional area of ​​the electromagnet is taken as the theoretical maximum magnetic force, which is approximately 48 Newtons. The lower boundary is set based on the remanence decay law of the core material near the Curie temperature. The residual magnetic force is measured to be approximately 0.3 Newtons after heating the sample to 420 degrees Celsius in a constant temperature chamber and removing the excitation current. This value is taken as the limiting residual magnetic force threshold. During the chassis assembly stage, the system conducts a calibration test for each batch of electromagnets, collects magnetic force data at both high and low temperatures, corrects the above two boundaries, and writes them into the configuration file. If the real-time magnetic force prediction value calculated by the support vector regression algorithm is lower than 0.3 Newtons, it is truncated to 0.3 Newtons; if it is higher than 48 Newtons, it is truncated to 48 Newtons; the prediction value within the range is kept as the original value output, thereby ensuring that the final output magnetic force data fully conforms to the actual physical characteristics of the hardware inside the chassis.

[0040] Step S103: Obtain the real-time magnetic force prediction value of each adsorption location at the current moment, calculate the magnetic force difference between adjacent adsorption locations, and generate a global magnetic force attenuation gradient map reflecting the overall force state. Map the 8x12 adsorption location array into a two-dimensional grid of nodes, with each node carrying a location number, the real-time magnetic force prediction value, and the average of the last three sampling periods. Adjacent nodes are connected by directed edges, with the edge weight being the difference between the magnetic force of the upstream node and the magnetic force of the downstream node, pointing towards the side with the lower value. When the absolute value of the edge weight exceeds a preset threshold of 0.35 N, the edge is marked as an attenuation edge and assigned light yellow, orange, and red colors for weights between 0.35 and 0.8 N, 0.8 and 1.5 N, and greater than 1.5 N, respectively. The gradient graph is stored in memory in an adjacency list structure. The fields include node coordinates, node magnetic force, edge weight, direction identifier, and color level. At the same time, it is output as a thermally superimposed visual grid through the front-end rendering layer and refreshed every 200 milliseconds. This allows subsequent steps to directly read the decay edge set and color distribution to determine the force state.

[0041] Based on the real-time magnetic field prediction values ​​and their corresponding spatial coordinates at each adsorption location, the Euclidean distance between each adsorption location and the other adsorption locations is calculated using spatial coordinates. Adsorption locations with Euclidean distances less than or equal to the neighborhood radius *r* are included in a candidate set. The top *k* adsorption locations in the candidate set are then retained as adjacent adsorption locations in ascending order of distance. Here, *r* can be 1.2 to 1.5 times the average distance between adjacent adsorption locations, and *k* is generally set to 4 to 8 to balance locality and stability. For each adsorption location and its adjacent adsorption locations, the absolute value of the difference between their magnetic field prediction values ​​is calculated as the magnetic field difference value *D*. Combined with the spatial coordinates, the Euclidean distance *L* between them is calculated. The magnetic field attenuation gradient value in that direction is obtained by dividing *G* by *L*. The gradient values ​​in all adjacent directions of the same adsorption location are then averaged to obtain the magnetic field attenuation gradient value for that adsorption location. The magnetic field attenuation gradient values ​​of all adsorption locations are combined according to their spatial coordinates to form a global magnetic field attenuation matrix. This matrix is ​​then mapped to a preset two-dimensional grid using bilinear interpolation to obtain the attenuation value corresponding to each grid node. The attenuation threshold T is used to filter nodes with significant magnetic attenuation, reflecting the intensity of the magnetic force decrease per unit distance. It serves as the boundary between normal attenuation regions and abnormally rapid attenuation regions, typically ranging from 0.05 to 0.2 mT / mm. This threshold can be dynamically adjusted based on the overall magnetic force level of the adsorption scenario; for example, the 75th percentile of the global attenuation value distribution can be used as T. When the attenuation value of a grid node exceeds T, the node is extracted, and its location and numerical information are preserved. Finally, a global magnetic attenuation gradient map reflecting the overall force state is plotted based on the extracted set of grid nodes.

[0042] In one implementation, after analyzing the nonlinear mapping relationship of each node inside the chassis, the system further performs the operation of acquiring the real-time magnetic force prediction value of each adsorption position at the current moment. This prediction value serves as the basis for subsequent spatial force analysis. The system arranges the real-time magnetic force prediction values ​​of each point distributed at different locations on the chassis according to the spatial coordinate order of the adsorption points, forming a one-dimensional numerical sequence of length equal to the total number of adsorption points, denoted as F_t, where t represents the current sampling time, the i-th component in the sequence corresponds to the real-time magnetic force prediction value of the i-th adsorption point at time t, the value of i ranges from 1 to N, and N is the total number of adsorption points, typically ranging from 16 to 64. This sequence is the real-time magnetic force distribution set at the current moment, distinguishing it from the initial magnetic force observations used as input to the SVR model mentioned earlier. The former is the result of spatial aggregation of the model output, while the latter is the raw signal collected at the model input. The two differ in data source, processing stage, and physical meaning. The system refreshes F_t at the end of each sampling period and sends it to the next stage of the spatial force analysis module.

[0043] Specifically, to deeply analyze the stress balance of the internal structure of the chassis, the system needs to calculate the magnetic force difference between adjacent adsorption positions. It should be noted that adjacent adsorption positions here do not simply refer to absolute physical proximity, but are comprehensively defined based on the topological connections of the internal metal components and stress transmission paths. The system first reads the 3D structural model from the CAD assembly file generated during the chassis design phase. The file format is STEP or IGES. The model records the center point coordinates xi, yi, zi of all electromagnets, as well as the geometric contours and assembly relationships of rigid components such as the chassis frame, reinforcing ribs, and bolted connecting plates. By parsing the assembly tree, the system extracts whether there are shared rigid connecting parts between any two electromagnets and constructs an undirected connected graph G with electromagnets as nodes and rigid connections as edges. Subsequently, the system sets a spatial distance threshold d0 = 150 mm and a topological connectivity determination condition: two adsorption positions are considered adjacent if the number of edges of the shortest path in graph G does not exceed 2, and the Euclidean distance dij = sqrt[(xi-xj)^2 + (yi-yj)^2 + (zi-zj)^2] is less than or equal to d0, where dij represents the straight-line distance between position i and position j. The system traverses each target adsorption position in the initial magnetic state set and filters out other adsorption positions adjacent to it according to the above conditions. For example, two electromagnets located on the same side plate and sharing a reinforcing rib are considered adjacent even if the straight-line distance is 120 mm because the number of edges is 1; conversely, if two electromagnets are located on different side plates and are not directly connected by rigid components, they are not considered adjacent even if the distance is only 100 mm. For each pair of adjacent adsorption sites, the system extracts their respective real-time magnetic force prediction values ​​Fi and Fj. By calculating the absolute difference |Fi-Fj| and dividing it by dij, the change in magnetic force per unit distance, Δij = |Fi-Fj| / dij, is calculated. This change is the magnetic force difference value, which intuitively quantifies the degree of imbalance of electromagnetic adsorption forces in a local area.

[0044] For example, in calculating the magnetic force difference, the system not only focuses on the absolute difference but also introduces the calculation dimension of the relative attenuation ratio. For instance, when the initial design magnetic forces of two adjacent adsorption positions are different, directly comparing the absolute difference may mask the actual force distortion. Therefore, the system divides the absolute difference by the average reference magnetic force of the two positions at standard room temperature to obtain a dimensionless relative difference coefficient r=|F1-F2| / ((B1+B2) / 2), where F1 and F2 are the measured magnetic force values ​​of the two positions under the current operating conditions, and B1 and B2 are the reference magnetic force values ​​at 25 degrees Celsius recorded in the factory calibration file. The file is measured and written into the chassis firmware by the production line after the electromagnets are assembled. The relative difference coefficient and absolute difference are weighted and fused to obtain the comprehensive magnetic force difference value D = 0.4*|F1-F2|+0.6*r*Bavg, where Bavg is the mean of B1 and B2 to restore the dimensions. The weights of 0.4 and 0.6 are the default configurations, which users can adjust according to the model. D is used as the sole input data for the subsequent gradient map construction. Further, based on the massive discrete comprehensive magnetic force difference values ​​obtained above, the system initiates the process of generating a global magnetic force attenuation gradient map reflecting the overall force state. The global magnetic force attenuation gradient map is essentially a two-dimensional or three-dimensional numerical matrix covering the entire adsorption plane inside the chassis. Each element in the matrix represents the degree of magnetic force attenuation within a corresponding small spatial region. To construct this gradient map, the system first divides the effective adsorption area of ​​the chassis into a grid, cutting it into numerous evenly distributed grid nodes, with a default node spacing of 20 mm. In a preferred embodiment, since the actual number of deployed electromagnets is limited and unevenly distributed, the system uses a spatial interpolation algorithm to calculate the magnetic force attenuation gradient values ​​of grid nodes where electromagnets are not directly installed. In practice, the system employs an inverse distance weighted interpolation method. For any given grid node to be calculated, the system searches outwards from that node for the nearest pairs of adjacent adsorption positions corresponding to the known combined magnetic force difference values ​​D. The system assigns different weight coefficients based on the spatial distance between these known position pairs and the node to be calculated. The closer the known data points are, the larger their weight coefficients, and the more significant their impact on the gradient value of that grid node. The system then performs a weighted summation of these known combined magnetic force difference values ​​D and their corresponding weight coefficients to calculate the continuous gradient value of that grid node. By performing this weighted summation calculation on each grid node, the system smoothly extends the originally discrete local difference data at each electromagnet position to the entire chassis space, thereby generating a continuous and complete global magnetic force attenuation gradient map.

[0045] Understandably, the generated global magnetic attenuation gradient map accurately depicts the fluctuations in the stress state inside the chassis through the magnitude and sign of the values ​​in the matrix. Areas with higher values ​​in the matrix indicate a large difference in magnetic attenuation between adjacent electromagnets, meaning that the chassis metal panel in that area is experiencing strong shear stress or torsional torque; while areas with values ​​close to zero indicate that the local magnetic attenuation is uniform and the stress state is relatively stable.

[0046] For example, in the central load-bearing beam area of ​​the chassis, due to the dense concentration of electromagnets and the easy and uniform conduction of heat, the gradient map generated by the system typically exhibits a gentle, low-value distribution in this area. However, in the area near the external cold air inlet at the edge of the chassis, the outer electromagnets maintain good magnetic force due to the influence of the cold air, while the adjacent inner electromagnets experience significant magnetic force attenuation due to their proximity to heat-generating components. At this point, the magnetic force difference calculated by the system is extremely large, which, reflected in the global magnetic force attenuation gradient map, forms a dense gradient band with a sharp increase in value in this edge region, clearly indicating potential structural deformation risk areas. In an optional implementation, after generating the gradient map, the system also applies a two-dimensional filtering algorithm to smooth the matrix data. By setting an appropriate smoothing window size, the system suppresses local abnormal peak values ​​in the gradient map, making the final output global magnetic force attenuation gradient map more consistent with the objective law of continuous stress change in the physical world.

[0047] S104. If the maximum magnetic difference value in the global magnetic attenuation gradient map is greater than the preset difference threshold, then the deformation feature vector of the side plate surface is extracted based on the global magnetic attenuation gradient map.

[0048] A global magnetic attenuation gradient map is obtained. The maximum magnetic difference value contained in the global magnetic attenuation gradient map is extracted, and it is determined whether the maximum magnetic difference value is greater than a preset difference threshold. If it is greater than the difference threshold, the global magnetic attenuation gradient map is determined as the target gradient map, where the difference threshold is set to 45.5. A two-dimensional convolution operation is performed on the target gradient map using a discrete differential operator to obtain the gradient vector field of the side plate surface region. The components corresponding to each pixel in the gradient vector field are denoted as gx and gy, representing the gradient components in the horizontal and vertical directions, respectively. Based on the generalized Hooke's law, the gradient vector field is substituted into the stress-strain relationship to obtain the strain tensor distribution map corresponding to the side plate surface. The strain tensor element at each pixel satisfies eij equal to 0.5*(du / dxj+duj / dxi), where i and j represent coordinate direction indices, and u represents the displacement component. The strain tensor distribution map is thresholded, retaining pixels with strain amplitudes greater than 0.0025 as candidate deformation regions. A connected component labeling algorithm is used to divide the candidate deformation regions into several connected components. For each connected region, its centroid x-coordinate and centroid y-coordinate are extracted. The deflection angle of the principal deformation axis is calculated based on the second-order central moment, and the strain peak value within the connected region is statistically analyzed. The centroid x-coordinate, centroid y-coordinate, deflection angle, and strain peak value are concatenated in a fixed order to generate a one-dimensional floating-point array as the deformation feature vector corresponding to the side plate surface.

[0049] The system scans the constructed global magnetic attenuation gradient map matrix in real time and locates the maximum magnetic difference value in the matrix using an extreme value search algorithm. Assuming the system finds the maximum difference value in the current matrix to be 52.3, it compares this value with the preset difference threshold of 45.5 in the configuration library. Since 52.3 is greater than 45.5, the system determines that there is significant local stress concentration in the current chassis side panel and immediately triggers the deformation feature vector extraction procedure. During the extraction process, the system first calls the Sobel discrete differential operator to perform a two-dimensional convolution operation on the gradient map matrix, calculating the partial derivatives gx and gy of each grid node in the horizontal and vertical directions, respectively, constructing a gradient vector field reflecting the direction of force. Then, a magnetic force-stress conversion factor k = 1.2 * 10^6 Pascals per Tesla per millimeter is introduced, mapping the gradient components to equivalent in-plane stresses σx = k * gx, σy = k * gy, τxy = k * (gx + gy) / 2, converting the dimensions from Tesla per millimeter to Pascals. Substituting the elastic modulus E = 71.0 GPa and Poisson's ratio ν = 0.3 into the side plate material, the strain components are solved using the Hooke's Law matrix form under plane stress: εx = (σx - ν * σy) / E, εy = (σy - ν * σx) / E, γxy = 2 * (1 + ν) * τxy / E, resulting in a two-dimensional strain tensor distribution map. The system calculates the maximum principal strain ε1 = (εx + εy) / 2 + sqrt(((εx - εy) / 2)^2 + (γxy / 2)^2) for each node as a scalar criterion, and uses the material yield strain σs / E as a threshold. In this example, the yield stress is taken as 142 MPa, so the threshold is set to 0.002. An eight-neighbor connected domain labeling algorithm is used to filter out dangerous regions where ε1 exceeds 0.002. For the largest hazardous connected region, weighted second-order central moments μ20, μ02, and μ11 are calculated with ε1 as the weight. Then, the principal axis deflection angle of 35.5 degrees is derived from θ = 0.5*arctan(2*μ11 / (μ20-μ02)). The geometric centroid's x-coordinate is given as 120.5, y-coordinate as 240.0, and the maximum principal strain peak value within the region is given as 0.0034. To eliminate dimensional differences, the system performs dimensionless normalization on the four components: the coordinates are divided by the image side length (800), the angle by 180, and the strain peak value by a threshold of 0.002. These normalized components are then concatenated in a fixed order to generate a one-dimensional floating-point feature vector, which is used as input for the subsequent sideplate health classification model.

[0050] S105. Using the deformation feature vector, the random forest algorithm is used to calculate the stress concentration area inside the side plate and determine the glass stress distribution matrix containing multi-dimensional stress information.

[0051] Deformation feature vectors are obtained, and the decision tree group included in the random forest algorithm is used to perform node splitting calculations on the deformation feature vectors to obtain the Gini coefficients corresponding to each node. It is determined whether the Gini coefficient is less than a preset coefficient threshold, which is 0.15. If it is less than the coefficient threshold, it indicates that the node has high purity and concentrated sample categories, and it is extracted as the target node. The target node is traversed using a depth-first approach to extract the leaf nodes corresponding to the target nodes. The stress concentration region is determined based on the region labels contained in the leaf nodes. The boundary of the stress concentration region is obtained, and the deformation feature vector corresponding to the region boundary is extracted. The triaxial strain components e1, e2, e3 and shear components e4, e5, e6 in the deformation feature vector are taken. Combined with the elastic modulus E and Poisson's ratio v of the glass material, the principal stress sigma_i = E*e_i / (1-v*v) and the shear stress tau_ij = E*e_ij / (2*(1+v)) are calculated according to Hooke's law. This set of force information is used as the force vector. The stress information is arranged using matrix mapping. The mapping rule is to fill in the six stress components at the same boundary sampling point row by row, and arrange the different sampling points in columns in a clockwise order according to the boundary, resulting in an initial stress matrix with 6 rows and the same number of columns as the number of sampling points. Tucker tensor decomposition is applied to the initial stress matrix, decomposing it into a core tensor and factor matrices in three directions. The multidimensional tensor corresponding to the initial stress matrix is ​​extracted. Based on the multidimensional tensor in terms of stress type, spatial location, and time series direction, the initial stress matrix is ​​dimensionally expanded, i.e., a third-order tensor structure is reconstructed using the factor matrices as basis vectors. This structure corresponds one-to-one with the multidimensional stress tuples carried by each grid node in the subsequent grid coordinate system, obtaining a glass stress distribution matrix containing multidimensional stress information.

[0052] In one implementation, after successfully extracting the deformation feature vector of the side plate surface, the system further transmits this one-dimensional floating-point array as input data to a pre-deployed stress analysis engine. The core logic of this engine is to use a random forest algorithm to deeply analyze the physical deformation laws contained in the feature vector, thereby accurately calculating the potential stress concentration areas inside the side plate.

[0053] Tempered glass side panels exhibit high light transmittance and brittle fracture characteristics. A significant nonlinear coupling relationship exists between their internal principal stress and external loads and assembly preload, making it difficult for a single regression model to maintain stable accuracy under all operating conditions. Therefore, the previously extracted deformation feature vector X=[x1,x2,...,x12] is used as the input to a random forest regressor, where x1 to x6 represent the normal displacements of six measurement points on the side panel, x7 to x12 represent the strain gradients at the corresponding measurement points, and the output Y is the predicted equivalent stress at that measurement point. Two hundred regression trees are constructed in the random forest. At each split, four features are randomly selected as candidates, with a minimum leaf node sample size of 5. Out-of-bag sampling is used to estimate the generalization error. Tree expansion stops when the out-of-bag mean square error decreases by less than 0.5% for 10 consecutive rounds. Finally, the arithmetic mean of the predicted values ​​from all trees is used as the output of the side panel stress. The predicted stress is compared with the allowable stress threshold of 45MPa. Measurement points exceeding this threshold are marked as high-risk areas and input into subsequent structural optimization processes.

[0054] Specifically, the random forest algorithm model used in the system is composed of multiple independent classification and regression trees. The historical stress test database comes from long-term monitoring records of in-service glass side panels on the production line and the results of destructive loading tests in the laboratory, with a cumulative sample size of no less than 50,000. Each sample contains three parts: first, a deformation feature vector collected by a high-precision visual measurement system, covering 12 dimensions such as centroid coordinate offset, deflection angle, principal strain peak, and strain gradient distribution; second, load condition parameters recorded simultaneously, such as loading direction, loading rate, and ambient temperature; and third, a stress state label jointly calibrated by stress meters and manual re-inspection, divided into four levels: safe, subcritical, critical, and failure. Each deformation feature vector corresponds one-to-one with the load condition parameters and state label according to the timestamp, forming a structured input-output mapping relationship. In the model initialization phase, the system extracts samples from the above database and constructs multiple different training subsets at a ratio of approximately 63.2% of the total sample size for each subset using a bootstrap sampling method, and assigns a decision tree to each subset for independent training. This integration mechanism can effectively reduce the risk of overfitting that a single decision tree is prone to when dealing with high-dimensional features, and ensure the model's ability to generalize to complex stress states of the glass side plate.

[0055] For example, when a deformation feature vector containing data such as centroid coordinates, deflection angle, and peak strain is input into a random forest model, each element in the vector is passed down through the root nodes of each decision tree. At each internal node, the algorithm randomly selects a portion of the feature vector as candidate splitting attributes and, based on the principle of minimizing Gini impurity or mean squared error, seeks the optimal feature threshold for node splitting.

[0056] For example, a decision tree might prioritize judging based on the deflection angle of the deformation principal axis. If the input deflection angle is greater than the splitting threshold set by that node, the data flows to the left subtree; otherwise, it flows to the right subtree. As the data deepens layer by layer in the tree structure, the feature vector eventually falls into the leaf nodes of each decision tree. Each leaf node is pre-bound to a specific stress concentration region coordinate range or stress intensity level, representing the independent prediction result of that decision tree for the current stress state of the glass side panel. Further, to obtain the final stress concentration region determination result, the system aggregates the outputs of all decision trees in the random forest. For the region localization task, the system uses a majority voting mechanism, statistically analyzing the region boundary coordinates predicted by all decision trees, and selecting the region with the highest overlap and the most votes as the final determined stress concentration region. For the stress intensity regression task, the system calculates the arithmetic mean of the stress prediction values ​​output by all decision trees, using this as the core stress index within the region. Through this multi-model fusion strategy, the system can effectively filter out misjudgments caused by noisy data from individual decision trees. After determining the specific location of the stress concentration area and the core stress index, the system needs to further extract physical parameters containing multi-dimensional stress information.

[0057] It should be noted that the multi-dimensional stress information here not only covers the basic normal stress and tangential shear stress, but also includes the stress birefringence and principal stress direction angles, which reflect changes in the internal microstructure of the glass material. Based on the determined coordinates of the stress concentration region, the system calls the built-in material mechanics transformation function to map the core stress indicators to the aforementioned multi-dimensional stress parameters, thereby comprehensively characterizing the stress state of that region. The conversion function takes the stress tensor components sxx, syy, and txy from the core stress index as input and outputs four types of parameters sequentially according to the following process: First, sxx and syy are projected and superimposed along the outer normal direction of the glass side plate to obtain the normal stress sn, which generally ranges from 0.5MPa to 40MPa; Second, txy is extracted along the tangential direction to obtain the tangential shear stress tt, which generally ranges from 0.2MPa to 25MPa; Third, the principal stress difference s1-s2 between sxx and syy is calculated and linearly amplified according to the photoelastic constant C of the glass material to obtain the stress birefringence dn, which generally ranges from 1e-7 to 5e-6; Fourth, the principal stress direction angle theta is obtained by using the arctangent operation of 2*txy and sxx-syy, which ranges from -90 degrees to 90 degrees. In a preferred embodiment, the system restructures the multi-dimensional stress information obtained above to construct the final glass stress distribution matrix. The system first establishes a two-dimensional discretized grid coordinate system based on the physical dimensions of the chassis glass side panel, with a grid step size of 2mm. Then, the system maps the identified stress concentration regions onto the corresponding grid nodes in this coordinate system. For each grid node within a stress concentration region, the system assigns a multidimensional data tuple containing four parameters—sn, tt, dn, and theta—arranged in a fixed order. For grid nodes outside the stress concentration region, the system assigns them default background stress reference values, where sn and tt default to 0.05MPa, dn defaults to 1e-8, and theta defaults to 0 degrees. By traversing all grid nodes in the coordinate system, the system ultimately generates a high-dimensional tensor structure, namely the glass stress distribution matrix. Based on the above steps, the generated glass stress distribution matrix is ​​serialized into a binary data stream in row-major order and stored in pre-divided fixed-length data blocks in memory, with the matrix's row and column dimensions and a timestamp recorded at the beginning of each block. The subsequent stress monitoring module, alarm judgment module, and finite element verification module directly read the matrix by node coordinates or parameter index through a unified data interface, and incrementally overwrite the tuple data of the corresponding node with each mechanical state update, thereby completing the state refresh.

[0058] S106. Obtain the glass stress distribution matrix, extract the peak stress data from the stress distribution matrix, and determine whether the peak stress data exceeds the preset safety threshold.

[0059] A glass stress distribution matrix is ​​obtained, and peak stress is extracted from it. If the peak stress is greater than a preset critical value of 85.5 MPa, the value corresponding to the peak stress is determined to be an outlier. The row and column indices of the outlier in the stress matrix are used as the horizontal and vertical coordinates. The maximum absolute value of the stress difference between the outlier and its four neighboring stresses (up, down, left, and right) is used as the gradient value of that point. All outlier points are assigned grayscale values ​​according to their gradient values ​​in the same coordinate system to obtain an initial gradient map. The gradient direction points from low stress to high stress, and the gradient value reflects the intensity of local stress jumps. The initial gradient map is clustered using the DBSCAN density clustering algorithm, with a neighborhood radius of 3 pixels and a minimum sample size of 5. The output consists of several connected regions formed by clusters of high gradient points. For each connected region, an 8-neighborhood boundary tracing algorithm is used to extract the outer contour line, and the actual area value is calculated based on the number of pixels enclosed by the contour line. If the area value is greater than a preset area threshold of 20 square millimeters, it is determined to be a target connected region, and the arithmetic mean of all pixel coordinates within that connected region is taken as the centroid coordinates. Using the geometric center of the glass plate as the central reference point, the Euclidean distance between the centroid and the central reference point is calculated. If this distance is less than a preset distance threshold of 30 mm, the region containing the target connected domain is determined to be a risk zone. For example, in a certain detection, a peak stress of 85.6 MPa was extracted, exceeding the critical value of 85.5 MPa and thus classified as an outlier. After clustering, the area of ​​the target connected domain formed was 24.3 square millimeters, and the centroid was 18.7 mm from the central reference point, which is less than the distance threshold. Therefore, this region was marked as a risk zone and an early warning was issued.

[0060] The system reads pre-stored tensor structure data through a direct memory access channel, deserializes it into a two-dimensional array containing stress information of discretized grid nodes, and thus obtains the glass stress distribution matrix. During the parsing of this matrix, the system initiates a global extremum search process, traversing the multidimensional data tuples corresponding to each grid node in the matrix. For the normal stress and tangential shear stress values ​​within the tuples, the system calls the built-in fourth strength theory calculation module to synthesize the multidimensional stress components into a single equivalent stress scalar according to the von Mises yield criterion formula. Subsequently, the system compares the equivalent stress scalars of all nodes one by one in a linear traversal manner, using the variable sigma_max to record the currently traversed maximum value. Each time a new node's equivalent stress sigma_i is read, if sigma_i > sigma_max, then sigma_max = sigma_i is updated. After the traversal, sigma_max is the peak stress data in the stress distribution matrix; for example, the peak equivalent stress extracted at the coordinate node is 85.6 MPa. After obtaining the peak value, the system immediately sends a query request to the configuration database to retrieve the basic tensile strength threshold of 90.0 MPa for the current batch of tempered glass material. To improve the accuracy of the judgment, the system synchronously reads the data from the thermal sensor inside the chassis. Assuming the current ambient temperature is 45.5 degrees Celsius, the system retrieves a preset temperature-attenuation coefficient calibration table from the configuration database. This table records reference points with a step size of 10 degrees Celsius, such as 1.00 for 25 degrees Celsius, 0.98 for 35 degrees Celsius, 0.955 for 45 degrees Celsius, and 0.90 for 55 degrees Celsius. For the non-reference point temperature T of 45.5 degrees Celsius, the system performs linear interpolation between the two adjacent endpoints of 45 degrees Celsius and 55 degrees Celsius, according to the formula k=k1+T-T1 / T2-T1*k2-k1, where T1=45, T2=55, k1=0.955, and k2=0.90. The calculated temperature attenuation coefficient is approximately 0.95. The base threshold is then multiplied by this coefficient to dynamically generate a corrected safety critical value of 85.5 MPa. The system executes a floating-point comparison instruction, determines that the extracted peak stress data of 85.6 MPa is greater than the safety threshold of 85.5 MPa, and then writes the coordinates of the grid node and the over-limit difference into the abnormal status register.

[0061] S107. If the peak stress data exceeds the preset safety threshold, an independent electromagnetic compensation control signal is generated for each adsorption position based on the deviation between the real-time magnetic force prediction value and the preset standard magnetic force value.

[0062] During the stress monitoring of the glass substrate, the real-time acquired strain data is compared with the background reference value. Continuous grid cells with a deviation rate exceeding 15% are merged into risk zones, and the center coordinates and the length and width dimensions of the circumscribed rectangle of these risk zones are used as area identifiers. The default upper limit for the length and width dimensions is 200mm x 200mm. The peak stress data of the risk zones is acquired. If the peak stress data is greater than the preset safety threshold of 8MPa, the real-time predicted magnetic force value and the preset standard magnetic force value of each adsorption position on the glass adsorption device are acquired. The difference between the real-time predicted magnetic force value and the standard magnetic force value is calculated to obtain the magnetic force deviation data for each adsorption position. The magnetic force deviation data is acquired, and a product operation is performed based on the magnetic force deviation data and the coil impedance parameters corresponding to each adsorption position to obtain the compensation current value corresponding to each adsorption position. The calculation relationship is I = ΔF * Z, where I is the compensation current value, ΔF is the magnetic force deviation data, and Z is the coil impedance parameter. The compensation current value is obtained, and the compensation current value is encoded using pulse width modulation to determine an independent electromagnetic compensation control signal for each of the adsorption positions.

[0063] In one possible implementation, when the system determines that the extracted peak stress data is greater than the safety threshold, it indicates that the tempered glass of the current chassis is subjected to excessive local stress during the fixing process, and there is a potential risk of breakage. At this time, the system will immediately trigger the dynamic adjustment mechanism of electromagnetic adsorption force.

[0064] Specifically, the system first needs to obtain the real-time predicted magnetic force value and the preset standard magnetic force value under the current state. The preset standard magnetic force value refers to the basic adsorption force value that each electromagnetic adsorption node needs to maintain under ideal assembly conditions to ensure that the tempered glass is stably attached to the chassis frame without generating destructive stress. This value is solved by the system in the initialization phase based on the static model, node by node. The specific process is as follows: first, read the mass M of the glass, the length and width dimensions L*W, the chassis mounting tilt angle θ, and the coordinate distribution of the adsorption nodes on the frame. Then, based on the torque balance principle, the component M*g*cosθ of the total weight of the glass along the normal direction of the mounting surface is weighted and distributed inversely proportional to the distance from each node to the center of gravity to obtain the static bearing capacity Fi of each node. Then, a safety redundancy coefficient k is superimposed, with a value ranging from 1.2 to 1.5, and finally the standard magnetic force value Fs=k*Fi of the node is obtained and written to the control register. Taking a symmetrical arrangement of four corner points as an example, the standard magnetic force value of a single node generally falls in the range of 15N to 40N. The real-time magnetic force prediction value is not directly read from the force sensor, but is calculated in real time by the system calling the built-in magnetic force prediction model. The system simultaneously collects the real-time operating current I and terminal voltage U of the electromagnetic coil corresponding to each adsorption position, as well as the microscopic air gap thickness δ between the glass and electromagnet surfaces obtained by the laser ranging module, with a sampling period of 5ms. These multi-dimensional state data are used as input feature vectors and fed into a pre-trained hybrid prediction model of magnetoresistive network and BP neural network. The model's magnetoresistive network layer, based on the law of electromagnetic induction, treats the iron core, air gap, and glass substrate as equivalent series-parallel magnetoresistive units, outputting a preliminary magnetic flux estimate. The BP neural network layer corrects for nonlinear deviations such as magnetic saturation, leakage flux, and temperature drift. The network structure is a three-layer feedforward network with 12 hidden layer nodes, using ReLU as the activation function. The output layer is a linear function. Training samples are derived from over 10,000 sets of force calibration data measured on an assembly bench under different combinations of current, air gap, and tilt angle. Mean square error is used as the loss function, and the model converges iteratively using the Adam optimizer. The final model outputs the real-time magnetic force prediction value Fp for each node. Further, after acquiring the above two sets of values, the system initiates a deviation analysis module. For each independent adsorption position on the chassis frame, the system subtracts the corresponding real-time magnetic force prediction value Fp from the standard magnetic force value Fs to obtain the magnetic force deviation ΔF = Fp - Fs for that node. This deviation not only includes its numerical value but also its direction. A positive value indicates that the current actual adsorption force is too large, which is the main reason for the local stress exceeding the limit. A negative value indicates that the adsorption force is insufficient. The system integrates these deviations into a multi-dimensional deviation vector, which serves as the basis for generating compensation commands.

[0065] It should be noted that generating independent electromagnetic compensation control signals for each adsorption location is the core of the entire adjustment process. The system incorporates a multi-channel independent pulse width modulation controller, which can decouple the control of multiple electromagnetic adsorption nodes on the chassis. The system inputs the aforementioned calculated deviation vector into the adaptive proportional-integral-derivative (PID) controller. The adaptive mechanism of this regulator is reflected in real-time monitoring of the current deviation amplitude *e* and the deviation change rate *de / dt*, and online tuning of three parameters accordingly: when *e* is greater than 60% of the rated value, a large proportional coefficient *Kp* and a small integral coefficient *Ki* are used to accelerate the response, where *Kp* is linearly interpolated between 2.5 and 4.0, and *Ki* is placed between 0.05 and 0.15 to suppress integral saturation; when *e* is in the 20% to 60% range, *Kp* drops to 1.2 to 2.5, and *Ki* is increased to 0.15 to 0.35 to eliminate steady-state deviation; when *e* is less than 20%, *Kp* drops to 0.5 to 1.2, and the differential coefficient *Kd* is adjusted according to the *de / dt* amplitude between 0.02 and 0.08 to suppress overshoot and oscillation. The tuned output calculates the amount of magnetic compensation to be reduced for adsorption positions with positive deviations and maps it to a duty cycle reduction command for the control signal. The system uses a lookup table to access a pre-defined mapping table of magnetic force and duty cycle to determine the specific duty cycle adjustment step. This mapping table is based on offline calibration tests on the production line: under multiple operating conditions with air gaps fixed at 0.5mm to 2.0mm, the duty cycle is swept from 10% to 90% in 10% increments, while simultaneously collecting data from the coil current sensor and force gauge to obtain discrete sample points of the adsorption force F and duty cycle D. A two-dimensional lookup table is then formed through cubic spline interpolation, with the row index representing the air gap value and the column index representing the target magnetic force value. The elements in the table represent the corresponding duty cycle values. A typical mapping relationship is as follows: an air gap of 1.0mm and a target magnetic force of 80N correspond to a duty cycle of approximately 55%. Subsequently, the system generates an independent electromagnetic compensation control signal containing a specific frequency and the adjusted duty cycle, which is directly sent to the power drive circuit corresponding to that location. This reduces the average excitation current of the coil to weaken the local adsorption force, thereby releasing the concentrated stress in that area.

[0066] For example, suppose the standard magnetic force at the adsorption position in the upper left corner of the chassis is 45.0 N, while the real-time predicted magnetic force is 52.5 N, resulting in a positive deviation of 7.5 N. This node has undergone linear mapping calibration of magnetic force and duty cycle before leaving the factory, with a calibration coefficient K of approximately 0.94 N per percentage point. This means that a 1% change in duty cycle corresponds to a change in adsorption force of approximately 0.94 N, and the mapping range covers a duty cycle of 30% to 85%. After receiving this deviation data, the regulator uses an incremental PID algorithm, where the proportional coefficient Kp is 0.8, the integral coefficient Ki is 0.15, and the derivative coefficient Kd is 0.05. The sampling period is 10 ms, the input is the 7.5 N positive deviation, and the output is the duty cycle adjustment Δd. Calculations show that Δd is approximately -8%, therefore the duty cycle of the node's electromagnetic coil pulse width modulation signal is reduced from 60% to 52%, falling within the calibration range. The system then generates an independent electromagnetic compensation control signal for the top left node and drives the solid-state relay to perform an action, precisely reducing the adsorption force at that point.

[0067] In one embodiment, to prevent stress redistribution in other nodes when the adsorption force is significantly adjusted at a certain node, the system introduces a global torque balance constraint when generating independent electromagnetic compensation control signals. The system constructs a set of force balance equations for the tempered glass chassis based on the spatial coordinates xi,yi of each adsorption position and the current adsorption force Fi at that node, while simultaneously constraining the resultant force ΣFi=0, the torque around the x-axis Σyi*Fi=0, and the torque around the y-axis Σxi*Fi=0, where i represents the i-th adsorption node. When the compensation amount at a certain main adjustment point k is ΔFk, the system distributes -ΔFk to the remaining nodes according to the principle of diagonal priority followed by adjacent nodes. Diagonal nodes bear approximately 50% to 60%, two adjacent nodes each bear approximately 20% to 25%, and the remaining nodes bear less than 5%. The distributed amount is then converted into the duty cycle fine-tuning value Δdi for each node using the calibration curve of electromagnetic force and duty cycle. Taking a four-point adsorption layout as an example, if the duty cycle of the main adjustment point is increased from 45% to 60%, i.e. Δdk=15%, then the duty cycle of the diagonal nodes is simultaneously reduced by 8% to 9%, and the duty cycles of the two adjacent nodes are each reduced by about 3% to 4%. This ensures that the resultant force and torque residual of the glass converge to within the set values ​​of 0.05N and 0.02N*m after each adjustment, ensuring that while eliminating local peak stress, the overall stress state of the glass remains stable.

[0068] S108. Drive the electromagnetic coil at the corresponding adsorption position according to the independent electromagnetic compensation control signal, adjust the output current of the electromagnetic coil, and obtain a balanced target magnetic force state with consistent force at each adsorption position.

[0069] Independent electromagnetic compensation control signals are acquired, and the target current value corresponding to each adsorption site is analyzed. The target current value for each adsorption site ranges from 500 mA to 1200 mA; in this embodiment, 850 mA is used. The actual output current is collected by current sensors installed on the power supply circuit of the electromagnetic coil at each adsorption site. The collected data is then fed into a discrete Kalman filter algorithm in the embedded control unit for state estimation. After filtering out sampling noise, the smoothed output current corresponding to each adsorption site is obtained. The difference between the smoothed output current and the target current value is calculated to obtain the current deviation e. If the absolute value of e is greater than a preset deviation threshold of 20 mA, e is fed into a proportional-integral-differential algorithm to output the driving voltage value corresponding to each adsorption site, with a driving voltage range of 0 V to 24 V. If the absolute value of e does not exceed 20 mA, the current driving voltage value is maintained unchanged, and no further adjustment is made. The electromagnetic coil corresponding to each adsorption site is energized according to the driving voltage value to obtain the real-time magnetic field B corresponding to each adsorption site. The force application points are the contact detection positions of the electromagnetic coils at each adsorption site, directly facing the surface of the workpiece. Each adsorption site corresponds to a specific adsorption site. A Hall sensor is positioned at this location to collect the B value, and the force is calculated using the formula F = k * B * B * S / u, where F is the force value, S is the effective magnetic pole area of ​​the electromagnetic coil, u is the air permeability, and k is a calibrated structural correction coefficient. If F reaches the preset force threshold of 50 N, the adsorption site is determined to have reached the equilibrium target magnetic state.

[0070] Upon receiving an independent electromagnetic compensation control signal, the system's underlying power distribution module immediately activates a hardware-level response mechanism, converting a pulse-width modulation waveform with a specific duty cycle into a control voltage that drives the gate of the MOSFET. Taking the adsorption node numbered 4 in the lower right corner of the chassis as an example, the duty cycle is reduced from the initial 60% to 52%. Based on the linear mapping relationship between the excitation current and the duty cycle within the operating range of this drive circuit, the target value of the excitation current is simultaneously reduced from 1200 mA to 850 mA. Consequently, the conduction time of the switching transistor in the power drive circuit is shortened, and the transient excitation current flowing through the electromagnetic coil of this node gradually decreases from the initial 1200 mA. To avoid electromagnetic shocks caused by sudden current changes, the system embeds a discrete Kalman filter algorithm in the current regulation loop to filter out noise and estimate the state of the real-time current data collected by the built-in Hall current sensor at a frequency of 1000 times per second. When the absolute value of the difference between the filtered actual current value and the target current value of 850 mA set by the compensation signal is less than 15 mA for 10 consecutive sampling periods, the system determines that the coil current is adjusted to the appropriate level. During the synchronous dynamic adjustment of the output current of the electromagnetic coils at each node, the system continuously calls the thin-film pressure sensor array distributed around the chassis frame to collect real-time force data at each adsorption position. The adsorption nodes at the bottom of the chassis are evenly arranged in 4 rows and 4 columns, totaling 16 force acquisition points. The central processing unit substitutes the force values ​​of these 16 nodes into the overall standard deviation formula σ=sqrt(Σ(Fi-Favg)² / N), where Fi is the force at the i-th node, Favg is the arithmetic mean of the forces at the 16 nodes, and N is taken as 16, to evaluate the dispersion of the force distribution in real time. When σ smoothly decreases from the initial 18 Newtons and stabilizes within the preset convergence threshold of 3 Newtons, and the force range Fmax-Fmin at each node does not exceed 5 Newtons, the system determines that the internal stress of the tempered glass of the chassis has been completely released and has successfully entered the equilibrium target magnetic state with consistent forces at each adsorption position.

[0071] The above are only some preferred embodiments of the present invention, but the present invention is not limited thereto, and many improvements and modifications can be made. Any improvements and modifications made based on the basic principles of the present invention should be considered to fall within the protection scope of the present invention.

Claims

1. An electromagnetic adsorption control method, characterized in that, include: The real-time temperature sequence of each adsorption location inside the device casing and the initial magnetic force value corresponding to each adsorption location are obtained, and a temperature-magnetic force mapping set containing spatial coordinate information is constructed. Based on the temperature-magnetic mapping set, a support vector regression algorithm is used to process the nonlinear mapping relationship between the real-time temperature sequence and the initial magnetic force value, and to calculate the real-time magnetic force prediction value of each adsorption position at the current moment; the real-time magnetic force prediction value of each adsorption position at the current moment is obtained, the magnetic force difference value between adjacent adsorption positions is calculated, and a global magnetic force attenuation gradient map reflecting the overall stress state is generated; if the maximum magnetic force difference value in the global magnetic force attenuation gradient map is greater than a preset difference threshold, the deformation feature vector of the side panel surface is extracted based on the global magnetic force attenuation gradient map; through the deformation feature vector, a random forest algorithm is used to calculate the stress concentration region inside the side panel, and a transparent panel stress distribution matrix containing multi-dimensional stress information is determined; the transparent panel stress distribution matrix is ​​obtained, the peak stress data in the transparent panel stress distribution matrix is ​​extracted, and it is determined whether the peak stress data exceeds a preset safety threshold. If the peak stress data exceeds the preset safety threshold, an independent electromagnetic compensation control signal is generated for each of the adsorption positions based on the deviation between the real-time magnetic force prediction value and the preset standard magnetic force value. The electromagnetic coil at the corresponding adsorption position is driven according to the independent electromagnetic compensation control signal, and the output current of the electromagnetic coil is adjusted to obtain a balanced target magnetic force state with consistent force at each adsorption position.

2. The method as described in claim 1, characterized in that, The process of acquiring real-time temperature sequences at each adsorption location inside the device casing and initial magnetic force values ​​corresponding to each adsorption location, and constructing a temperature-magnetic mapping set containing spatial coordinate information, includes: acquiring at least one adsorption point on the device casing wall, collecting a first temperature value and a first magnetic force value corresponding to the adsorption point to obtain an initial data set; for the initial data set, determining if the first temperature value is greater than a preset temperature threshold, then removing the corresponding first magnetic force value to obtain a filtered data set; acquiring a second temperature value and a second magnetic force value from the filtered data set, and processing them using a support vector machine algorithm in conjunction with the coordinate values ​​of the adsorption point to determine a spatial cluster set; based on the spatial cluster set, extracting the sequence length and acquisition rate corresponding to the second temperature value to obtain a target sequence set; calculating the correlation degree between the second temperature value and the second magnetic force value using the target sequence set, and determining if the correlation degree is greater than a preset correlation threshold to obtain an initial mapping set; and using a random forest algorithm to classify the initial mapping set to determine the target mapping set.

3. The method as described in claim 1, characterized in that, The step of processing the nonlinear mapping relationship between the real-time temperature sequence and the initial magnetic force value using a support vector regression algorithm based on the temperature-magnetic force mapping set, and calculating the real-time magnetic force prediction value of each adsorption location at the current moment, includes: extracting the historical temperature sequence and initial magnetic force value corresponding to each spatial node from the temperature-magnetic force mapping set, and combining them to form a mapping sample; processing the mapping sample using a support vector regression algorithm, analyzing the nonlinear characteristics between the historical temperature sequence and the initial magnetic force value, and establishing a regression model; obtaining the real-time temperature sequence of the current time sequence node, performing nonlinear mapping processing on the real-time temperature sequence through the regression model, and calculating the initial magnetic force prediction value of each spatial node; obtaining a preset error compensation coefficient, using the error compensation coefficient to weight the initial magnetic force prediction value, and if the weighted value is greater than a preset magnetic force threshold, then the weighted value is determined as the real-time magnetic force prediction value of each spatial node at the current moment.

4. The method according to any one of claims 1-3, characterized in that, The step of obtaining the real-time magnetic force prediction value of each adsorption position at the current moment, calculating the magnetic force difference value between adjacent adsorption positions, and generating a global magnetic force attenuation gradient map reflecting the overall force state includes: obtaining the real-time magnetic force prediction value and corresponding spatial coordinates of each adsorption position at the current moment, and determining the adjacent adsorption positions corresponding to each adsorption position based on the spatial coordinates; calculating the magnetic force difference value between each adsorption position and its corresponding adjacent adsorption position based on the real-time magnetic force prediction value, and calculating the Euclidean distance between them based on the spatial coordinates; performing a division operation based on the magnetic force difference value and the Euclidean distance to calculate the magnetic force attenuation gradient value corresponding to each adsorption position, and combining the magnetic force attenuation gradient values ​​to generate a global magnetic force attenuation matrix; mapping the global magnetic force attenuation matrix through a preset two-dimensional grid to obtain the attenuation value corresponding to each grid node, and if the attenuation value is greater than a preset attenuation threshold, extracting the corresponding grid node, and generating a global magnetic force attenuation gradient map reflecting the overall force state based on the extracted grid node.

5. The method according to any one of claims 1-3, characterized in that, The step of extracting the deformation feature vector of the side panel surface based on the global magnetic attenuation gradient map includes: obtaining a global magnetic attenuation gradient map, extracting the maximum magnetic difference value contained in the global magnetic attenuation gradient map, determining whether the maximum magnetic difference value is greater than a preset difference limit, and if it is greater than the difference limit, determining the global magnetic attenuation gradient map as the target gradient map; extracting the initial contour line corresponding to the side panel surface using an edge detection algorithm based on the target gradient map; obtaining each pixel point contained in the initial contour line and calculating the curvature value corresponding to each pixel point; determining whether the curvature value is greater than a preset curvature threshold, and if it is greater than the curvature threshold, extracting the corresponding pixel point as a deformation pixel point; calculating the normal vector and deflection angle corresponding to each deformation pixel point based on the deformation pixel point; and performing dimensionality reduction processing on the normal vector and the deflection angle using a principal component analysis algorithm to obtain the deformation feature vector corresponding to the side panel surface.

6. The method according to any one of claims 1-3, characterized in that, The step of calculating the stress concentration region inside the side panel using the deformation feature vector and employing a random forest algorithm to determine the stress distribution matrix of the transparent panel containing multi-dimensional stress information includes: obtaining the deformation feature vector, performing node splitting calculation on the deformation feature vector using a decision tree group included in the random forest algorithm to obtain the Gini coefficient corresponding to each node; determining whether the Gini coefficient is greater than a preset coefficient threshold, and if it is greater than the coefficient threshold, extracting the corresponding node as a target node, extracting the leaf node corresponding to the target node by depth-first traversal, and determining the stress concentration region based on the region label contained in the leaf node; obtaining the region boundary contained in the stress concentration region, extracting the deformation feature vector corresponding to the region boundary, calculating the stress information based on the deformation feature vector, arranging the stress information using matrix mapping to determine the initial stress matrix; performing tensor decomposition on the initial stress matrix, extracting the multi-dimensional tensor corresponding to the initial stress matrix, and expanding the dimensions of the initial stress matrix based on the multi-dimensional tensor to obtain the stress distribution matrix of the transparent panel containing multi-dimensional stress information.

7. The method according to any one of claims 1-3, characterized in that, The process of obtaining the stress distribution matrix of the transparent panel, extracting peak stress data from the stress distribution matrix, and determining whether the peak stress data exceeds a preset safety threshold includes: obtaining the stress distribution matrix of the transparent panel, extracting peak stress from the stress distribution matrix, and determining that the peak stress value is an outlier if the peak stress is greater than a preset threshold; obtaining the outlier, extracting the coordinate system corresponding to the outlier in the stress distribution matrix of the transparent panel, and constructing an initial gradient map based on the coordinate system; obtaining the initial gradient map, performing clustering calculations on the initial gradient map using a density clustering algorithm to obtain multiple connected components; obtaining the multiple connected components, extracting the contour lines of the connected components, and calculating the area value of the enclosed region based on the contour lines; obtaining the area value, and if the area value is greater than a preset area threshold, extracting the connected component corresponding to the area value as a target connected component, and calculating the centroid of the target connected component; obtaining the centroid, calculating the distance between the centroid and a preset center reference point, and determining that the region where the target connected component is located is a risk area if the distance is less than a preset distance threshold.

8. The method according to any one of claims 1-3, characterized in that, The step of generating independent electromagnetic compensation control signals for each adsorption position based on the deviation between the real-time magnetic force prediction value and the preset standard magnetic force value includes: acquiring peak stress data of the risk zone; if the peak stress data is greater than a preset safety threshold, acquiring the real-time magnetic force prediction value and the preset standard magnetic force value for each adsorption position on the panel adsorption device; calculating the difference between the real-time magnetic force prediction value and the standard magnetic force value to obtain magnetic force deviation data for each adsorption position; acquiring the magnetic force deviation data and performing a product operation based on the magnetic force deviation data and the coil impedance parameters corresponding to each adsorption position to obtain the compensation current value corresponding to each adsorption position; acquiring the compensation current value and encoding the compensation current value using pulse width modulation to determine the independent electromagnetic compensation control signal for each adsorption position.

9. A computer case with a magnetically attached transparent side panel, characterized in that, include: Chassis body; A transparent side panel is detachably mounted on one side of the main chassis body; Multiple electromagnetic adsorption components are respectively disposed on the main body of the chassis at each adsorption position corresponding to the transparent side panel, and each electromagnetic adsorption component includes an electromagnetic coil. A controller is connected to each of the electromagnetic coils. The controller is used to execute the electromagnetic adsorption control method according to any one of claims 1-8 to adjust the output current of each of the electromagnetic coils so that the force at each adsorption position is balanced.