Server visual inspection equipment and method thereof

Abstract

The invention relates to visual detection equipment for a server, and relates to the technical field of server detection. The three-axis moving assembly is installed on the top face of the equipment workbench and erected above the conveying belt in a crossing mode, and a first visual camera is arranged at the free end of the three-axis moving assembly; preliminarily detecting the quality of the cleaning assembly, and if the detection value is consistent, continuously conveying to the lower part of the three-axis moving assembly for detection; if not, inputting the environmental parameter information acquired by the acquisition module into an analysis module, and if cleaning is needed, starting a preliminary quality inspection cleaning assembly to clean the device; and the quality inspection and sorting mechanism is arranged on the two sides of the conveying belt and is close to one end of the qualified product discharging end, and unqualified products and qualified products are separately discharged by performing secondary quality inspection and sorting on the servers detected by the first visual camera. According to the invention, intelligent pre-adjustment and adaptive optimization of the detection environment are realized, and the technical effects of improving the image quality and reducing the false detection rate are achieved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of server detection, in particular to a server visual detection device and a method thereof. BACKGROUND

[0002] With the rapid development of information technology, servers as the core equipment of data centers, their performance and stability directly affect the operation efficiency and service quality of the entire information system. In order to ensure that the server meets high-quality requirements before leaving the factory, each server must be strictly and carefully detected. The traditional server detection method mainly relies on manual visual inspection and simple automatic test tools, which not only consumes time and effort, but also is difficult to ensure the consistency and accuracy of detection. In addition, the traditional detection method can only detect specific parts or single fault types, and lacks comprehensiveness and flexibility. Therefore, developing a server visual detection device that is efficient, accurate and can automatically complete multi-dimensional detection tasks has become a problem to be solved.

[0003] In modern manufacturing industry, especially for precise electronic equipment such as servers, it is crucial to ensure that each component meets strict quality standards. However, the existing server detection methods have many shortcomings: first, due to the limitations of manual operation, it is easily affected by human factors, resulting in unstable detection results; second, the existing automatic detection system can only realize detection in local area or fixed mode, and cannot meet the detection needs of full range and high precision under complex structure; third, for server components sensitive to environmental conditions (such as dust, light), detection in an undesirable environment may introduce additional errors, affecting the accuracy of the final judgment. In the traditional visual detection system, the server is directly fed into the detection station after feeding, and there is no pre-judgment and adjustment structure for the environment state (such as surface dust, environmental light) before detection. SUMMARY

[0004] The purpose of the present application is to solve the above problems by providing a server visual detection device and a method thereof.

[0005] To achieve the above requirements, the technical solution adopted by the present application to solve its technical problems is:

[0006] A server visual detection device is provided, comprising:

[0007] A device workbench, the top surface of which is equipped with a conveying belt, one end of the conveying belt is set as a server feeding end, and the other end is set as a qualified product discharging end;

[0008] A three-axis moving assembly is installed on the top surface of the equipment workbench and across above the conveying belt, and a free end terminal of the three-axis moving assembly is provided with a first visual camera;

[0009] A preliminary quality inspection and cleaning assembly is arranged above the top surface of the equipment workbench and close to one end of the server feeding end, and is used for preliminarily detecting whether the environmental parameters of the feeding server meet the subsequent detection environment (including detecting whether the dust of the server to be detected component is clean, and whether the environmental brightness and light are sufficient), if yes, the server is continuously conveyed to below the three-axis moving assembly for detection, if not, the environmental parameter information collected by the collection module is input into the analysis module, and if cleaning is needed, the preliminary quality inspection and cleaning assembly is started to clean the device.

[0010] A quality inspection and sorting mechanism is arranged on both sides of the conveying belt and close to one end of the qualified product discharging end, and the server detected by the first visual camera is subjected to secondary quality inspection and sorting, so that the unqualified product and the qualified product are discharged separately.

[0011] In the embodiment, inverted U-shaped supports are symmetrically installed on both sides of the conveying belt and on the top surface of the equipment workbench, and the three-axis moving assembly is installed on the top ends of the two U-shaped supports and above the conveying belt.

[0012] In the embodiment, the three-axis moving assembly includes X-axis guide rails respectively installed on the top ends of the two U-shaped supports, a Y-axis guide rail slidingly installed between the top ends of the two X-axis guide rails, and a Z-axis hydraulic column slidingly installed on the outer wall of the Y-axis guide rail, the first visual camera is fixedly installed on the bottom lifting free end of the Z-axis hydraulic column, and the first visual camera faces the conveying belt.

[0013] In the embodiment, the Y-axis guide rail is fixedly installed with a first cross beam and a second cross beam on one side close to the server feeding end and between the top ends of the two U-shaped supports, and the preliminary quality inspection and cleaning assembly is installed on the first cross beam and the second cross beam.

[0014] In the embodiment, the preliminary quality inspection and cleaning assembly includes a second cross beam installed below the first cross beam, cleaning cotton for cleaning dust parts of the server, and an LED light supplementing lamp for adjusting the brightness or color temperature of the light supplementing lamp.

[0015] In the embodiment, the bottom surface of the second cross beam is provided with a collection module, the top end of the second cross beam is connected with the free end of the collection module, the bottom surface of the second cross beam is fixedly provided with a hydraulic lifting rod, the bottom lifting end of the hydraulic lifting rod is fixedly provided with an inverted U-shaped lifting frame, the cleaning cotton is rotatably installed in the U-shaped lifting frame, a servo motor is installed on one side outer wall of the U-shaped lifting frame, and the output shaft of the servo motor is connected with the cleaning cotton.

[0016] In this embodiment, an L-shaped lifting rod is welded to one side of the hydraulic lifting rod and on the outer wall of the second crossbeam. The LED supplementary light is installed at the bottom end of the L-shaped lifting rod. An integrated module box is installed on the outer wall of one of the U-shaped brackets. The integrated module box contains: a detection module: using a first vision camera or other imaging device (such as an infrared camera) to capture the appearance image of the server or the working status of the internal components, and using computer vision algorithms to process and analyze the collected image data (such as detecting physical damage, overheating, or abnormal indicator light status through comparative analysis);

[0017] Fault prediction and diagnosis module: Based on a machine learning model, it performs fault prediction and diagnosis on information extracted from images (such as identifying hardware faults and cable connection problems); when a fault is detected, the system compares and analyzes to determine the type of problem, and then drives the corresponding pusher cylinder based on the identified problem result (attached). Figure 1 The three cylinders shown represent three different problems. For example, when problem A is detected, cylinder A is activated to push the defective server onto the conveyor belt in area a of the corresponding defective product unloading area, and so on. This allows for the classification and processing of defective servers, which is beneficial for subsequent problem repair.

[0018] In this embodiment, the quality inspection and sorting mechanism includes multiple pusher cylinders equidistantly arranged on one side of the conveyor belt and multiple non-conforming product unloading conveyor belts equidistantly arranged on the other side of the conveyor belt, with each pusher cylinder aligned with a non-conforming product unloading conveyor belt.

[0019] In this embodiment, a fixed upright plate is welded to one side of the conveyor belt and located on the top surface of the equipment workbench. Multiple pushing cylinders are fixedly installed on the inner wall of the fixed upright plate. A pushing block is fixed to the free end of the piston rod of each pushing cylinder, and a silicone protective pad is adhered to the outer wall of the pushing block. This allows the pushing cylinder to drive the piston rod to extend and push the defective product server through the silicone protective pad onto the corresponding defective product unloading conveyor belt for unloading. A vision camera initially checks whether the environmental parameters of the detection area meet the requirements for subsequent detection (including whether the server component to be detected is dusty and whether the ambient light is sufficient). If it meets the requirements, the system continues to transport the product to the area below the Z-axis for detection. If it is not suitable, the environmental parameter information collected by the acquisition module is input into the analysis module. If dust is present, the motor is activated to drive the cleaning cotton for cleaning; if the light is insufficient, the brightness of the adjustable LED supplementary light is controlled.

[0020] A detection method for a server visual inspection device includes the following steps:

[0021] S1: Place the server to be inspected from the server loading end onto the conveyor belt, and start the conveyor belt to transport the server to the area below the preliminary quality inspection and cleaning component;

[0022] S2: Collect the environmental parameters of the server parts to be detected by the acquisition module, including surface cleanliness and environmental light intensity; if there is dust on the surface, the servo motor is started to drive the cleaning cotton to rotate, and the hydraulic lifting rod drives the cleaning cotton to descend to contact the surface of the server for cleaning; if the environmental light is insufficient, the brightness or color temperature of the LED fill light at the bottom of the L boom is adjusted to meet the detection requirements;

[0023] S3: After the environmental parameters meet the standards, the server continues to be transported under the three-axis moving assembly, and the first visual camera performs multi-angle and high-precision image acquisition on the server under the cooperation of the Z-axis hydraulic column, Y-axis guide rail and X-axis guide rail;

[0024] S4: The collected image data is transmitted to the detection module in the integrated module box, and the computer vision algorithm is used to analyze the appearance and internal component state of the server to determine whether there is physical damage, overheating area or abnormal indicator light;

[0025] S5: The analysis result is input into the fault prediction and diagnosis module, and the potential hardware fault or connection problem is predicted and classified based on the machine learning model;

[0026] S6: If the server is determined to be qualified, it continues to be transported to the qualified product discharge end for discharge; if it is determined to be unqualified, the corresponding push cylinder is matched according to the fault type, the control system starts the corresponding cylinder, and the server is pushed to the corresponding number of unqualified product discharge conveyor belt through the push block and silica gel protective pad, realizing classified discharge;

[0027] S7: After sorting is completed, the equipment is reset, and the next round of detection operation is prepared.

[0028] The beneficial effects of the present application are:

[0029] The server visual detection equipment, by setting the preliminary quality inspection and cleaning assembly near the feeding end above the equipment workbench, combining the real-time monitoring of the environmental parameters (such as dust and light) of the server parts to be detected by the acquisition module, and based on the analysis result, the design of automatically starting the cleaning cotton to clean or adjusting the brightness of the LED fill light, makes the environmental conditions before detection always keep in the best state, avoids the problems of image blur, feature loss caused by surface pollution or insufficient light, realizes intelligent pre-adjustment and self-adaptive optimization of the detection environment, and achieves the technical effects of improving image quality and reducing false detection rate.

[0030] The three-dimensional linkage structure designed by the triaxial moving assembly composed of the X-axis guide rail, the Y-axis guide rail and the Z-axis hydraulic column enables the first visual camera to freely move in space and accurately position to any detection point, and the Z-axis lifting function is combined to realize focal length adjustment and close-range shooting, so that multi-dimensional and multi-scale imaging of the complex structure (such as a slot, an interface and a heat sink) of the server at different heights and angles can be realized, flexible detection coverage under complex working conditions is realized, and the technical effects of comprehensively identifying micro defects and hidden faults are achieved.

[0031] By arranging the quality inspection sorting mechanism including a plurality of pushing cylinders and a corresponding number of unqualified product discharge conveyors, and combining the fixed vertical plate to realize stable installation of the cylinder, after the first visual camera completes image acquisition and the fault diagnosis module determines that it is unqualified, the system can accurately drive the corresponding numbered pushing cylinder to act according to the fault type, and the server is gently pushed to the specified unqualified product discharge conveyor through the pushing block and the silica gel protective pad, so that classified discharge and accurate shunting of different types of defective products are realized, mixing and manual secondary sorting are avoided, and the effects of improving sorting accuracy, supporting subsequent targeted maintenance and quality tracing are achieved.

[0032] By arranging the detection module, the fault prediction and diagnosis module and the control logic in the integrated module box, the image data are automatically analyzed and classified by using computer vision algorithm and machine learning model, and the result is directly converted into the execution instruction of the pushing cylinder, so that the whole detection-judgment-sorting process is completely automated and intelligent, and the closed-loop control from image acquisition to result output can be completed without manual intervention, the autonomous decision-making ability of the detection system is realized, and the comprehensive benefits of improving production rhythm, reducing labor cost and enhancing system scalability are achieved. BRIEF DESCRIPTION OF DRAWINGS

[0033] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the present application will be further described below with reference to the drawings and embodiments. The drawings in the following description are only some embodiments of the present application, and those skilled in the art can obtain other drawings according to these drawings without creating any creative labor:

[0034] Figure 1 The structure of the present application is shown in the figure;

[0035] Figure 2 The installation structure of the cleaning cotton of the present application is shown in the figure;

[0036] Figure 3 The structure of the LED light supplementing lamp of the present application is shown in the figure;

[0037] Figure 4 The structure of the pushing cylinder of the present application is shown in the figure.

[0038] Marked for explanation:

[0039] In the figure: 1, equipment workbench; 2, conveying belt; 3, server feeding end; 4, qualified product discharging end; 5, U-shaped support; 6, X-axis guide rail; 7, Y-axis guide rail; 8, Z-axis hydraulic column; 9, first vision camera; 10, first cross beam;

[0040] 11, second cross beam; 12, second cross beam; 13, fixed vertical plate; 14, pushing cylinder; 15, unqualified product discharging conveying belt; 16, three-axis moving assembly; 17, quality inspection sorting mechanism; 18, acquisition module; 19, integrated module box; 20, hydraulic lifting rod;

[0041] 21, cleaning cotton; 22, U-shaped lifting frame; 23, servo motor; 24, L-shaped lifting rod; 25, LED light supplementing lamp; 26, pushing block; 27, silica gel protective pad. DETAILED DESCRIPTION

[0042] The terms "first", "second", "third", and "fourth" and the like in the description and in the claims of the present application and the accompanying drawings are used for distinguishing between similar objects, not necessarily described in a particular order. Also, the terms "comprise", "comprising", "including", and "having" and any variations thereof are intended to cover a non-exclusive inclusion, such that a process, method, system, product, or apparatus that comprises a list of steps or units is not necessarily limited to the steps or units that are listed, but can optionally include additional steps or units not listed, or can also include steps or units inherent to such process, method, product, or apparatus.

[0043] Reference herein to "an embodiment" means that a particular feature, structure, or characteristic described in connection with the embodiment can be included in at least one embodiment of the application. The appearances of the phrase in various places in the specification are not necessarily all referring to the same embodiment, nor are they necessarily mutually exclusive or alternative embodiments. It is expressly understood that the embodiments described herein are merely examples from a whole class of comparable embodiments which those skilled in the art will readily appreciate. It is also expressly understood that the embodiments described herein with either the same, corresponding or analogous elements are interchangeable such that the implications of embodiments recited in the alternative are also implicitly implied.

[0044] "Multiple" refers to two or more. "And / or", describes the association relationship of the associated objects, which means that there can be three relationships, for example, A and / or B, which can represent the three cases of A alone, A and B together, and B alone. The character " / " generally represents that the front and rear associated objects are in an "or" relationship.

[0045] Moreover, the terms "upper, lower, left, right, upper end, lower end, longitudinal" and the like indicating the orientation are all with reference to the attitude position of the device or equipment described in the scheme in normal use.

[0046] In order to make the purposes, technical solutions and advantages of the embodiments of the present application clearer, the following will make a clear and complete description of the technical solutions in the embodiments of the present application. Obviously, the described embodiments are some but not all of the embodiments of the present application. Based on the embodiments of the present application, all other embodiments obtained by those of ordinary skill in the art without creative effort belong to the protection scope of the present application.

[0047] The embodiment discloses a server visual detection device as shown in Figures 1 to 4 The embodiment discloses a server visual detection device as shown in

[0048] The device workbench 1 has a conveying belt 2 horizontally arranged on the top surface, one end of the conveying belt 2 is arranged as a server feeding end 3, and the other end is arranged as a qualified product discharging end 4.

[0049] The three-axis moving assembly 16 is arranged on the top surface of the device workbench 1 and horizontally arranged above the conveying belt 2, and the free end of the three-axis moving assembly 16 is provided with the first visual camera 9.

[0050] The preliminary quality inspection and cleaning assembly is arranged on the top surface of the device workbench 1 and close to one end of the server feeding end 3, and is used for preliminarily detecting whether the feeding environment parameters meet the subsequent detection environment (including whether the dust of the server to be detected component is clean, and whether the environmental brightness and light are sufficient), if yes, the server is continuously conveyed to the detection below the three-axis moving assembly 16; if not, the environmental parameter information collected by the collection module is input into the analysis module, and if cleaning is needed, the preliminary quality inspection and cleaning assembly is started to clean the device.

[0051] The quality inspection and sorting mechanism 17 is arranged on both sides of the conveying belt 2 and close to one end of the qualified product discharging end 4, and is used for secondarily detecting and sorting the server after the detection by the first visual camera 9, so that the unqualified product and the qualified product are discharged separately.

[0052] In the embodiment, the inverted U-shaped supports 5 are symmetrically arranged on both sides of the conveying belt 2 and on the top surface of the device workbench 1, and the three-axis moving assembly 16 is arranged on the top ends of the two U-shaped supports 5 and above the conveying belt 2.

[0053] In the embodiment, the three-axis moving assembly 16 includes the X-axis guide rail 6 arranged on the top ends of the two U-shaped supports 5, the Y-axis guide rail 7 slidingly arranged between the top ends of the two X-axis guide rails 6, and the Z-axis hydraulic column 8 slidingly arranged on the outer wall of the Y-axis guide rail 7, and the first visual camera 9 is fixedly arranged on the bottom lifting free end of the Z-axis hydraulic column 8, and the first visual camera 9 faces the conveying belt 2.

[0054] In this embodiment, the Y-axis guide rail 7 is fixedly installed with a first cross beam 10 and a second cross beam 11 at the top of the two U-shaped supports 5 on the side close to the server feeding end 3.

[0055] In this embodiment, the preliminary quality inspection and cleaning assembly includes a second cross beam 12 installed below the first cross beam 10, cleaning cotton 21 for cleaning the dust part of the server, and an LED light supplementing lamp 25 for adjusting the brightness or color temperature of the light supplementing lamp.

[0056] In this embodiment, the bottom surface of the second cross beam 12 is installed with a collection module 18, the top end of the second cross beam 12 is connected with the free end of the collection module 18, the bottom surface of the second cross beam 11 is fixedly installed with a hydraulic lifting rod 20, the bottom lifting end of the hydraulic lifting rod 20 is fixedly installed with an inverted U-shaped lifting frame 22, the cleaning cotton 21 is rotatably installed in the U-shaped lifting frame 22, a servo motor 23 is installed on one side outer wall of the U-shaped lifting frame 22, and the output shaft of the servo motor 23 is connected with the cleaning cotton 21.

[0057] In this embodiment, an inverted L-shaped lifting rod 24 is welded on one side of the hydraulic lifting rod 20 and located on the outer wall of the second cross beam 11, and the LED light supplementing lamp 25 is installed at the bottom end of the L-shaped lifting rod 24. One of the outer walls of the U-shaped support 5 is installed with an integrated module box 19, which has: a detection module: using a first visual camera 9 or other imaging equipment (such as an infrared camera) to capture the appearance image of the server or the working state of the internal components, using a computer vision algorithm to process and analyze the collected image data (such as detecting whether there is physical damage, overheating phenomenon, or abnormal indicator light state through comparison analysis);

[0058] a fault prediction and diagnosis module: based on a machine learning model, the information extracted from the image is used for fault prediction and diagnosis (such as identifying hardware failure, cable connection problem); when detecting a fault problem, the system compares and analyzes which problem, and then drives the corresponding push cylinder based on the identified problem result (as shown in the figure, three cylinders represent three different problem corresponding cylinders, for example: when detecting A problem, the A cylinder is started to push the unqualified server to the a area conveying belt of the corresponding unqualified product discharge area, and so on, so as to classify and process the unqualified server, which is beneficial to the repair of subsequent problems). Figure 1

[0059] The detection module is used to collect image data of the server surface and internal components through the first visual camera (9) or other imaging equipment (such as an infrared thermal imaging camera), and to process and analyze the image by using a computer vision algorithm to determine whether there is physical damage, abnormal indicator light state or abnormal heat dissipation. ​

[0060] Image acquisition and pre-processing:

[0061] Visible light band images of the server appearance are acquired using a high-resolution industrial camera (first vision camera 9).

[0062] If the overheating problem needs to be detected, an infrared thermal imaging camera is enabled to synchronously acquire a temperature distribution map.

[0063] The pre-processing steps include:

[0064] Grayscale: convert the color image I rgb (x,y) to a grayscale image I gray (x,y):

[0065] I gray (x,y) = 0.299R(x,y) + 0.587G(x,y) + 0.114B(x,y) I gray (x,y): represents the grayscale image pixel value at coordinate position (x,y). The value is an integer between 0 (black) and 255 (white), representing the brightness intensity of the point.

[0066] R(x,y): represents the red channel (Red) pixel value of the original color image at coordinate (x,y).

[0067] G(x,y): represents the green channel (Green) pixel value of the original color image at coordinate (x,y).

[0068] B(x,y): represents the blue channel (Blue) pixel value of the original color image at coordinate (x,y).

[0069] 0.299, 0.587, 0.114: are weighting coefficients reflecting the sensitivity of the human eye to different colors of light (the human eye is most sensitive to green, followed by red, and least sensitive to blue). These coefficients are internationally recognized standards (such as ITU-R BT.601) for reasonably integrating RGB three-channel information into a single grayscale value.

[0070] In this technical solution: when the first vision camera 9 captures the color image of the server surface, the system first converts it to a grayscale image to facilitate subsequent image processing operations such as edge detection and template matching, reducing computational complexity and improving processing efficiency.

[0071] Denoising: Gaussian filtering or median filtering is used to suppress image noise:

[0072]

[0073] where: w(i,j) is the Gaussian kernel weight;

[0074] I filtered (x, y): denotes the pixel value of the denoised image at coordinate (x, y) after filtering.

[0075] I gray (x+i, y+j): denotes the pixel value of the neighborhood in the original grayscale image centered at (x, y) with offset (i, j).

[0076] w(i, j): denotes the weight coefficient in the filter kernel (or convolution kernel), which determines the influence of the neighborhood pixels on the center pixel. For example:

[0077] If it is a Gaussian filter, w(i, j) follows a two-dimensional Gaussian function distribution, with the center weight being the largest and the edge gradually decreasing; if it is a median filter, this formula is not directly applicable (median filtering is a nonlinear operation), but its effect is also to suppress noise.

[0078] k: denotes the radius of the filter window. For example, k = 1 means using a 3x3 neighborhood window for calculation.

[0079] In this technical solution: During server detection, due to environmental light fluctuations or camera sensor noise, the collected images may appear "salt and pepper noise" or "Gaussian noise". Through this filtering algorithm, the image can be effectively smoothed and random noise points can be removed, improving the accuracy of subsequent feature extraction, especially when identifying small scratches or indicator light states.

[0080] Light normalization: using histogram equalization or CLAHE (contrast limited adaptive histogram equalization) to improve image contrast and eliminate environmental light effects.

[0081] Feature extraction and target detection: using a deep learning target detection model (such as YOLOv5, Faster R-CNN) to identify key components:

[0082] The detected objects include: power indicator light, hard drive status light, fan, interface terminal, heat sink, etc.

[0083] The model outputs the bounding box B = (x c ,y c ,w,h) and class label C for each target.

[0084] Where B: denotes the bounding box of a target detection, used to frame the identified key components in the image.

[0085] x c : denotes the horizontal coordinate (abscissa) of the center point of the bounding box in the image, in pixels.

[0086] y cx: represents the horizontal coordinate of the center point of the bounding box in the image, in pixels.

[0087] w: represents the width of the bounding box, i.e., the pixel span of the target in the horizontal direction.

[0088] h: represents the height of the bounding box, i.e., the pixel span of the target in the vertical direction.

[0089] In this technical solution: after the first visual camera 9 takes the server image, the "power indicator", "hard disk status light", "interface terminal" and other key components are detected by YOLO or Faster R-CNN model, and the position is accurately positioned by the bounding box, providing ROI (Region of Interest) for subsequent color recognition and state analysis.

[0090] Using convolutional neural network (CNN) to extract image features:

[0091] F l =σ(W l *F l-1 +b l )

[0092] where F l : represents the feature map output by the l-th convolutional neural network, which is a two-dimensional or three-dimensional matrix, recording the abstract features (such as edges, textures, shapes, etc.) extracted from the image at this layer.

[0093] F l-1 : represents the output feature map of the previous layer (l-1 layer), which is the input of the current layer.

[0094] W l : represents the convolution kernel weight matrix (Filter / KernelWeights) of the l-th layer, which is a parameter learned during network training, determining how to extract new features from input features.

[0095] *: represents the convolution operation, i.e., the convolution kernel slides on the input feature map, point-by-point computing the weighted sum, used to extract local spatial features.

[0096] b l : represents the bias term of the l-th layer, which is a scalar or vector, used to adjust the baseline of the activation value.

[0097] σ: represents the activation function, introducing nonlinearity to enable the network to fit complex functions. Commonly used: ReLU: σ(x) = max(0, x).

[0098] In this technical solution: after the server image is input, it undergoes feature extraction through multiple layers of CNN, and finally outputs high-level semantic information that can be used for target detection and defect segmentation. For example, shallow layers extract edges and corners, while deep layers identify whether "indicator lights are on" or "screws are missing".

[0099] Abnormal state judgment logic:

[0100] Indicator light status recognition: Color segmentation (HSV space) is performed on the detected LED area:

[0101]

[0102] Where H represents the hue value of a pixel in the image, which is a component of the HSV color space and reflects the basic color type (such as red, green, and blue). The value range is usually [0°, 360°] or [0, 180] (the latter is commonly used in OpenCV).

[0103] H∈[0°,10°]∪[350°,360°]: Represents red, which the system interprets as an "alarm state".

[0104] H∈[60°,80°]: Represents green, which the system interprets as "normal operation".

[0105] H∈[100°,140°]: Represents blue, which the system identifies as "standby mode".

[0106] The HSV color space is more consistent with human color perception than RGB, making it particularly suitable for color recognition under varying lighting conditions. The system first converts the RGB image to HSV, then performs a threshold judgment on the H value of the LED area, thereby achieving stable state recognition and avoiding misjudgments caused by changes in brightness.

[0107] In this technical solution, the system automatically identifies whether the server power indicator, hard drive activity indicator, etc., are in normal working condition, eliminating the need for manual inspection and improving the level of automation.

[0108] Physical damage detection: Using edge detection algorithms (Canny) or defect segmentation models (U-Net) to identify anomalies such as scratches, deformation, and detachment.

[0109]

[0110] If there is a sudden change in edge strength or a discontinuous shape, it is considered a damage.

[0111] The fault prediction and diagnosis module is based on a machine learning model that fuses and analyzes multidimensional features extracted from images to achieve fault prediction and classification.

[0112] in, Gradient magnitude of image I at a point, reflecting the intensity change of pixels around that point. The larger the gradient, the more obvious the edge.

[0113] Partial derivative of image in horizontal direction (x-axis), i.e. pixel difference in column direction, reflecting horizontal edges.

[0114] Partial derivative of image in vertical direction (y-axis), i.e. pixel difference in row direction, reflecting vertical edges.

[0115] I: represents the input grayscale image or preprocessed image function, I(x,y) is the pixel value at coordinates (x,y).

[0116] In this technical solution: use Canny edge detection algorithm to calculate the gradient value, identify scratches, cracks, deformation and other physical damage on the server shell, interface, circuit board. If the gradient value of a certain area suddenly changes and forms a line, it is determined as an edge; if the edge is broken, misaligned or abnormally dense, it is determined as a damage.

[0117] Feature vector construction: extract the following features from the output of the detection module to form the input vector

[0118] Visual features: number of missing parts, number of abnormal indicator lights, scratch area ratio, color offset degree max Thermal imaging features: maximum temperature T max T avg

[0119] Historical data: average failure rate of similar servers, usage time, maintenance record code.

[0120] Model selection and training: use integrated learning model (such as XGBoost) or deep neural network (DNN) for classification prediction: XGBoost model objective function:

[0121]

[0122] Where:

[0123] Objective function of the model, i.e. loss value to be minimized, used to guide model training.

[0124] θ: represents the set of all learnable parameters of the model, including the structure and weight of each tree.

[0125] N: represents the total number of training samples. For example, the system accumulates N server detection data of known fault types.

[0126] yi : True label of the i-th sample, such as "power failure", "hard disk loose", etc.

[0127] : Predicted output value of the i-th sample by the model. In classification tasks, it is usually the class probability or logit value.

[0128] : Loss function that measures the gap between the predicted value and the true value. Commonly used are:

[0129] Logistic Loss (binary classification):

[0130] Cross-Entropy Loss (multi-classification)

[0131] K: represents the total number of regression trees (weak learners) constructed in the XGBoost model.

[0132] f k : The k-th regression tree, which is a mapping function from the feature space to the predicted value.

[0133] : The final predicted value is the sum of the outputs of all trees, embodying the additive model idea of ensemble learning.

[0134] Ω(f k ): represents the complexity penalty term of the k-th tree, used to prevent overfitting.

[0135]

[0136] T: the number of leaf nodes of the tree, which controls the structural complexity of the tree, and γ is the leaf node penalty coefficient.

[0137] ‖w‖ 2 : The sum of the squares of all leaf node output weights w, and λ is the L2 regularization coefficient, used to smooth the predicted value.

[0138] In this technical solution: through the objective function, the system can train a high-precision classification model that can accurately identify the current fault and generalize to new samples, supporting intelligent fault classification of servers.

[0139] DNN structure example:

[0140] Input layer: n-dimensional feature vector

[0141] Hidden layer: 3 fully connected layers, ReLU activation

[0142] Output layer: Softmax classification, output the probability of each fault:

[0143]

[0144] where P(c j | x): represents the probability of the server belonging to the jth failure class given the input features x. For example, P(fan failure | x) = 0.87.

[0145] c j : represents the jth pre-defined failure class, such as:

[0146] c1: power module abnormality

[0147] c2: hard disk connection loose

[0148] c3: heat sink shedding

[0149] C: represents the total number of failure classes.

[0150] z j : represents the linear combination result of the jth output neuron, which is the input of Softmax.

[0151] W j : represents the jth row of the weight matrix connecting the input layer and the output layer, which is the feature weight vector corresponding to the jth failure class.

[0152] x: input feature vector (same as above).

[0153] b j : the bias term (Bias) of the jth class, used to adjust the classification boundary.

[0154] Exponential transformation of the logit value to convert it to a non-negative number.

[0155] Denominator Normalization factor to ensure the sum of probabilities of all classes is 1.

[0156] Contrastive analysis and sorting control logic: image contrastive analysis is performed on the image to be tested I test and the standard template image I ref to calculate the similarity:

[0157]

[0158] Where: if S < 0.85, it is determined to be an appearance anomaly.

[0159] S: represents the image similarity score, which is a real number between 0 and 1, reflecting the structural similarity between the image to be tested and the standard template.

[0160] S→1: indicates that the two images are highly similar.

[0161] S→0: indicates the difference is very large.

[0162] Decision threshold: if S < 0.85, the system determines that the appearance is abnormal, such as missing parts, misassembly, label error, etc.

[0163] I test : indicates the current detection image of the server to be tested, which is the gray or pre-processed image data of the first visual camera (9) in real time, usually a two-dimensional pixel matrix.

[0164] Each pixel point is denoted as I test (x,y), which represents the brightness value (0-255) at coordinate (x,y).

[0165] I ref : indicates the standard template image (ReferenceTemplateImage), which is the ideal image of the "qualified product" server under the same viewing angle, which is used as a comparison reference.

[0166] It can be obtained by learning multiple qualified product images and averaging to reduce the influence of individual differences.

[0167] Indicates the average gray value (MeanIntensity) of the image to be tested, and the calculation method is:

[0168]

[0169] Where N is the total number of image pixels. It is used to remove the influence of overall brightness offset of the image.

[0170] Sorting execution logic, control system receives fault category c *

[0171] Look up the mapping table:

[0172] Fault type Pushing cylinder number Defective product discharging conveyor Class A fault 14-1 15-1 Class B fault 14-2 15-2 Class C fault 14-3 15-3

[0173] The output control signal drives the corresponding push cylinder 14 to act, the stroke d = 100 mm, the pushing force F = 200 N.

[0174] In this embodiment, the quality inspection sorting mechanism 17 includes a plurality of push cylinders 14 arranged equidistantly on one side of the conveying belt 2, and a plurality of unqualified product discharge conveyors 15 arranged equidistantly on the other side of the conveying belt 2, each push cylinder 14 aligning with an unqualified product discharge conveyor 15.

[0175] In this embodiment, one side of the conveying belt 2 is welded with a fixed vertical plate 13 on the top surface of the equipment workbench 1, and a plurality of pushing cylinders 14 are fixedly installed on the inner wall of the fixed vertical plate 13. The free end of the piston rod of each pushing cylinder 14 is fixed with a pushing block 26, and the outer wall of the pushing block 26 is bonded with a silica gel protective pad 27. The pushing cylinder 14 drives the piston rod to extend through the silica gel protective pad 27 to push the unqualified server to the corresponding unqualified server discharging conveying belt 15 for discharging. The visual camera preliminarily detects whether the environmental parameters of the detection area meet the subsequent detection environment (including whether the dust on the server to be detected component is clean, and whether the environmental brightness and light are sufficient), if yes, the server continues to be conveyed to the detection below the Z-axis; if not, the environmental parameter information collected by the collection module is input into the analysis module, if there is dust, the motor is started to drive the cleaning cotton to clean, and if the light is insufficient, the brightness of the adjustable LED fill light is controlled.

[0176] Preliminary quality inspection environmental parameter analysis and control logic: environmental brightness detection and light supplement control, the light sensor in the collection module (18) collects the environmental illumination L (unit: lux), and the threshold value L is set th = 300 lux th , the LED light supplement lamp (25) is started, and the brightness I is adjusted:

[0177] I = k p (L th -L) + I0

[0178] Wherein k p is the proportional gain, and I0 is the basic brightness.

[0179] A detection method of a server visual detection device, comprising the following steps:

[0180] S1: placing the server to be detected on the conveying belt 2 from the server feeding end 3, and starting the conveying belt 2 to convey the server to below the preliminary quality inspection cleaning assembly;

[0181] S2: collecting the environmental parameters of the server to be detected part by the collection module 18, including surface cleanliness and environmental illumination intensity; if there is dust on the surface, the servo motor 23 is started to drive the cleaning cotton 21 to rotate, and the hydraulic lifting rod 20 drives the cleaning cotton 21 to descend to contact the surface of the server for cleaning; if the environmental illumination is insufficient, the brightness or color temperature of the LED light supplement lamp 25 at the bottom end of the L hanger 24 is adjusted to meet the detection requirements;

[0182] S3: after the environmental parameters meet the standards, the server continues to be conveyed to below the three-axis moving assembly 16, and the first visual camera 9 performs multi-angle and high-precision image collection on the server under the cooperation of the Z-axis hydraulic column 8, the Y-axis guide rail 7 and the X-axis guide rail 6;

[0183] S4: The collected image data is transmitted to the detection module in the integrated module box 19, and the computer vision algorithm analysis server appearance and internal component state is used to judge whether there is physical damage, overheating area or abnormal indicator light;

[0184] S5: The analysis result is input into the fault prediction and diagnosis module, and the potential hardware failure or connection problem is predicted and classified based on the machine learning model;

[0185] S6: If the server is determined to be qualified, it continues to be conveyed to the qualified product discharge end 4 for discharge; if it is determined to be unqualified, the corresponding push cylinder 14 is matched according to the fault type, the control system starts the corresponding cylinder, and the server is pushed to the corresponding number of unqualified product discharge conveyor belt 15 through the push block 26 and the silica gel protective pad 27, so as to realize classified discharge;

[0186] S7: After the sorting is completed, the equipment is reset, and the next round of detection operation is prepared.

[0187] It should be understood that, for those skilled in the art, improvements or changes can be made according to the above description, and all these improvements and changes shall belong to the protection scope of the appended claims of the present application.

Claims

1. A server visual inspection device, characterized in that, include: The equipment workbench has a conveyor belt spanning across its top surface. One end of the conveyor belt is set as the server loading end, and the other end is set as the qualified product unloading end. A three-axis motion assembly is installed on the top surface of the equipment workbench and spans across the conveyor belt. A first vision camera is provided at the free end of the three-axis motion assembly. The preliminary quality inspection and cleaning component includes one end located above the top surface of the equipment workbench and near the server loading end, for preliminary inspection of the loaded server; The quality inspection and sorting mechanism is located on both sides of the conveyor belt and near the end of the qualified product unloading. It performs secondary quality inspection and sorting on the server after it has been inspected by the first vision camera, so that unqualified products and qualified products are unloaded separately.

2. The server visual inspection device according to claim 1, characterized in that: The conveyor belt is symmetrically mounted on both sides and on the top of the equipment workbench with inverted U-shaped brackets, and the three-axis moving assembly is mounted on the top of the two U-shaped brackets and located directly above the conveyor belt.

3. The server visual inspection device according to claim 2, characterized in that: The three-axis moving assembly includes an X-axis guide rail mounted on the top of two U-shaped supports, a Y-axis guide rail slidably mounted between the tops of the two X-axis guide rails, and a Z-axis hydraulic column slidably mounted on the outer wall of the Y-axis guide rail. The first vision camera is fixedly mounted on the bottom lifting free end of the Z-axis hydraulic column.

4. The server visual inspection device according to claim 3, characterized in that: The Y-axis guide rail is fixedly installed on the side near the loading end of the server and between the tops of the two U-shaped brackets. The first crossbeam and the second crossbeam are respectively fixedly installed on the first crossbeam and the second crossbeam. The preliminary quality inspection and cleaning component is installed on the first crossbeam and the second crossbeam.

5. A server visual inspection device according to claim 4, characterized in that: The preliminary quality inspection and cleaning components include a second crossbeam installed below the first crossbeam, a cleaning cotton installed below the second crossbeam for cleaning dusty parts of the server, and an LED fill light for adjusting the brightness or color temperature of the fill light.

6. The server visual inspection device according to claim 5, characterized in that: A data acquisition module is installed on the bottom surface of the second crossbeam, and the top of the second crossbeam is connected to the free end of the data acquisition module. A hydraulic lifting rod is fixed on the bottom surface of the second crossbeam, and an inverted U-shaped hoisting frame is fixed at the bottom lifting end of the hydraulic lifting rod. The cleaning cotton is rotatably installed in the U-shaped hoisting frame, and a servo motor is installed on one outer wall of the U-shaped hoisting frame. The output shaft of the servo motor is connected to the cleaning cotton.

7. A server visual inspection device according to claim 6, characterized in that: An L-shaped lifting rod is welded to one side of the hydraulic lifting rod and on the outer wall of the second crossbeam. The LED supplementary light is installed at the bottom end of the L-shaped lifting rod. An integrated module box is installed on the outer wall of one of the U-shaped brackets.

8. A server visual inspection device according to claim 7, characterized in that: The quality inspection and sorting mechanism includes multiple pusher cylinders equidistantly arranged on one side of the conveyor belt and multiple non-conforming product unloading conveyor belts equidistantly arranged on the other side of the conveyor belt, with each pusher cylinder aligned with a non-conforming product unloading conveyor belt.

9. A server visual inspection device according to claim 8, characterized in that: A fixed upright plate is welded to one side of the conveyor belt and located on the top surface of the equipment workbench. Multiple pusher cylinders are fixedly installed on the inner wall of the fixed upright plate. A pusher block is fixed to the free end of the piston rod of each pusher cylinder. A silicone protective pad is adhered to the outer wall of the pusher block.

10. A server visual inspection method, characterized in that, Includes the following steps: S1: Place the server to be inspected from the server loading end onto the conveyor belt, and start the conveyor belt to transport the server to the area below the preliminary quality inspection and cleaning component; S2: The acquisition module collects environmental parameters of the server's inspection area, including surface cleanliness and ambient light intensity; if there is dust on the surface, the servo motor is activated to rotate the cleaning cotton, while the hydraulic lifting rod drives the cleaning cotton to descend to contact the server surface for cleaning; If the ambient light is insufficient, adjust the brightness or color temperature of the LED supplementary light at the bottom of the L-shaped boom to meet the testing requirements; S3: After the environmental parameters meet the standards, the server continues to be transported to the bottom of the three-axis moving component. The first vision camera performs multi-angle, high-precision image acquisition on the server under the coordinated action of the Z-axis hydraulic column, Y-axis guide rail and X-axis guide rail. S4: The acquired image data is transmitted to the detection module in the integrated module box. The computer vision algorithm is used to analyze the appearance of the server and the status of its internal components to determine whether there is physical damage, overheating area or abnormal indicator light. S5: Input the analysis results into the fault prediction and diagnosis module, and predict and classify potential hardware faults or connection problems based on the machine learning model; S6: If the server is deemed qualified, it continues to be conveyed to the qualified product unloading end for unloading; if it is deemed unqualified, the corresponding pusher cylinder is matched according to the fault type, the control system starts the corresponding cylinder, and pushes the server to the corresponding numbered unqualified product unloading conveyor belt through the pusher block and silicone protective pad to achieve classified unloading. S7: After sorting is completed, the equipment is reset to prepare for the next round of testing.

Images