A precision grinding control method for fuses that integrates target location and defect classification

Abstract

This invention discloses a precision grinding control method for fuses that integrates target localization and defect classification, relating to the fields of industrial automation and precision machining technology. The invention includes the following steps: S1: Acquiring three-dimensional point cloud data and high-resolution RGB images of the fuse using a multimodal vision sensor; S2: Target localization of the RGB image based on a lightweight convolutional neural network, outputting the pixel coordinates of the fuse, and mapping them to the robotic arm coordinate system through a perspective transformation matrix to calculate the initial gripping and grinding points; S3: Constructing a multi-scale feature fusion network to extract and classify features of the located defect area. This invention achieves sub-millimeter-level fuse positioning accuracy by introducing a perspective transformation matrix and an improved YOLOv8-Nano network. Even with a random deviation of ±13μm in the loading posture, it can accurately calculate the grinding points in the robotic arm coordinate system, solving the problems of high rigidity and poor adaptability in traditional mechanical positioning.

Description

Technical Field

[0001] This invention belongs to the field of industrial automation and precision machining technology, specifically a precision grinding control method for fuses that integrates target positioning and defect classification. Background Technology

[0002] In power systems, fuses are critical protective components, and the surface finish and geometric precision of fuses directly affect their heat dissipation performance and insulation strength. In traditional fuse manufacturing processes, cast fuses often have defects such as burrs, oxide scale, and micro-cracks.

[0003] Existing grinding technologies mainly suffer from the following problems: 1. Low positioning accuracy: Traditional methods often use mechanical tooling for positioning or simple photoelectric switch triggering, which cannot cope with random posture deviations during fuse loading, resulting in inaccurate contact positions between the grinding head and the workpiece, easily causing "over-grinding" damage to the substrate or "under-grinding" residual defects; 2. Defect identification and processing disconnect: Existing visual inspection is usually only used for post-processing sorting, lacking the ability to convert defect types such as deep scratches, pores, and inclusions into processing parameters in real time; for defect areas of different hardness, if a constant grinding pressure and speed are used, soft defects will be removed too quickly while hard areas will be removed insufficiently; 3. Open-loop control strategy: The grinding process is usually based on fixed G-codes or timed control, without considering the wear of the grinding head, uneven workpiece material, and dynamic changes in real-time defect depth, making it impossible to achieve true "precision" and "adaptive" control.

[0004] Therefore, there is an urgent need for a technical solution that can integrate precise target positioning, intelligent defect classification, and real-time closed-loop control of grinding process parameters to solve the above problems.

[0005] The information disclosed above in this background section is only for enhancing the understanding of the background section of this invention, and therefore may include prior art that is not known to those skilled in the art. Summary of the Invention

[0006] The purpose of this invention is to provide a precision grinding control method for fuses that integrates target positioning and defect classification, so as to solve the problems in the prior art.

[0007] To achieve the above objectives, the present invention provides the following technical solution: a precision grinding control method for fuses that integrates target localization and defect classification, comprising the following steps: S1: acquiring three-dimensional point cloud data and high-resolution RGB images of the fuse through a multimodal vision sensor; S2: performing target localization on the RGB images based on a lightweight convolutional neural network, outputting the pixel coordinates of the fuse, and mapping them to the robotic arm coordinate system through a perspective transformation matrix to calculate the initial gripping and grinding points; S3: constructing a multi-scale feature fusion network to extract and classify features of the located defect area, identifying the defect type and quantifying the defect depth level, and generating a defect feature vector V. def S4: Based on the defect feature vector V def Based on the initial grinding position, an adaptive grinding path is generated using the Signed Distance Field (SDF) algorithm, and the feed speed v of the grinding head is planned. f S5: Start the grinding operation, using a force sensor to collect the grinding contact force F in real time, along with the spindle speed n; real Based on the preset material removal rate model, the feed depth d of the grinding head is dynamically adjusted by an adaptive PID controller until the defects are removed and the surface roughness meets the standard.

[0008] Preferably, the lightweight convolutional neural network described in step S2 adopts an improved YOLOv8-Nano architecture and introduces a coordinate attention mechanism (CA), whose loss function for localizing coordinates is L. loc The bounding box regression loss, which incorporates IoU, is formulated as follows:

[0009]

[0010] Where IoU is the intersection-over-union ratio of the predicted bounding box and the ground truth bounding box, ρ is the Euclidean distance between the center points of the two boxes, and c is the minimum diagonal distance between the two bounding boxes. For bounding box parameters, Here are the actual parameters, and α is the weighting coefficient.

[0011] By minimizing L loc Obtain the optimal bounding box [x min ,y min ,x max ,y max The pixel coordinates (u,v) are then transformed into physical coordinates (X,Y,Z) in the robotic arm's base coordinate system using the following perspective transformation matrix H:

[0012]

[0013] Where K is the camera intrinsic parameter matrix, and [R|T] is the extrinsic parameter rotation and translation matrix of the camera relative to the end of the robotic arm.

[0014] Preferably, the multi-scale feature fusion network in step S3 includes one main branch and two attention branches. The main branch uses ResNet-50 to extract basic features, and the attention branches process texture features and geometric features respectively.

[0015] The probability distribution for defect classification is calculated using the Softmax function, with the following formula:

[0016]

[0017] Among them, z i Let logit be the score of the i-th type of defect, and C be the total number of defect categories (including no defects).

[0018] Defect depth level D level Quantization employs regression head prediction, combined with depth camera data D raw Calculated using the weighted fusion formula:

[0019]

[0020] Where freg(⋅) is the prediction depth output by the image regression network, and β is the confidence weighting factor, with a value range of [0.6, 0.9].

[0021] The final generated defect feature vector V def =[P(c1),...,P(c C ),D level ,Area], where Area is the area of ​​the defective connected domain.

[0022] Preferably, the specific process of generating the adaptive polishing path based on the symbolic distance field (SDF) in step S4 is as follows:

[0023] Construct a two-dimensional mesh with the center of the defect region as the origin, calculate the shortest distance ϕ(x) from each point in the mesh to the defect boundary, and define the SDF function:

[0024]

[0025] Where Ω represents the defect region and d represents the Euclidean distance;

[0026] The grinding path is planned to expand spirally from the inside to the outside along ∇ϕ(x), and the feed speed vf is inversely proportional to the local curvature k, while the spindle speed n is directly proportional to the defect hardness Hv. The specific control formula is as follows:

[0027]

[0028] Among them, v base The base feed rate, γ is the sensitivity coefficient, and H v(x) represents the estimated Vickers hardness value obtained by looking up the table based on the defect type.

[0029] Preferably, the material removal rate (MRR) model described in step S5 is based on the improved Preston equation, taking into account the wear factor of the grinding head, and the formula is as follows:

[0030]

[0031] Among them, K p Where P is the Preston coefficient, v is the contact pressure, and v is the contact pressure. r The relative velocity is η(t), and the tool wear correction factor is η(t), which decreases with the grinding time t.

[0032]

[0033] Where λ is the maximum wear coefficient and μ is the wear rate constant;

[0034] The output u(t) of the adaptive PID controller is used to adjust the feed depth d, and its control law is:

[0035]

[0036] Among them, error , The target contact force is calculated based on the material removal rate model.

[0037] PID parameter K p ,K i ,K d Dynamically adjusts with error e(t):

[0038] .

[0039] Preferably, the multimodal vision sensor in step S1 includes a linear scanning camera and a structured light 3D camera, which are synchronously acquired via hardware triggering. Before data fusion, the 3D point cloud needs to be denoised and registered, outliers are removed using statistical filtering, and the current frame point cloud is aligned with the standard fuse CAD model using the ICP algorithm. The objective function is:

[0040]

[0041] Where, p i For the source point cloud (collected data), q i Let R be the target point cloud (CAD model), and R and t be the rotation matrix and translation vector to be solved.

[0042] Preferably, it also includes a digital twin prediction module, which, before the polishing begins, uses the defect feature vector V generated in step S3 to predict the defect feature vector V. defThe input is fed into a pre-trained polishing result prediction network to predict the surface morphology after polishing;

[0043] The prediction network adopts a generative adversarial network (GAN) structure. The generator G takes the defect image and polishing parameters as input and outputs the predicted polished image. The discriminator D judges the authenticity of the image.

[0044] The loss function includes adversarial loss and content loss:

[0045]

[0046] If the prediction results indicate a risk of over-wearing, i.e., the area of ​​missing matrix material in the predicted image is greater than the threshold, then the feed rate v will be automatically reduced. f Reduce by 10%-20% and replan the route.

[0047] Preferably, a thermo-mechanical coupling safety constraint is introduced in step S5 to prevent the fuse temperature from becoming too high during the polishing process, which could lead to annealing or ablation.

[0048] Establish an instantaneous temperature field model:

[0049]

[0050] in, These are the material density, specific heat capacity, and thermal conductivity, respectively. For the rate of frictional heat generation, ;

[0051] When the predicted temperature T is close to the material phase transformation temperature T critical When the temperature drops to a safe level, the cooling air circuit should be forcibly activated and grinding should be paused. The tool should be retracted and the operation should continue.

[0052] Preferably, it also includes a grinding head wear compensation strategy: establishing a grinding head diameter D tool Mapping table LUT between grinding time t and polishing time t wear ;

[0053] Real-time monitoring of current signal I motor , when I motor When the feed current exceeds a certain multiple of the no-load current I0 and the feed depth d remains unchanged, it is determined to be tool wear;

[0054] The compensation algorithm is as follows:

[0055]

[0056]

[0057] Where, d cmd D represents the commanded feed depth. nom Dt is the nominal diameter. ool_estFor LUT wear Estimated current diameter, k w This is the wear compensation gain coefficient.

[0058] Preferably, the entire control process constitutes a closed-loop feedback: during the polishing process, a local image is re-acquired at fixed time intervals Δt, and the current defect residual rate R is calculated. res :

[0059]

[0060] If R res <ϵ (ϵ is the set residual rate threshold) and surface roughness R a If the measured value is less than the target value, the grinding at this station will end; otherwise, continue iteratively executing steps S4 to S5 until the processing requirements are met.

[0061] Compared with the prior art, the present invention has the following beneficial effects:

[0062] 1. This invention achieves sub-millimeter-level fuse positioning accuracy by introducing a perspective transformation matrix and an improved YOLOv8-Nano network. Even with a random deviation of ±13μm in the loading posture, it can accurately calculate the grinding point in the robotic arm coordinate system, thus solving the problems of strong rigidity and poor adaptability of traditional mechanical positioning.

[0063] 2. This invention is not blind grinding, but accurately identifies the type and quantifies the depth of defects through a multi-scale feature fusion network. The path planning algorithm based on the symbolic distance field can automatically generate a spiral grinding path according to the geometry of the defect, and the spindle speed and feed rate can change in real time with the hardness of the defect, realizing flexible processing with one strategy for one defect, which greatly improves the consistency of processing quality.

[0064] 3. This invention introduces a digital twin prediction module and thermo-mechanical coupling safety constraints, which can pre-simulate the processing results before actual grinding. If the AI ​​predicts that there is a risk of over-grinding, the system will automatically reduce the speed, effectively avoiding damage to the expensive fuse substrate. At the same time, the introduction of the temperature field model prevents the material properties of the workpiece from degrading due to overheating.

[0065] 4. By establishing a material removal rate model and a tool wear mapping table, the system can compensate for the decrease in cutting force caused by grinding wheel wear, maintain a constant removal rate, and, combined with a closed-loop feedback mechanism based on the residual rate, ensure that defects are completely removed while minimizing the amount of material removed and extending the workpiece life.

[0066] 5. In summary, compared with the traditional fixed-parameter grinding method, the present invention improves the defect removal rate of fuses to over 99%, increases the surface roughness Ra stability by 30%, shortens the grinding cycle of a single part by 15%-20%, and significantly reduces the scrap rate and production cost. Attached Figure Description

[0067] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments recorded in this invention. For those skilled in the art, other drawings can be obtained based on these drawings.

[0068] Figure 1 This is a flowchart of a precision grinding control method for fuses that integrates target location and defect classification according to the present invention.

[0069] Figure 2 Example image before end face grinding;

[0070] Figure 3 Example image before side polishing;

[0071] Figure 4 Example image after end face grinding;

[0072] Figure 5 This is an example image after the side has been sanded. Detailed Implementation

[0073] To enable those skilled in the art to better understand the technical solutions of the present invention, the present invention will be further described in detail below with reference to the accompanying drawings. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention.

[0074] Example 1: This example provides a precision grinding control system for fuses that integrates target location and defect classification, and its workflow, specifically including the following implementation steps:

[0075] 1. System hardware platform setup

[0076] The hardware system in this embodiment consists of a six-degree-of-freedom industrial robotic arm, a multimodal vision sensor mounted at the end effector, a force-controlled grinding head, a high-speed electric spindle, and an industrial computer. The multimodal vision sensor includes a 5-megapixel color industrial camera and a line laser structured light 3D profilometer, which achieve microsecond-level synchronous data acquisition via external hardware trigger signals. The grinding head is equipped with a six-dimensional force / torque sensor to provide real-time feedback of the contact force F. real The electric spindle is equipped with a Hall current sensor to monitor the spindle motor current I. motor .

[0077] 2. Data Acquisition and Preprocessing

[0078] The fuses are randomly placed below the vision station via a loading tray. The industrial control computer sends a trigger signal, and the structured light 3D camera projects multi-line laser stripes to acquire 3D point cloud data P of the fuse surface. raw Simultaneously, the RGB camera synchronously acquires high-resolution two-dimensional color images. rgb .

[0079] For the acquired original point cloud P raw First, a statistical outlier removal algorithm is used: for each point, the average distance μ and standard deviation σ of its k=20 nearest neighbors are calculated. If the average distance of a point's neighborhood exceeds μ+2σ, it is identified as an outlier and removed, resulting in a denoised point cloud P. denoised Subsequently, the target point cloud Q generated from the standard CAD model of the fuse is loaded. cad Registration and alignment are performed using the Iterative Closest Point (ICP) algorithm, with the objective function being:

[0080]

[0081] Where, p i For the source point cloud (collected data), q i Let R be the target point cloud (CAD model), and R and t be the rotation matrix and translation vector to be solved. i ∈P denoised q i ∈Q cad After obtaining the rotation matrix R and translation vector t, the denoised point cloud is transformed to the CAD model coordinate system to complete the 3D spatial calibration and provide a reference for subsequent coordinate mapping.

[0082] 3. Target localization and coordinate mapping based on improved YOLOv8-Nano

[0083] RGB image I rgb The input is fed into a pre-trained lightweight convolutional neural network, namely an improved YOLOv8-Nano. The backbone feature extraction network has coordinate attention (CA) modules embedded at its ends to enhance sensitivity to fuse edges and small features. The network's bounding box regression loss function uses an improved loss L10000 combined with IoU. loc The specific expression is as follows:

[0084]

[0085] The weighting coefficient α is set to 0.25. Through model inference, the pixel-level bounding box [x] of the fuse in the image is output. min ,y min ,x max ,y max ], and with the center point of the box (u c , v c () is used as the initial reference point for crawling.

[0086] To achieve the transformation from pixel coordinates to the robot arm's base coordinate system, the camera's intrinsic parameter matrix K and its extrinsic parameter matrix R|T relative to the robot arm's end effector are first obtained through hand-eye calibration. Coordinate mapping is then performed using the perspective transformation matrix H.

[0087]

[0088] Where K is the camera intrinsic parameter matrix, and [R|T] is the extrinsic parameter rotation and translation matrix of the camera relative to the end effector of the robotic arm. The transformed coordinates (X, Y, Z) are directly used for motion planning of the end effector of the robotic arm to generate the initial grinding approach point. In this embodiment, the positioning accuracy has been tested to reach ±0.3mm.

[0089] 4. Multi-scale feature fusion network and defect feature vector generation

[0090] After the robotic arm moves the fuse to the grinding station, the vision system performs a second high-resolution image capture of the area within the positioning frame to obtain a local defect image. def .

[0091] A multi-scale feature fusion network is constructed for defect analysis. The main network structure is as follows:

[0092] Main branch: A ResNet-50 pre-trained on ImageNet is used as the backbone network. The last fully connected layer is removed, retaining only the feature map F output up to conv5_x. base The size is 1 / 32 of the input image.

[0093] Texture Attention Branch: A Local Binary Pattern (LBP) guided attention module is connected after the output layers of ResNet-50's conv3_x and conv4_x to generate texture enhancement feature maps F. tex .

[0094] Geometric attention branch: Utilizing local point cloud depth maps D acquired synchronously by a structured light camera raw Calculate the surface normal vector change map and encode it as a feature map F. geo .

[0095] F base F tex and F geoAfter concatenation along the channel dimension, the data is fed into a fully connected layer and two output headers:

[0096] (1) Classification Header: Outputs a C=5-dimensional logit vector z, with preset defect categories: no defects, burrs, oxide scale, pores, and scratches. The probability P(c) of each category is calculated using the Softmax function. i For example, for a certain pore defect, the probability distribution output by the classification head is 0.02, 0.05, 0.03, 0.87, 0.03.

[0097] (2) Regression head: Outputs the predicted depth value f reg (I def ). Set the confidence weighting factor β=0.8, and combine it with the raw depth camera data D. raw (Taking the median of the defect region), the final defect depth level D is obtained by weighted fusion. level For example, f reg Predicted to be 1.2mm, D raw If it is 1.0mm, then D level =0.8×1.2 + (1-0.8)×1.0 = 1.16mm, quantized to a precision level of 0.1mm.

[0098] Simultaneously, connected component analysis is performed on the defect regions with the highest classification confidence, and the area (in mm²) is calculated. Finally, a defect feature vector V is generated. def :

[0099] V def = 0.02,0.05,0.03,0.87,0.03,1.16,15.4.

[0100] 5. Adaptive Grinding Path Planning Based on Symbolic Distance Field (SDF)

[0101] With the geometric center of the defective connected domain (x) c , y c Using ∇ϕ(x) as the origin, construct a two-dimensional grid Ω with a spacing of 0.5 mm on the grinding plane. For any point x within the grid, calculate the shortest Euclidean distance from x to the boundary ∂Ω of the defect region, generating an SDF function ϕ(x) (negative inside the defect and positive outside). Define the grinding path as a series of equidistant lines spiraling outward along the gradient direction of ∇ϕ(x).

[0102] In this embodiment, the basic feed rate v base =20 mm / s, sensitivity coefficient γ=0.8. From the table, the estimated Vickers hardness value H for porosity defects is obtained. v ≈80 HV, spindle base speed n base =15000 rpm. Therefore, the process parameters at path point x are calculated as follows:

[0103] feed rate It automatically slows down in areas with greater curvature.

[0104] spindle speed .

[0105] The generated grinding path is sent to the robotic arm controller in G-code format.

[0106] 6. Force control and adaptive PID adjustment during the grinding process

[0107] Start the grinding operation, with a preset initial feed depth d0 = 0.3mm. A six-dimensional force sensor collects the normal contact force F in real time. real An improved Preston Material Removal Rate (MRR) model is used to predict the removal amount and calculate the target contact force. The Preston coefficient K is taken as... p =2.5×10 -6 mm / (N·m / min), initial tool wear correction factor η(0)=1, wear parameter λ=0.15, μ=0.01 s -1 Set the desired removal rate to MRR. target =0.8 mm 3 / s, back-calculate target contact force F target ≈12 N.

[0108] The input error of the adaptive PID controller is e(t) = F target -F real The output control quantity u(t) adjusts the feed depth d. The PID parameters are dynamically adjusted in segments according to the error magnitude.

[0109] When |e|>5 N, (K p , K i , K d = (0.08, 0.02, 0.005);

[0110] When |e|≤5 N, switch to (0.04, 0.01, 0.002).

[0111] During the grinding process, the current feed depth command value is calculated in real time: dcmd(t) = d cmd (t -1 ) + u(t).

[0112] 7. Grinding head wear compensation and thermo-coupling safety constraints

[0113] The system backend maintains a wear mapping table (LUT) based on historical data. wear Record the diameter D of the grinding head. toolThe relationship between the cumulative grinding time t and the actual operation. The spindle motor current I is monitored in real time. motor The no-load current I0 = 1.2 A. When it is detected that at the same feed depth d, I motor When the value consistently exceeds 2.5 × I0 (i.e., 3.0 A) without showing a decreasing trend, the tool is considered to have entered the wear stage. The compensation algorithm is then activated: from the LUT... wear Query the current estimated diameter D tool_est Assuming the nominal diameter D nom =6.0 mm, wear compensation gain k w =1.2, then the compensated feed depth d comp = d cmd × (6.0 / Dtool_est) 1.2 This is to maintain a constant actual cutting depth.

[0114] Meanwhile, to prevent the fuse material (such as epoxy resin or thermosetting plastic) from softening or ablation due to frictional heat, a thermo-coupled safety constraint module is activated. The transient heat conduction equation is solved in real time using the finite difference method, where the frictional heat generation rate q... f = μ f ·P·v r friction coefficient μ f Take 0.45. When the predicted grinding point temperature T reaches the material phase transformation temperature T... critical When the temperature reaches 90% of 120°C (i.e., 108°C), the industrial control computer immediately triggers a digital output signal to open the compressed air cooling valve and blow air onto the grinding area. At the same time, the robotic arm performs a tool-lifting action to pause the grinding process. The processing resumes once the infrared temperature sensor detects that the temperature has dropped below 80°C.

[0115] 8. Digital Twin Prediction and Overwear Prevention

[0116] Before the actual polishing begins, the defect feature vector V is... def , Proposed grinding parameters (v f A pre-trained generative adversarial network (GAN) is constructed using the inputs of (n) and local images of defects. The generator G, based on a Pix2Pix architecture, outputs a predicted image of the polished surface topography I. pred Using a trained semantic segmentation model to analyze I pred The system analyzes and calculates the ratio of the area of ​​missing pixels in the substrate material to the total area. If this ratio exceeds a preset threshold of 5%, it is determined that there is a risk of over-wearing. The system automatically adjusts the feed speed v. f Reduce the speed by 15% (e.g., from 20 mm / s to 17 mm / s) and adjust the SDF path spacing accordingly, then regenerate the grinding program to prevent workpiece damage at the source.

[0117] 9. Closed-loop iteration and processing termination conditions

[0118] The grinding process is not completed一次性. Every Δt = 5 seconds, the robotic arm moves the workpiece back to the vision inspection station, repeats the local image acquisition and defect analysis in step S3, calculates the current defect residue rate R res (current defect area / original defect area × 100%), and measures the Ra value of the grinding area with a contact roughness meter. If R res < ϵ = 2% and R a < 1.6 μm, it is determined that the grinding at this station is qualified, and the robotic arm places the workpiece in the blanking area; otherwise, update the defect feature vector V def ', and repeat steps S4 to S5 to perform secondary fine grinding. This closed-loop control strategy ensures that even in the face of unknown material fluctuations or defect depth prediction deviations, the processing quality requirements can be finally achieved through iteration.

[0119] Through the method described in this embodiment, in the actual production line test, for a certain type of fuse (the material is DMC unsaturated polyester bulk molding compound), the one-time defect removal rate reaches 99.2%, the average grinding cycle per piece is shortened from 85 seconds of the traditional method to 68 seconds, and the surface roughness R a The mean value is improved from 1.8 μm to 1.2 μm and the consistency is significantly improved.

[0120] Only some exemplary embodiments of the present invention have been described by way of illustration above. Without doubt, for those of ordinary skill in the art, the described embodiments can be modified in various different ways without departing from the spirit and scope of the present invention. Therefore, the above drawings and description are illustrative in nature and should not be construed as limiting the scope of the claims of the present invention.

Claims

1. A precision grinding control method for fuses that integrates target location and defect classification, characterized in that, Includes the following steps: S1: Acquire 3D point cloud data and high-resolution RGB images of the fuse using a multimodal vision sensor; S2: Based on a lightweight convolutional neural network, target localization is performed on the RGB image, the pixel coordinates of the fuse are output, and the initial gripping and polishing positions are calculated by mapping the perspective transformation matrix to the robotic arm coordinate system. S3: Construct a multi-scale feature fusion network to extract and classify features from the located defect region, identify the defect type and quantify the defect depth level, and generate a defect feature vector V. def ; S4: Based on the defect feature vector V def Based on the initial grinding position, an adaptive grinding path is generated using the Signed Distance Field (SDF) algorithm, and the feed rate v of the grinding head is planned. f With spindle speed n; S5: Start the grinding operation, using a force sensor to collect the grinding contact force F in real time. real Based on the preset material removal rate model, the feed depth d of the grinding head is dynamically adjusted by an adaptive PID controller until the defects are removed and the surface roughness meets the standard.

2. The precision grinding control method for fuses integrating target positioning and defect classification according to claim 1, characterized in that: The lightweight convolutional neural network described in step S2 adopts an improved YOLOv8-Nano architecture and introduces a coordinate attention (CA) mechanism, whose loss function for localizing coordinates is L. loc The bounding box regression loss, which incorporates IoU, is formulated as follows: Where IoU is the intersection-over-union ratio of the predicted bounding box and the ground truth bounding box, ρ is the Euclidean distance between the center points of the two boxes, and c is the minimum diagonal distance between the two bounding boxes. For bounding box parameters, Here are the actual parameters, and α is the weighting coefficient. By minimizing L loc Obtain the optimal bounding box [x min ,y min ,x max ,y max The pixel coordinates (u,v) are then transformed into physical coordinates (X,Y,Z) in the robotic arm's base coordinate system using the following perspective transformation matrix H: Where K is the camera intrinsic parameter matrix, and [R|T] is the extrinsic parameter rotation and translation matrix of the camera relative to the end of the robotic arm.

3. The precision grinding control method for fuses integrating target positioning and defect classification according to claim 1, characterized in that: The multi-scale feature fusion network described in step S3 includes one main branch and two attention branches. The main branch uses ResNet-50 to extract basic features, and the attention branches process texture features and geometric features respectively. The probability distribution for defect classification is calculated using the Softmax function, with the following formula: Among them, z i Let logit be the score of the i-th type of defect, and C be the total number of defect categories (including no defects). Defect depth level D level Quantization employs regression head prediction, combined with depth camera data D raw Calculated using the weighted fusion formula: Where freg(⋅) is the prediction depth output by the image regression network, and β is the confidence weighting factor, with a value range of [0.6, 0.9]. The final generated defect feature vector V def =[P(c1),...,P(c C ),D level ,Area], where Area is the area of ​​the defective connected domain.

4. The precision grinding control method for fuses integrating target positioning and defect classification according to claim 1, characterized in that: The specific process of generating the adaptive grinding path based on the symbolic distance field (SDF) in step S4 is as follows: Construct a two-dimensional mesh with the center of the defect region as the origin, calculate the shortest distance ϕ(x) from each point in the mesh to the defect boundary, and define the SDF function: Where Ω represents the defect region and d represents the Euclidean distance; The grinding path is planned to expand spirally from the inside to the outside along ∇ϕ(x), and the feed speed vf is inversely proportional to the local curvature k, while the spindle speed n is directly proportional to the defect hardness Hv. The specific control formula is as follows: Among them, v base The base feed rate, γ is the sensitivity coefficient, and H v (x) represents the estimated Vickers hardness value obtained by looking up the table based on the defect type.

5. The precision grinding control method for fuses integrating target positioning and defect classification according to claim 1, characterized in that: The Material Removal Rate (MRR) model described in step S5 is based on the improved Preston equation, taking into account the wear factor of the grinding head. The formula is as follows: Among them, K p Where P is the Preston coefficient, v is the contact pressure, and v is the contact pressure. r The relative velocity is η(t), and the tool wear correction factor is η(t), which decreases with the grinding time t. Where λ is the maximum wear coefficient and μ is the wear rate constant; The output u(t) of the adaptive PID controller is used to adjust the feed depth d, and its control law is: Among them, error , The target contact force is calculated based on the material removal rate model. PID parameter K p ,K i ,K d Dynamically adjusts with error e(t): 。 6. The precision grinding control method for fuses integrating target positioning and defect classification according to claim 1, characterized in that: The multimodal vision sensor in step S1 includes a linear scanning camera and a structured light 3D camera, which are synchronously acquired via hardware triggering. Before data fusion, the 3D point cloud needs to be denoised and registered. Statistical filtering is used to remove outliers, and the ICP algorithm is used to align the current frame point cloud with the standard fuse CAD model. The objective function is: Where, p i For the source point cloud, q i Let R be the target point cloud, and R and t be the rotation matrix and translation vector to be solved.

7. The precision grinding control method for fuses integrating target positioning and defect classification according to claim 1, characterized in that: It also includes a digital twin prediction module, which, before the polishing begins, uses the defect feature vector V generated in step S3. def The input is fed into a pre-trained polishing result prediction network to predict the surface morphology after polishing; The prediction network adopts a generative adversarial network (GAN) structure. The generator G takes the defect image and polishing parameters as input and outputs the predicted polished image. The discriminator D judges the authenticity of the image. The loss function includes adversarial loss and content loss: If the prediction results indicate a risk of over-wearing, i.e., the area of ​​missing matrix material in the predicted image is greater than the threshold, then the feed rate v will be automatically reduced. f Reduce by 10%-20% and replan the route.

8. The precision grinding control method for fuses integrating target positioning and defect classification according to claim 1, characterized in that: In step S5, a thermo-mechanical coupling safety constraint is introduced to prevent the fuse from overheating during the grinding process, which could lead to annealing or ablation. Establish an instantaneous temperature field model: in, These are the material density, specific heat capacity, and thermal conductivity, respectively. For the rate of frictional heat generation, ; When the predicted temperature T is close to the material phase transformation temperature T critical When the temperature drops to a safe level, the cooling air circuit should be forcibly activated and grinding should be paused. The tool should be retracted and the operation should continue.

9. The precision grinding control method for fuses integrating target positioning and defect classification according to claim 1, characterized in that: It also includes a grinding head wear compensation strategy: establishing the grinding head diameter D tool Mapping table LUT between grinding time t and polishing time t wear ; Real-time monitoring of current signal I motor , when I motor When the feed current exceeds a certain multiple of the no-load current I0 and the feed depth d remains unchanged, it is determined to be tool wear; The compensation algorithm is as follows: Where, d cmd D is the commanded feed depth. nom Dt is the nominal diameter. ool_est For LUT wear Estimated current diameter, k w This is the wear compensation gain coefficient.

10. A precision grinding control method for fuses integrating target positioning and defect classification according to any one of claims 1-9, characterized in that: The entire control process constitutes a closed-loop feedback: during the polishing process, a local image is re-acquired at fixed time intervals Δt, and the current defect residual rate R is calculated. res : If R res <ϵ (ϵ is the set residual rate threshold) and surface roughness R a If the measured value is less than the target value, the grinding at this station will end; otherwise, continue iteratively executing steps S4 to S5 until the processing requirements are met.

Images