Optimization Method for Polishing Path of Cladding Layer on Inner Surface of Cylindrical Barrel Based on Machine Learning

Abstract

A method for optimizing the polishing path of the inner surface cladding layer of a cylindrical barrel based on machine learning, belonging to the technical field of cladding layer polishing. The steps are as follows: First, key operating parameters are collected through sensors installed on the polishing equipment, including polishing pressure, speed, and the contact angle between the polishing head and the material surface. Subsequently, the obtained data is preprocessed to filter out noise and standardize it, preparing for the training of the subsequent deep learning model. The deep learning model adopted in the present invention combines a convolutional neural network (CNN) and a recurrent neural network (RNN), and can automatically optimize the polishing path according to the processed data. During the polishing process, the system can adjust the polishing parameters and path in real time to ensure that the predetermined polishing quality standard is achieved. The present invention can complete the polishing task efficiently and with high quality, is applicable to the polishing of the inner surface cladding layer of workpieces such as the servomotor of a hydraulic turbine, and has important practical application value.

Description

Technical Field

[0001] The present invention belongs to the technical field of the inner surface treatment of a water turbine servomotor, and particularly relates to an optimization method for the polishing path of a cladding layer on the inner surface of a cylindrical barrel based on machine learning. Background Art

[0002] In the field of traditional water turbine servomotor manufacturing, polishing is an important surface treatment process. Traditional polishing methods are mainly divided into two categories: manual polishing and simple automated polishing.

[0003] Manual polishing: Manual polishing relies on the skills and experience of technicians. By using hand-held tools, the surface of the workpiece is ground and polished. Polishing efficiency: It takes about 3 hours to polish per square meter manually, and the average surface roughness Ra value is about 0.8 microns. A simple automated polishing machine polishes the workpiece by controlling the polishing head through a preset program. Its surface roughness Ra value can generally reach below 0.8 microns. It takes about 2 hours per square meter, and the average Ra value is 0.6 microns.

[0004] As the core equipment of a hydropower station, the inner surface quality of the servomotor of a water turbine directly affects the efficiency and service life of the entire equipment. Traditional cladding layer polishing technology mainly relies on manual operation by experienced technicians. This is not only inefficient but also the result quality is easily affected by human factors, making it difficult to ensure consistency and repeatability. In addition, for complex inner surface shapes, it is even more difficult to achieve the desired effect with manual polishing.

[0005] In recent years, with the development of artificial intelligence technology, deep learning has shown powerful capabilities in multiple fields such as image recognition, speech processing, and natural language processing. By establishing complex neural network models, it can learn patterns and features in a large amount of data, thereby achieving optimization for specific tasks. In the field of industrial polishing, the application of deep learning technology has begun to receive attention. Especially in the optimization of the polishing path, by automatically analyzing the 3D model of the polishing object through deep learning algorithms, the areas that need to be polished can be effectively identified, and the most optimized polishing path can be calculated, thereby achieving efficient and high-quality polishing effects. For example, Zhao-sheng Li et al. proposed an offline programming system to generate the tool path for robot polishing. This system can generate precise and uniform paths for free-form surfaces, which is applicable to the polishing of turbine blade surfaces and propeller blade surfaces, demonstrating the potential of deep learning in improving polishing efficiency and quality.

[0006] Therefore, the technical problem to be solved by the present invention is: Based on a deep learning optimization model, how to develop an optimization method for the polishing path to solve the problems of low traditional polishing efficiency, low polishing quality, and high labor costs. Summary of the Invention

[0007] The technical problem to be solved by the present invention is to provide an optimization method for the polishing path of the cladding layer on the inner surface of a cylindrical barrel based on machine learning. By optimizing the polishing path, the efficiency of the polishing process and the surface treatment quality are improved, while the need for manual operation and related costs are reduced.

[0008] To solve the above technical problems, the technical solution adopted by the present invention is:

[0009] An optimization method for the polishing path of the cladding layer on the inner surface of a cylindrical barrel based on machine learning, comprising the following steps:

[0010] Step 1: Collect the geometric data of the inner surface of the cylindrical barrel through a 3D scanner:

[0011] Step 1.1: Fix the cylindrical barrel on a rotating platform;

[0012] Step 1.2: Adjust the scanning parameters of the 3D scanner according to the size and complexity of the cylindrical barrel. The scanning parameters include scanning speed, laser intensity, and resolution; then use the 3D scanner to collect geometric data;

[0013] Step 1.3: Save the geometric data as raw point cloud data in point cloud format. Each point contains the coordinates (X, Y, Z) in space and the surface reflection intensity of the point;

[0014] Step 1.4: Correct the raw point cloud data to eliminate the deviation caused by scanning angle and equipment error; merge the raw point cloud data scanned from multiple angles to ensure the coherence of the overall geometric data;

[0015] Step 1.5: Perform noise reduction and simplification processing on the corrected raw point cloud data to remove irrelevant information;

[0016] Step 2: Convert the collected raw point cloud data into a format suitable for analysis by a deep learning model, and use the deep learning model to process the geometric data to identify the key areas that need to be polished:

[0017] Step 2.1: Data cleaning: Preliminarily process the collected 3D scan data, remove incomplete or damaged data points, and ensure data integrity;

[0018] Step 2.2: Denoising: Use a filter to smooth the raw point cloud data and remove the noise generated during the scanning process;

[0019] Step 2.3: Feature extraction: Based on 3D geometric analysis methods, extract key feature information from the processed geometric data. The key feature information includes the curvature, edges, and geometric shape of the surface;

[0020] Step 2.4, Data Standardization: Using the Z-score (standard score method), convert the geometric data into a format with zero mean and unit variance;

[0021] Step 2.5, Data Format Conversion: Convert the processed point cloud data into a voxel grid, discretize the three-dimensional space, and provide a standardized input form for the deep learning model;

[0022] Step 3, Use the supervised learning method to train the deep learning model, and use the stochastic gradient descent optimization algorithm to optimize the loss function. The stochastic gradient descent optimization algorithm is used to optimize the loss function, which is used to measure the difference between the polished path predicted by the deep learning model and the actual optimal path; To avoid overfitting, introduce regularization terms and dropout (dropout method). The deep learning model is iteratively trained on multiple batches of data until the value of the loss function converges;

[0023] Step 4, The trained deep learning model is used to real-time identify a certain area on the inner surface of the cylindrical barrel and generate a polishing path. The deep learning model is used to identify the area that needs to be polished and possible surface defects. The surface defects include depressions and protrusions; Based on the recognition results, the deep learning model further calculates the optimal polishing path covering all recognized areas; The optimal polishing path includes parameters such as path shape, order, and polishing head speed.

[0024] Preferably, in Step 1, the scanning parameters further include the scanning interval. The scanning parameters are adjusted according to the material properties, dimensions, and complexity of the cladding layer of the inner surface of the cylindrical barrel. During the scanning process, multiple angles and multiple positions are used for scanning to capture all visible areas on the inner surface of the cylindrical barrel, ensure the complete coverage of the geometric data, avoid dead corners, and obtain complete three-dimensional information.

[0025] Preferably, in Step 1, the formula (1) is used to calculate the nearest distance d min (P i , P j ) between points in the point cloud data to assist subsequent data processing and path optimization:

[0026]

[0027] where P i and P j are any two points in the point cloud data, and (x i , y i , z i ) and (x j , y j , z j ) are their respective coordinates.

[0028] Preferably, in step 2.1, since irrelevant data points will be generated during the scanning process, the irrelevant data points include noise caused by environmental reflection or equipment error. Through threshold segmentation and statistical analysis methods, the standard deviation and mean are used to identify and remove outliers.

[0029] Preferably, in step 2.2, the Gaussian filtering algorithm is used to denoise the data to smooth the data and reduce the influence caused by scanning error. The formula of the Gaussian filtering algorithm is as follows:

[0030]

[0031] where F represents the original scanned data, F filtered represents the filtered data, W is the weight matrix of the filter, (x, y) is the coordinate of the data point, and (i, j) is the position in the weight matrix;

[0032] Then, the original point cloud data is preliminarily processed using formula (3) to eliminate noise and smooth the data:

[0033]

[0034] where P smooth represents the smoothed point, and N(P) is the neighborhood set of point P.

[0035] Preferably, in step 2.4, formula (4) is used to normalize the data to ensure that the training of the deep learning model is not affected by different scanning scales:

[0036]

[0037] where P is the original data point, P min and P max are the minimum and maximum values in the dataset respectively.

[0038] Preferably, after step 2.5, boundary detection is performed: for the boundary regions that need to be precisely polished, an edge detection algorithm is used to identify and mark these regions:

[0039]

[0040] where E(x, y) represents the edge detection result, I represents the image intensity, and are the gradients of the image intensity in the x and y directions respectively.

[0041] Preferably, the deep learning model uses a convolutional neural network and a recurrent neural network as the basic architecture to process and identify the spatial hierarchical information in the image data. The deep learning model includes an input layer, a convolutional layer, a pooling layer, and a fully connected layer;

[0042] Among them, the input layer: adapts to the network architecture by accepting the two-dimensional depth map converted from three-dimensional scan data or the processed point cloud data;

[0043] Convolutional layer: Implement multiple convolutional layers. Each layer uses multiple convolutional kernels to extract local features of the data. The formula of the convolutional layer is as follows:

[0044]

[0045] Among them, is the j-th feature response in the i-th feature map of the (l + 1)-th layer, σ represents the activation function, and are the bias and convolutional kernel weights of the l-th layer respectively, and * represents the convolution operation;

[0046] Pooling layer: Introduce the max pooling of the pooling layer to reduce the feature dimension and extract a more abstract feature representation;

[0047] Fully connected layer: Apply a fully connected layer at the backend of the network to map the distribution of the polished area into a form recognizable by the output layer. The formula of the fully connected layer is:

[0048] [O i = σ(b i + ∑ j W ij H j )] (7);

[0049] Among them, O i is the i-th neuron of the output layer, H j is the output of the previous layer, W ij and b i represent the weight and bias respectively, and σ is the activation function;

[0050] The deep learning model uses the cross-entropy loss function to evaluate the model performance:

[0051]

[0052] Among them, M is the number of classification categories, y o,c is the actual label, and p o,c is the probability predicted by the model.

[0053] Preferably, in step 3, the preprocessed data set is divided into a training set, a validation set, and a test set. The training process is monitored and the model hyperparameters are adjusted by evaluating the model performance on the validation set. The model is trained using the batch gradient descent method, and the weights and biases are iteratively updated:

[0054]

[0055] Among them, θ represents the model parameters, α is the learning rate, is the gradient of the loss function with respect to the parameters.

[0056] The present invention can achieve the following beneficial effects:

[0057] 1. By optimizing the polishing path, the efficiency of the polishing process and the surface treatment quality are improved, while the need for manual operation and related costs are reduced, which is applicable to the fields of hydropower station and water turbine manufacturing and maintenance.

[0058] 2. It can not only automatically identify and analyze the complex structure of the inner surface of the water turbine servo cylinder, but also generate an efficient and accurate polishing path to improve the polishing efficiency, reduce the labor cost, and ensure the consistency and repeatability of the polishing quality.

[0059] 3. By applying deep learning algorithms to optimize the polishing path of the cladding layer on the inner surface of the water turbine servo cylinder, the polishing efficiency and surface quality are improved. Traditional polishing methods rely on manual operation or simple automation programs, which are not only inefficient but also difficult to ensure the polishing quality. To address this issue, the present invention adopts advanced deep learning technologies to automatically identify the characteristics of the cladding layer, calculate the optimized polishing path, and achieve high-precision and high-efficiency polishing operations. BRIEF DESCRIPTION OF THE DRAWINGS

[0060] The present invention will be further described below in conjunction with the drawings and embodiments:

[0061] Figure 1 : Overall system architecture diagram.

[0062] Figure 2 : Deep learning model architecture diagram.

[0063] Figure 3 : Flowchart of the polishing path planning algorithm.

[0064] Figure 4 : Schematic diagram of the deep learning training process.

[0065] Figure 5 : Diagram of the optimization algorithm evaluation criteria.

[0066] Figure 6 : Diagram of the polishing effect evaluation index.

[0067] Figure 7 : Flowchart of error analysis and adjustment. DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS

[0068] The preferred solution is as Figures 1 to 7 shown, a method for optimizing the polishing path of the cladding layer on the inner surface of a cylindrical barrel based on machine learning, Figure 1Used to display the overall framework of the technology for the polishing path of the inner surface of the servomotor of a hydraulic turbine optimized based on deep learning, including the main modules and data flow directions. Figure 2 Used to display the deep neural network architecture for polishing area recognition and path generation, including the configuration and functions of each layer. Figure 3 Used to elaborate on the calculation method and optimization strategy for automatically generating the polishing path according to the output of the deep learning model. Figure 4 Used to illustrate how to train the deep learning model using a large amount of data, including the preparation of the training set, the model training, and the validation process. Figure 5 Used to describe how to evaluate the performance of the polishing path optimization algorithm, including criteria such as efficiency, coverage rate, and surface quality. Figure 6 Used to display the quality comparison of the inner surface of the servomotor before and after polishing, including the quantitative evaluation of parameters such as surface roughness and surface shape accuracy. Figure 7 Used to display how the system automatically detects and corrects the deviation during the polishing process to ensure the consistency of the polishing effect.

[0069] The specific steps are as follows:

[0070] Step 1: Collect the geometric data of the inner surface of the cylindrical barrel through a 3D scanner:

[0071] Step 1.1: Fix the cylindrical barrel on the rotating platform;

[0072] Step 1.2: Adjust the scanning parameters of the 3D scanner according to the size and complexity of the cylindrical barrel. The scanning parameters include scanning speed, laser intensity, and resolution;

[0073] The scanning parameters also include the scanning interval. The scanning parameters are adjusted according to the material properties, size, and complexity of the cladding layer on the inner surface of the cylindrical barrel. During the scanning process, multiple angles and multiple positions are used for scanning to capture all visible areas of the inner surface of the cylindrical barrel and ensure the complete coverage of the geometric data, avoiding dead angles, in order to obtain complete 3D information.

[0074] In Step 1, the formula (1) is used to calculate the closest distance between points in the point cloud data to assist subsequent data processing and path optimization:

[0075]

[0076] where P i and P j are any two points in the point cloud data, and (x i , y i , z i ) and (x j , y j , z j ) are their respective coordinates.

[0077] Step 1.3: Save the geometric data as the original point cloud data in point cloud format. Each point contains the coordinates (X, Y, Z) in space and the surface reflection intensity of the point.

[0078] Step 1.4: Calibrate the original point cloud data to eliminate the deviations caused by scanning angles and equipment errors; merge the original point cloud data scanned from multiple angles to ensure the coherence and consistency of the overall geometric information.

[0079] Step 1.5: Perform noise reduction and simplification processing on the calibrated original point cloud data to remove irrelevant information.

[0080] Step 2: Convert the collected original point cloud data into a format suitable for analysis by a deep learning model, and use the deep learning model to process the geometric data to identify the key areas that need to be polished.

[0081] Step 2.1: Data cleaning: Preliminarily process the collected 3D scan data, remove incomplete or damaged data points to ensure data integrity.

[0082] In Step 2.1, since irrelevant data points will be generated during the scanning process, the irrelevant data points include the noise caused by environmental reflection or equipment errors. Through threshold segmentation and statistical analysis methods, the standard deviation and mean are used to identify and remove outliers.

[0083] Step 2.2: Denoising: Use the Savitzky-Golay filter to smooth the original point cloud data and remove the noise generated during the scanning process.

[0084] In Step 2.2, the Gaussian filtering algorithm is used to denoise the data to smooth the data and reduce the influence caused by scanning errors. The Gaussian filtering algorithm formula is as follows:

[0085]

[0086] where F represents the original scan data, F filtered represents the filtered data, W is the weight matrix of the filter, (x, y) are the coordinates of the data point, and (i, j) are the positions in the weight matrix.

[0087] Then, the original point cloud data is preliminarily processed using formula (3) to eliminate noise and smooth the data:

[0088]

[0089] where P smooth represents the smoothed point, and N(P) is the neighborhood set of point P.

[0090] Step 2.3, Feature Extraction: Based on the three-dimensional geometric analysis method, extract key feature information from the processed geometric data. The key feature information includes the curvature, edges, and geometric shape of the surface;

[0091] Step 2.4, Data Standardization: Use the Z-score (standard score method) to convert the geometric data into a format with zero mean and unit variance;

[0092] In Step 2.4, formula (4) is used to normalize the data to ensure that the training of the deep learning model is not affected by different scanning scales:

[0093]

[0094] where P is the original data point, P min and P max are the minimum and maximum values in the dataset respectively.

[0095] Step 2.5, Data Format Conversion: Convert the processed point cloud data into a voxel grid, discretize the three-dimensional space, and provide a standardized input form for the deep learning model;

[0096] After Step 2.5, perform boundary detection: For the boundary regions that need to be precisely polished, use an edge detection algorithm to identify and mark these regions:

[0097]

[0098] where E(x, y) represents the edge detection result, I represents the image intensity, and are the gradients of the image intensity in the x and y directions respectively.

[0099] Step 3, Use the supervised learning method to train the deep learning model, use the stochastic gradient descent optimization algorithm to optimize the loss function. The stochastic gradient descent optimization algorithm to optimize the loss function is used to measure the difference between the polished path predicted by the deep learning model and the actual optimal path; To avoid overfitting, introduce a regularization term and dropout (dropout method). The deep learning model is iteratively trained on multiple batches of data until the value of the loss function converges; The deep learning model uses a convolutional neural network and a recurrent neural network as the basic architecture to process and identify the spatial hierarchical information in the image data. The deep learning model includes an input layer, a convolutional layer, a pooling layer, and a fully connected layer,

[0100] Among them, the input layer: accepts the two-dimensional depth map converted from the three-dimensional scan data or the processed point cloud data to adapt to the network architecture;

[0101] Convolutional layer: Multiple convolutional layers are implemented. Each layer uses multiple convolutional kernels to extract local features of the data. The formula representation of the convolutional layer is as follows:

[0102]

[0103] Where, is the j-th feature response in the i-th feature map of the (l + 1)-th layer, σ represents the activation function, and are the bias and convolutional kernel weights of the l-th layer respectively, and * represents the convolution operation.

[0104] Pooling layer: Max pooling of the pooling layer is introduced to reduce the feature dimension and extract more abstract feature representations.

[0105] Fully connected layer: The fully connected layer is applied at the backend of the network to map the distribution of the polished area into a form recognizable by the output layer. The formula of the fully connected layer is:

[0106] [O i =σ(b i +∑ j W ij H j )] (7);

[0107] Where, O i is the i-th neuron of the output layer, H j is the output of the previous layer, W ij and b i represent the weight and bias respectively, and σ is the activation function.

[0108] The deep learning model uses the cross-entropy loss function to evaluate the model performance:

[0109]

[0110] Where, M is the number of classification categories, y o,c is the actual label, and p o,c is the probability predicted by the model.

[0111] The preprocessed dataset is divided into a training set, a validation set, and a test set. The training process is monitored and the model hyperparameters are adjusted by evaluating the model performance on the validation set. The model is trained using the batch gradient descent method, and the weights and biases are iteratively updated:

[0112]

[0113] Where, θ represents the model parameters, α is the learning rate, is the gradient of the loss function with respect to the parameters.

[0114] Step 4: The trained deep learning model is used to identify specific areas on the inner surface of the servomotor in real time and generate a polishing path. The deep learning model is used to identify areas that need to be polished and possible surface defects, where the surface defects include depressions and protrusions. Based on the recognition results, the deep learning model further calculates the optimal polishing path covering all recognized areas. The optimal polishing path includes parameters such as path shape, sequence, and polishing head speed.

[0115] Example 1:

[0116] In this example, the cylindrical barrel is the cylinder body of the turbine servomotor. The present invention uses a deep convolutional network (DCN) to extract the features of the cladding layer on the inner surface of the turbine servomotor cylinder body and trains a network model for specific polishing requirements.

[0117] Network architecture design: A deep convolutional network with a multi-level structure is designed, and each layer is designed to capture different levels of polishing-related features.

[0118] Input layer configuration: The input layer is configured to receive the converted two-dimensional depth image, and each pixel value represents the distance of the point on the servomotor surface from the three-dimensional scanning device.

[0119] Data preprocessing: Perform a normalization operation on the input image, scale the pixel values to between 0 and 1, and the formula is expressed as:

[0120]

[0121] where, I represents the original image, I′ represents the normalized image, I min and I max represent the minimum and maximum pixel values in the image, respectively.

[0122] Convolution layer definition: Define multiple convolution layers to extract features. The first convolution layer captures basic edge and contour information, and subsequent layers gradually build more complex patterns.

[0123] Activation function application: Apply the ReLU activation function after each convolution operation to introduce non-linearity and enhance the expression ability of the network. The formula is as follows:

[0124] [f(x) = max(0, x)];

[0125] Pooling layer introduction: Use the max pooling layer to reduce the spatial size of the feature map, retain the most important feature information, and reduce the subsequent calculation amount.

[0126] Batch normalization layer: Add a batch normalization layer between the convolution layer and the activation function to stabilize the training process and accelerate the convergence speed.

[0127] Fully-connected layer construction: A fully-connected layer is constructed at the end of the network to map high-level features to the final output for predicting the areas that need to be polished.

[0128] Loss function design: A cross-entropy loss function is designed to evaluate the gap between the predicted output and the true label, and optimize the model.

[0129]

[0130] where M is the total number of classes, y o,c is the true label, is the predicted probability.

[0131] Optimizer selection: The Adam (Adaptive Moment Estimation) optimizer is selected to adjust the network weights, which has advantages such as adaptive learning rate, momentum consideration, and bias correction.

[0132] Regularization technique: The L2 regularization technique is adopted to reduce model overfitting and improve generalization performance.

[0133]

[0134] where Θ represents the model parameters, α is the learning rate, is the gradient of the loss function, and λ is the regularization parameter.

[0135] Data augmentation strategy: The robustness of the model for detecting polished areas is improved through data augmentation strategies of random rotation and translation.

[0136] Verification and testing: The performance of the model is tested on an independent validation set to ensure that it accurately identifies the areas on the inner surface of the servomotor that need to be polished.

[0137] Optimized polishing path algorithm design: The present invention proposes a polishing path optimization algorithm based on the prediction results of a deep learning model, and fully considers the specific requirements of the polishing process.

[0138] Path generation initialization: Initialize the path planning algorithm, use the areas predicted by the deep learning model that need to be polished as the input, and set the initial polishing path.

[0139] Path planning criterion: Define the objective function of path planning, aiming to minimize the polishing time and maximize the polishing quality.

[0140] [J(p) = αT(p) + βQ(p)];

[0141] where J(p) is the objective function, T(p) represents the predicted time under the polishing path p, Q(p) represents the predicted polishing quality, and α and β are weight factors.

[0142] Spatial Constraint Consideration: Set the spatial constraint conditions for the polishing path to ensure that the end effector of the robot does not collide with other parts of the servomotor.

[0143] Optimization Algorithm Selection: Select the genetic algorithm (GA) heuristic algorithm to iteratively optimize the polishing path.

[0144] Fitness Function Definition: Design a fitness function to evaluate the efficiency and quality of the polishing path.

[0145]

[0146] Polishing Parameter Adjustment: Introduce an adaptive mechanism to adjust the polishing parameters according to the actual polishing effect and optimize the path.

[0147] Path Iterative Optimization: Use an iterative algorithm to optimize the polishing path, and search for a higher fitness function value in each iteration.

[0148] Collision Detection: Implement a collision detection algorithm to ensure that the generated path avoids any potential mechanical interference.

[0149] Path Smoothing: Smooth the optimized path through curve fitting or other smoothing techniques to reduce the vibration of the polishing robot.

[0150] User Interface Integration: Develop a user interface for the operator to set and adjust the path optimization parameters in real time.

[0151] Multi-objective Optimization: Consider multi-objective optimization to simultaneously meet the requirements of polishing speed, cost, and polishing quality.

[0152] Integration of Polishing Process Parameters: Integrate the pressure and moving speed in the polishing process parameters into the path planning.

[0153] Real-time Feedback Loop: Establish a real-time feedback system to use the data collected during the polishing process for real-time adjustment of the polishing path.

[0154] Dynamic Update of Path Optimization: When the polishing robot is running, dynamically update the path according to the sensor feedback.

[0155] Conversion of the Optimized Polishing Path: This invention describes how to convert the optimized polishing path generated by the deep learning model into the motion instructions of the robot to control the precise polishing process.

[0156] Path Parsing: Parse the optimized path data and decompose it into a motion sequence that the robot can understand.

[0157] Coordinate System Conversion: Implement the conversion from the polishing path coordinate system to the robot body coordinate system.

[0158]

[0159] Among them, P robot is a point in the robot coordinate system, is the transformation matrix from the world coordinate system to the robot coordinate system, and P path is a point on the path.

[0160] Motion planning: Develop motion planning algorithms to plan the optimal motion trajectory from one polishing point to another.

[0161] Dynamic interpolation: Implement dynamic interpolation technology for generating smooth robot motion trajectories.

[0162] Velocity control: Adjust the motion speed of the robot according to the polishing process requirements.

[0163] v(t) = v0 + at;

[0164] Among them, v(t) is the velocity at time t, v0 is the initial velocity, and a is the acceleration.

[0165] Acceleration control: Set an appropriate acceleration to avoid sudden acceleration or deceleration of the robot's actions and protect the life of the robotic arm.

[0166] Tool tip control: Calculate and control the attitude of the tool tip end effector to ensure that it always maintains the correct angle with the surface during polishing.

[0167] Error compensation: Perform real-time error compensation through sensor data to ensure the accuracy of path execution.

[0168] Force control: Apply force control strategies to adjust the polishing pressure according to surface feedback during polishing.

[0169] F applied = k(F desired - F measured );

[0170] Among them, F applied is the applied force, F desired is the desired force, F measured is the measured force, and k is the control gain.

[0171] Vision system integration: Integrate a vision system to verify the execution of the polishing path and make real-time adjustments.

[0172] Polishing quality inspection: After polishing, use a vision sensor to detect and evaluate the quality of the polished surface to ensure that it meets the predetermined standards.

[0173] Recording and analysis: Record the polishing process data and perform subsequent analysis to optimize future polishing path planning.

[0174] Algorithm iterative update: Based on the feedback results of polishing quality, the path planning algorithm is iteratively updated to continuously improve the quality and efficiency of the polishing path.

[0175] Real-time monitoring of polishing effect: This embodiment provides a real-time monitoring system for tracking and evaluating the effect during the polishing process of the turbine servomotor and dynamically adjusting the polishing path according to the monitored data. Configuration is as follows:

[0176] Monitoring system setup: Configure multiple sensors, including force sensors, vibration sensors, and vision sensors, to collect relevant data during the polishing process.

[0177] Data acquisition: Real-time collect data on the contact pressure, speed between the polishing head and the surface of the servomotor, and the temperature of the polishing area.

[0178] Data analysis: Apply data analysis algorithms to the collected data to determine whether the polishing process is carried out according to the established parameters.

[0179] R actual = f(Data force , Data vibration , Data temperature );

[0180] Where R actual represents the actual polishing result, and Data force , Data vibration and Data temperature represent the collected force, vibration, and temperature data respectively.

[0181] Effect evaluation: Use machine learning algorithms to evaluate the collected data to determine whether the polishing quality meets the predetermined standards.

[0182] Path adjustment strategy: Design a path adjustment strategy to optimize the motion trajectory of the polishing head to ensure the uniformity and high quality of the polishing effect.

[0183] Feedback control algorithm: Utilize the feedback control algorithm PID control in cybernetics to adjust the polishing path according to real-time data.

[0184]

[0185] Where, P adjust (t) is the path adjustment amount at time t, e(t) is the deviation between the target path and the actual path, and K i , K i and K d are the proportional, integral, and derivative gains respectively.

[0186] Dynamic update mechanism: During the polishing process, the polishing path is dynamically updated according to the feedback data.

[0187] Decision Support Systems: Implement decision support systems to help operators or automated systems decide when and how to adjust paths.

[0188] Experimental Feedback Loops: Set up experimental feedback loops to validate and fine-tune feedback control parameters through laboratory testing or simulations.

[0189] Visualization tools: Develop visualization tools to assist in monitoring polishing effects and intuitively display path adjustment results.

[0190] Consumables management: By monitoring the wear of the polishing head, the service life of the polishing head is automatically calculated and a prompt is provided when replacement is required.

[0191] Adaptive control strategy: The present invention includes an adaptive control strategy that can monitor the polishing process in real time and adjust polishing parameters based on feedback data to ensure that the polishing effect is always in an optimal state.

[0192] Parameter Monitoring: The sensor system is configured to monitor key polishing parameters in real time, including polishing pressure, speed, and contact angle between the polishing head and the material surface.

[0193] Feedback data collection: Real-time collection of sensor data, including polishing pressure, speed, and contact angle between the polishing head and the material surface, provides the required feedback for the control strategy.

[0194] Control model construction: Construct a control model that describes the relationship between polishing parameters and polishing effects. The formula is as follows:

[0195] C eff =g(P force ,V speed ,θ contact );

[0196] Among them, C eff represents the polishing effect, g is the model function, P force , V speed , and θ contact represent polishing pressure, speed and contact angle respectively.

[0197] Real-time control law design: Design real-time control law to adjust polishing parameters to ensure the predetermined polishing quality.

[0198] P new =P current +ΔP(f feedback );

[0199] Among them, P new is the adjusted polishing parameter, Pcurrent is the current parameter, and ΔP is the parameter change based on the feedback function f feedback .

[0200] Adaptive algorithm application: Apply the Adaptive Neuro-Fuzzy Inference System (ANFIS) and Savitzky-Golay filter to optimize the control strategy.

[0201] Fuzzy logic control: Use a fuzzy logic controller to handle the uncertainty and ambiguity of feedback data and improve the robustness of the control strategy.

[0202] Optimization algorithm integration: Integrate multiple optimization algorithms, adopt gradient descent and evolutionary algorithms to achieve the optimal adjustment of polishing parameters.

[0203] PID controller adjustment: Use a Proportional-Integral-Derivative (PID) controller to adjust the movement of the robot in response to real-time feedback.

[0204]

[0205] where e(t) represents the real-time error, K p , K i , K d represent the proportional, integral, and derivative gains respectively.

[0206] Learning rate adjustment: Adjust the learning rate of the model according to real-time feedback to maintain a fast and accurate control response.

[0207] Steady-state error reduction: Design a control strategy to reduce the steady-state error and ensure the consistency of the polishing effect during long-term operation.

[0208] Data-driven adjustment: Implement a data-driven method to adjust polishing parameters and use historical data analysis to guide future control strategies.

[0209] Polishing path adjustment: Update the polishing path in real time in response to the adjustment of polishing parameters to ensure polishing quality and efficiency.

[0210] Polishing pressure feedback control: Adjust the pressure applied by the polishing head according to real-time pressure feedback to adapt to the characteristics of different materials.

[0211] P adjusted = P target - k(P measured - P target );

[0212] where P adjusted is the adjusted pressure, P target is the target pressure, P measured is the measured pressure, and k is the adjustment coefficient.

[0213] Vibration control: Monitor the vibration during the polishing process, adjust the polishing speed and pressure according to the vibration data, and reduce surface defects.

[0214] Performance index development: The key performance indicators (KPIs) are set as the dimensional accuracy and cylindricity of the inner surface of the servomotor to evaluate the efficiency and quality of the polishing process.

[0215] Iterative learning: Apply the iterative learning control (ILC) strategy to gradually improve the control performance in consecutive polishing cycles.

[0216] Online optimization of control parameters: An algorithm is proposed in the present invention, which uses the following formula to online optimize the control parameters of the polishing robot to achieve the optimal polishing effect.

[0217] Definition of control parameters: Determine the set of key control parameters Θ, including the polishing pressure, speed, and the contact angle between the polishing head and the material surface.

[0218] Setting of optimization objective: Define a clear performance objective function J(Θ) for the optimization of control parameters, which quantifies the expected performance of the polishing effect, and set the surface roughness Ra to 0.1 micrometers and the surface shape accuracy cylindricity to 0.10 millimeters.

[0219] Real-time monitoring: Implement a real-time monitoring system to continuously measure the current value Θ of the control parameters current .

[0220] Performance evaluation: The present invention develops a performance evaluation module, which uses the polishing result data to calculate the value of the performance objective function J in real time.

[0221] Gradient calculation: Apply numerical methods to calculate the gradient of the objective function with respect to the control parameters

[0222]

[0223] Adaptive adjustment: Adjust the control parameters according to the gradient information to drive the performance index towards the minimum value of the objective function.

[0224]

[0225] where α is the learning rate, which determines the step size of the adjustment.

[0226] Performance difference feedback: Calculate the performance difference using the current and historical performance data as the feedback for optimization.

[0227] ΔJ = J(Θ current ) - J(Θ old );

[0228] Bias correction: Implement a bias correction mechanism to correct the bias between model predictions and actual performance.

[0229] Robustness enhancement: Enhance the robustness of the control parameter optimization algorithm against unknown disturbances by introducing robust control theory.

[0230] Multi-parameter collaborative optimization: Develop a multi-parameter collaborative optimization strategy to adjust multiple control parameters simultaneously to achieve the best polishing effect.

[0231] Optimization algorithm verification: The present invention verifies the effectiveness of the control parameter optimization algorithm in the laboratory simulation environment of the servomotor cylinder and applies it under actual working conditions.

[0232] Prediction model integration: Integrate the prediction model into the control parameter optimization and use historical data to predict future performance changes.

[0233] Visualization tool application: Adopt the Tableau visualization tool to analyze the dynamic changes of control parameter optimization.

[0234] Iterative path optimization: The present invention describes a method of continuously optimizing the polishing path using an iterative formula to gradually improve the polishing effect and make it more in line with the design goal.

[0235] Initialization of the iterative process: Set an initial polishing path and define the design goal.

[0236] Quantification of the design goal: Determine the specific indicators of polishing quality, surface roughness, and surface shape accuracy.

[0237] Definition of the iterative formula: According to the optimization goal, formulate an iterative formula to adjust the polishing path.

[0238] p new = p old + γ·Δp;

[0239] where p new and p old represent the old and new paths respectively, Δp represents the path adjustment amount, and γ is the step size coefficient.

[0240] Calculation of the path adjustment amount: Calculate the path adjustment amount using the performance evaluation function and feedback data.

[0241]

[0242] where, is the gradient of the performance evaluation function under the current path.

[0243] Setting of the performance evaluation function: Design a performance evaluation function to measure the deviation between the current path and the design goal.

[0244] J(p) = function of surface quality metrics;

[0245] Gradient calculation: Calculate the gradient of the performance evaluation function using numerical methods or analytical methods.

[0246] Polishing effect monitoring: Monitor the polishing effect after each iteration and provide feedback data for path adjustment.

[0247] Step size coefficient adjustment: Adjust the step size coefficient γ according to the change of the polishing effect during the iteration to achieve more effective convergence.

[0248] Convergence determination: Determine whether the iteration converges, that is, judge that the iteration ends when the magnitude of Δp is small enough or the change of the performance evaluation function is no longer significant.

[0249] Termination condition setting: Set the termination condition of the iteration, such as reaching the maximum number of iterations or the performance evaluation index meeting the predetermined threshold.

[0250] Iteration result evaluation: Evaluate the polishing path and polishing effect after each iteration to confirm whether the optimization direction is correct.

[0251] Dynamic adjustment mechanism: A dynamic learning rate adjustment mechanism is established, allowing the path to be adjusted according to the real-time changes of the polishing task during the iteration.

[0252] Automated adjustment process: A fully automated iteration adjustment process is realized, reducing manual intervention and improving the polishing efficiency.

[0253] Algorithm verification: Through the polishing process test of the laboratory servomotor cylinder block, the effectiveness of the iterative algorithm is verified, and it is determined that this method can be actually applied to the polishing process.

[0254] Iterative learning application: Apply iterative learning control (ILC) to continuously improve the path planning in continuous polishing tasks.

[0255] Through a series of embodiments, the present invention can reduce the polishing time by about 40%, and make the inner surface roughness and surface shape accuracy of the servomotor cylinder better than the target performance indicators.

[0256] Experimental design: Design a control experiment, with one group using the optimized polishing path technology of the present invention and the other group using the traditional polishing method.

[0257] Sample selection: Select the same type and size of the water turbine servomotor as the experimental object to ensure the consistency and comparability of the results.

[0258] Polishing time measurement: Use a precise timer to record the polishing times of two groups of experiments, and calculate the average polishing time respectively.

[0259]

[0260]

[0261] Among them, \(t\) new,i and \(t\) trad,i are the polishing times of the \(i\)-th sample using the technology of the present invention and the traditional technology respectively, and \(N\) is the number of samples.

[0262] Surface roughness: Use a surface roughness measuring instrument to test the polished surface, and record the Ra value of the surface roughness index.

[0263] Surface shape accuracy: Use an optical microscope to inspect the polished surface, evaluate the cylindricity of the surface shape accuracy after polishing, and record the data using a quantitative analysis method.

[0264] Data statistical analysis: Conduct statistical analysis on the data of polishing time, surface roughness, and surface shape accuracy, including calculating the mean value, standard deviation, and confidence interval.

[0265]

[0266] Among them, \(X\) is the observed value, \(\mu\) is the average value, \(N\) is the number of samples, and \(\sigma\) is the standard deviation.

[0267] Calculation of efficiency improvement: Calculate the percentage of efficiency improvement of the technology of the present invention compared with the traditional technology in terms of polishing time.

[0268]

[0269] General description of the technical solution: The present invention integrates deep learning technology and automated polishing equipment, providing a revolutionary solution for the polishing work of the inner surface cladding layer of the water turbine servomotor cylinder.

[0270] Development of deep learning model: Develop a dedicated deep learning model that can accurately identify the polishing area, automatically plan the optimal polishing path, and reduce human intervention.

[0271] Data-driven decision-making: Adopt a data-driven method to collect and analyze the polishing process data in real time, and dynamically adjust the polishing parameters according to the analysis results.

[0272] Improvement of polishing efficiency: The test results show that compared with the traditional method, the present invention can reduce the polishing time by about 40%.

[0273] Polishing quality improvement: By precisely controlling the polishing process, this technical solution significantly improves the surface roughness and surface shape accuracy of the inner surface of the servomotor, thus enhancing the product quality.

[0274] Enhanced process reliability: By using adaptive control and real-time feedback adjustment, the stability and reliability of the polishing process are ensured, and the defect rate is reduced.

[0275] Ease of operation: A user-friendly operation interface is designed to simplify the operation process and reduce the technical requirements for operators.

[0276] Example 2:

[0277] Based on the existing test platform and equipment, the present invention designs a system including a Sinsung GCR20-1100 collaborative robot and a customized polishing head, and conducts actual application tests in the polishing task of the cladding layer on the inner surface of the turbine servomotor cylinder. The test results show that through the control execution unit of the present invention, the collaborative robot can accurately execute the polishing path generated by the deep learning model and the path planning algorithm, significantly improving the polishing efficiency and surface quality.

[0278] Step1: Preparation stage

[0279] Before officially starting the polishing operation, preparation work is the key to ensuring the smooth progress of the polishing task and achieving the expected effect. The preparation stage mainly includes the following key steps:

[0280] 1) Equipment inspection and calibration:

[0281] Thoroughly inspect the Sinsung GCR20-1100 collaborative robot and the polishing tool of its end effector to ensure that all equipment is in good working condition.

[0282] Inspect and calibrate the three-dimensional scanning equipment to ensure that the geometric data of the inner surface of the turbine servomotor can be accurately captured.

[0283] 2) Data acquisition environment setting:

[0284] Prepare a suitable data acquisition environment, including ensuring that the lighting conditions meet the requirements of three-dimensional scanning, and providing sufficient space between the scanning equipment and the object to be scanned.

[0285] Place the servomotor to ensure its fixed position, avoiding displacement or rotation during the scanning process, which may affect the accuracy of the data.

[0286] 3) Pretreatment of the polishing area:

[0287] Conduct a preliminary cleaning of the inner surface of the servomotor to remove surface impurities and dust, ensuring the effectiveness of the polishing work and the accuracy of the data in the polishing area.

[0288] 4) Software preparation:

[0289] Ensure that all relevant software systems, including the deep learning model, path planning algorithm, and software of the control execution system, have been updated and configured.

[0290] Load the pre-trained deep learning model and configuration parameters to prepare for real-time polishing path calculation and optimization.

[0291] Step2: Data acquisition and processing

[0292] In the implementation process of the present invention, data acquisition and processing are the key steps to ensure the success of polishing path optimization. This stage involves the following detailed steps:

[0293] 1) 3D data acquisition:

[0294] Use a high-precision 3D scanning device to comprehensively scan the inner surface of the water turbine servomotor to ensure the integrity of the data.

[0295] 2) Data quality assessment:

[0296] Preliminarily evaluate the quality of the acquired 3D data, including data integrity, accuracy, and resolution. If the data does not meet the expectations, re-scanning is required.

[0297] 3) Data preprocessing:

[0298] Denoise the 3D scan data to remove noise points caused by the scanning environment or equipment. Data format conversion: Convert the original scan data into the format of a 3D mesh model. Perform data normalization to ensure the consistency of data input and improve the efficiency of model training and analysis.

[0299] 4) Feature extraction and annotation:

[0300] Extract key features from the preprocessed data according to the specific requirements of the polishing task, including surface roughness, surface shape accuracy, etc.

[0301] For the data required for training the deep learning model, perform manual data annotation work to clearly mark the areas to be polished and the areas to avoid polishing.

[0302] 5) Data augmentation:

[0303] To improve the generalization ability of the model, perform data augmentation on the processed dataset, including operations such as rotation and translation, to simulate various situations that may be encountered during the polishing process.

[0304] Step3: Model training and verification

[0305] The core of the present invention lies in applying deep learning technology to accurately predict the optimal polishing path for the cladding layer on the inner surface of the servomotor cylinder of a hydraulic turbine.

[0306] 1) Preparation of the training dataset:

[0307] Based on the data processed and enhanced in Phase 3.2, a training dataset is constructed. The dataset includes 3D models with polished areas and non-polished areas marked, as well as the corresponding polishing paths.

[0308] The dataset is divided into a training set, a validation set, and a test set, with 70% for training, 15% for validation, and 15% for testing, to support the training and performance evaluation of the model.

[0309] 2) Model training:

[0310] A deep learning model that combines a convolutional neural network (CNN) and a recurrent neural network (RNN) is used for feature extraction and path prediction. The hyperparameters of the model are configured, including the learning rate, batch size, and number of training epochs.

[0311] The gradient descent algorithm is applied for model training. The model is trained using the training set data and its performance is evaluated on the validation set to monitor overfitting and adjust the hyperparameters.

[0312] Network structure design: Design the network structure, including convolutional layers, pooling layers, fully connected layers, etc., and select the ReLU activation function.

[0313] 3) Performance evaluation and optimization:

[0314] Evaluation metrics such as accuracy, recall, and F1-score are used to comprehensively evaluate the performance of the model on the validation set. Based on the performance evaluation results, the model structure or hyperparameters are adjusted for iterative training until the model performance reaches the predetermined standard.

[0315] 4) Model validation:

[0316] The accuracy and generalization ability of the model are further verified on the test set. Ensure that the model can accurately identify the polished areas and generate effective polishing paths.

[0317] For cases where there are biases in the model recognition or prediction, analyze the reasons and perform model fine-tuning, including increasing the diversity of the dataset, optimizing the feature extraction layer, etc.

[0318] 5) Practical application test:

[0319] Apply the trained model to the polishing task of the cladding layer on the inner surface of the servomotor cylinder of a hydraulic turbine in the laboratory, and observe the degree of coincidence between the predicted polishing path of the model and the actual polishing effect.

[0320] Collect feedback information in actual applications, including surface roughness, surface shape accuracy, and efficiency, for the final evaluation of the model and guidance on future improvement directions.

[0321] Step4: Polishing Execution and Monitoring

[0322] After completing the training of the deep learning model, path planning, and all preparatory work, the next step is the execution and monitoring of the polishing task. The following are the detailed operations:

[0323] 1) Loading and Execution of the Polishing Path:

[0324] Import the optimized polishing path into the control execution unit. The collaborative robot performs actual polishing operations based on these paths, precisely moving the polishing head along the predetermined trajectory to complete the polishing task. During the polishing process, the collaborative robot automatically adjusts parameters such as polishing pressure and speed to adapt to different polishing requirements and ensure the polishing effect.

[0325] 2) Deployment of the Real-time Monitoring System:

[0326] Monitor the polishing process in real time through sensors and cameras installed around the collaborative robot and the working area. These devices collect data on polishing pressure, speed, surface temperature, etc., and feed the data back to the control system. The control system analyzes this data in real time and compares it with the expected polishing effect.

[0327] 3) Dynamic Adjustment and Optimization:

[0328] Based on the real-time monitoring data, the control system can dynamically adjust the polishing parameters, including the polishing path, and adjust the polishing speed or pressure.

[0329] During the polishing process, if it is found that the polishing effect is different from the expectation, the system will automatically pause the operation.

[0330] 4) Evaluation and Feedback of the Polishing Result:

[0331] After completing the polishing task, evaluate the polished surface through high-precision surface quality inspection equipment, including measuring surface roughness and surface shape accuracy. The results show that the polishing effect obtained by the method of the present invention is better than that of traditional polishing.

[0332] The present invention significantly optimizes the polishing process of the clad layer on the inner surface of the turbine servomotor cylinder by integrating advanced deep learning technology, precise data processing methods, and intelligent path planning algorithms. Through strict implementation steps and actual application tests, the present invention proves that this polishing method can achieve a surface roughness of Ra 0.1 micron or lower on the clad layer of the inner surface of the servomotor cylinder, and the cylindricity of the surface shape accuracy is not greater than 0.10 mm.

[0333] Example 3:

[0334] An optimization system for the polishing path of the cladding layer on the inner surface of a cylindrical barrel based on machine learning adopts an optimization method for the polishing path of the cladding layer on the inner surface of a cylindrical barrel based on machine learning, including a data acquisition module, a data preprocessing module, a deep learning model, a path planning algorithm model, and a control execution unit; where:

[0335] 1. Data acquisition module:

[0336] The main function of the data acquisition module is to comprehensively collect the geometric information of the inner surface of the turbine servomotor using high-precision three-dimensional scanning technology. This module ensures that the acquired data has sufficient resolution and accuracy for subsequent data preprocessing and deep learning model analysis.

[0337] 1.1 Equipment selection: A combination of a Leica AT960-LR laser scanner and a T-Scan 5 handheld scanner is adopted. This combined technology can achieve a spatial resolution of up to 0.02 mm, ensuring that the collected data can accurately reflect the microscopic features and complex structures of the inner surface of the servomotor. For the characteristics of the cladding layer on the inner surface of the servomotor cylinder, we optimized the scanning parameters, including scanning speed and resolution, to obtain the best scanning effect.

[0338] 1.2 Scanning process: First, fix the turbine servomotor to be scanned on a rotating platform to ensure stability during the scanning process. Then, according to the size and complexity of the servomotor, adjust the scanning parameters of the laser scanner, including scanning speed, laser intensity, and resolution, etc., to obtain the best scanning effect.

[0339] 1.3 Data format: After the scanning is completed, the raw data will be saved in point cloud format. Each point contains the coordinates (X, Y, Z) in space and the surface reflection intensity of this point. The point cloud data can accurately reflect the three-dimensional structure of the inner surface of the servomotor.

[0340] 1.4 Data correction: To improve the accuracy of the data, the collected raw point cloud data needs to be corrected. This includes eliminating the deviations caused by scanning angles and equipment errors, and merging the data scanned from multiple angles to ensure the coherence of the overall geometric information.

[0341] 1.5 Data optimization: According to the requirements of the subsequent deep learning model, the corrected point cloud data is denoised and simplified to remove irrelevant information.

[0342] 2. Data preprocessing module:

[0343] The main purpose of the data preprocessing module is to convert the original 3D scan data collected by the data acquisition module into a format suitable for analysis by deep learning models. This process includes steps such as data cleaning, denoising, feature extraction, and data normalization.

[0344] Main steps:

[0345] Step1: Data cleaning: Initially process the collected 3D scan data, remove incomplete or damaged data points, and ensure data integrity.

[0346] Step2: Denoising: Use the Savitzky-Golay filter to smooth the original point cloud data and remove the noise generated during the scanning process.

[0347] Step3: Feature extraction: Based on 3D geometric analysis techniques, extract key feature information from the processed data, including the curvature, edges, and geometric shapes of the surface.

[0348] Step4: Data normalization: Adopt the Z-score (standard score method) to convert the data into a format with zero mean and unit variance.

[0349] Step5: Data format conversion: Convert the processed point cloud data into a voxel grid, discretize the 3D space, and provide a standardized input form for the deep learning model.

[0350] 3 Deep learning model:

[0351] The core of the present invention is a deep learning model that deeply combines a convolutional neural network (CNN) and a recurrent neural network (RNN), and is specifically designed for identifying and processing the high-precision machining of the cladding layer on the inner surface of the servomotor cylinder of a hydraulic turbine. The model structure includes several convolutional layers, pooling layers, normalization layers, and fully connected layers. The convolutional layers are used to extract local features of the inner surface of the servomotor; the pooling layers are used to reduce the spatial dimension of the features and improve the generalization ability of the model; the normalization layers are used to accelerate the convergence speed of model training and prevent overfitting; the fully connected layers are finally used to integrate the features from the previous layers and output the optimized decision of the polishing path.

[0352] 3.1 Training method:

[0353] The training of the deep learning model adopts the supervised learning method, and the training data consists of the inner surface data of the servomotor obtained by three-dimensional scanning and the polishing paths manually designed by experts. First, the data is preprocessed and converted into a format acceptable to the model. Then, the Stochastic Gradient Descent (SGD) optimization algorithm is used to optimize the loss function, which measures the difference between the polishing path predicted by the model and the actual optimal path. To avoid overfitting, regularization terms and dropout techniques are introduced. The model is iteratively trained on multiple batches of data until the value of the loss function converges.

[0354] 3.2 Application to polishing path optimization:

[0355] The trained deep learning model is used to identify specific regions on the inner surface of the servomotor in real time and generate polishing paths. Specifically, the model can identify the regions that need to be polished and possible surface defects such as depressions or protrusions. Based on these identification results, the model further calculates the optimal polishing path covering all identified regions, and the path planning optimization includes parameters such as the shape, sequence of the path, and the speed of the polishing head to ensure the uniformity of the polishing effect and the high-quality completion.

[0356] 3.3 Optimization and adjustment:

[0357] In the actual application process, the performance of the model will be continuously optimized according to the polishing results. By collecting the inner surface data of the servomotor after polishing and comparing it with the prediction results, the prediction accuracy of the model is analyzed to find the optimization space. Then, the model is fine-tuned to continuously improve the prediction accuracy of the model and the efficiency of path planning.

[0358] 4 Path planning algorithm model

[0359] 4.1 Algorithm design and implementation:

[0360] The path planning algorithm of the present invention is implemented based on the key region identification results output by the deep learning model. The goal of this algorithm is to generate an efficient, effective and feasible polishing path, which can ensure that all necessary polishing regions are evenly and precisely processed, while minimizing the moving distance and time of the robot, thereby improving the efficiency of the entire polishing process.

[0361] The algorithm first calculates the optimal moving path between these regions based on the polishing regions identified by the deep learning model using the Dijkstra algorithm for the shortest path in graph theory. Then, the algorithm uses heuristic methods to optimize the path and adjusts the path sequence to reduce the idle running time of the robot polishing head.

[0362] 4.2 Optimization strategies:

[0363] To further optimize the polishing path, several key optimization strategies are introduced in this algorithm:

[0364] 1) Area coverage optimization: By adjusting the direction and sequence of the polishing path, ensure uniform coverage of the inner surface area of complex shapes.

[0365] 2) Collision avoidance: Utilize the robot kinematic model to predict potential collisions and adjust the path accordingly to ensure the safety of the operation.

[0366] 3) Efficiency improvement: By analyzing the size and shape of the polishing area, intelligently allocate the polishing time. For smaller or simpler-shaped areas, increase the overall efficiency by reducing the polishing time.

[0367] 5 Control execution unit:

[0368] 5.1 Selection and configuration of collaborative robots:

[0369] The present invention uses a collaborative robot (Cobot) as the main body for polishing execution, and its selection is based on the significant advantages of collaborative robots in terms of precise control, flexibility, and safety. The selected collaborative robot has high-precision positioning capabilities and programmable multi-axis collaborative working characteristics, and can meet the requirements of complex inner surface polishing tasks. The collaborative robot performs the polishing operation through a dedicated polishing head attachment, and different polishing tools can be replaced according to different polishing requirements.

[0370] 5.2 Control system design:

[0371] The core of the control execution unit is an advanced control system, which is responsible for converting the polishing path received from the path planning algorithm module into motion instructions for the collaborative robot. The control system uses a real-time operating system (RTOS) to ensure the immediate response and efficient execution of instructions during the polishing process. Through the graphical user interface (GUI), the operator can monitor the polishing process in real time and perform necessary manual interventions.

[0372] 5.3 Dynamic adjustment and feedback mechanism:

[0373] During the polishing operation, the control system collects real-time status information of the collaborative robot and feedback on the polishing effect, including but not limited to polishing pressure, polishing speed, and surface quality data. Using this information, the control system can dynamically adjust polishing parameters (such as polishing pressure and speed) to cope with any abnormal situations that occur during the polishing process, ensuring the consistency and optimization of the polishing quality.

[0374] The above embodiments are only the preferred technical solutions of the present invention and should not be regarded as limitations on the present invention. The protection scope of the present invention should be the technical solutions recorded in the claims, including equivalent replacement solutions of the technical features in the technical solutions recorded in the claims. That is, equivalent replacement improvements within this scope are also within the protection scope of the present invention.

Claims

1. A polishing path optimization method for the inner surface cladding layer of a cylindrical barrel based on machine learning, characterized in that The following steps are involved: Step 1: Collect the geometric data of the inner surface of the cylindrical barrel through a 3D scanner: Step 1.1, fix the cylindrical barrel on the rotating platform; Step 1.2, according to the size and complexity of the cylindrical barrel, adjust the scanning parameters of the 3D scanner, including scanning speed, laser intensity and resolution; and then use the 3D scanner to collect geometric data; Step 1.3, save the geometric data as raw point cloud data in point cloud format, where each point contains the coordinates (X, Y, Z) in space and the surface reflection intensity of the point; Step 1.4: Correct the original point cloud data to eliminate the deviation caused by scanning angle and equipment error; merge the original point cloud data scanned at multiple angles to ensure the consistency of the overall geometric data; Step 1.5, denoise and simplify the corrected original point cloud data to remove irrelevant information; Step 2: Convert the collected raw point cloud data into a format suitable for deep learning model analysis, use the deep learning model to process the geometric data and identify the key areas that need polishing: Step 2.1, data cleaning: preliminarily process the collected 3D scanning data, remove incomplete or damaged data points, and ensure data integrity; Step 2.2, denoising: Use a filter to smooth the original point cloud data and remove the noise generated during the scanning process; Step 2.3, feature extraction: based on the 3D geometric analysis method, extract key feature information from the processed geometric data, the key feature information includes the curvature, edge and geometric shape of the surface; Step 2.4, data standardization: Use the standard score method to transform the geometric data into a format with zero mean and unit variance; Step 2.5, data format conversion: convert the processed point cloud data into voxel grids, discretize the three-dimensional space, and provide a standardized input form for the deep learning model; Step 3: Use supervised learning method to train the deep learning model, and use stochastic gradient descent optimization algorithm to optimize the loss function. The stochastic gradient descent optimization algorithm is used to optimize the loss function to measure the difference between the polishing path predicted by the deep learning model and the actual optimal path. To avoid overfitting, regularization terms and discarding methods are introduced. The deep learning model is iteratively trained on multiple batches of data until the value of the loss function converges. Step 4: The trained deep learning model is used to identify a certain area on the inner surface of the cylindrical barrel in real time and generate a polishing path. The deep learning model is used to identify the area that needs to be polished and possible surface defects, including depressions and protrusions. Based on the recognition results, the deep learning model further calculates the optimal polishing path covering all the recognized areas. The optimal polishing path includes parameters such as path shape, sequence, and polishing head speed. In step 1, the scanning parameters also include a scanning interval, and the scanning parameters are adjusted according to the material properties, size and complexity of the cladding layer of the inner surface of the cylindrical barrel. During the scanning process, multiple angles and multiple positions are used to scan to capture all visible areas of the inner surface of the cylindrical barrel and ensure complete coverage of geometric data to avoid dead angles, so as to obtain complete three-dimensional information; In step 1, formula (1) is used to calculate the shortest distance between points in the point cloud data: , to assist in subsequent data processing and path optimization: （1）； in, and are any two points in the point cloud data, for The coordinates of for The coordinates of In step 2.1, since irrelevant data points are generated during the scanning process, including noise caused by environmental reflection or equipment error, the outliers are identified and removed by using the standard deviation and mean through threshold segmentation and statistical analysis methods; In step 2.2, the Gaussian filter algorithm is used to denoise the data to smooth the data and reduce the impact of scanning errors. The formula of the Gaussian filter algorithm is as follows: （2）； in, represents the original scan data, represents the filtered data, is the weight matrix of the filter, are the coordinates of the data points, is the position in the weight matrix; The original point cloud data is then processed using formula (3) to eliminate noise and smooth the data: （3）； in, represents the smoothed points, Yes The neighborhood set of ; In step 2.4, the data is normalized using formula (4) to ensure that the training of the deep learning model is not affected by different scanning scales: （4）； in, is the original data point, and are the minimum and maximum values ​​in the data set, respectively; After step 2.5 is completed, perform boundary detection: For the boundary areas that need to be precisely polished, use edge detection algorithms to identify and mark these areas: （5）； in, represents the edge detection result, represents the image intensity, and are the gradients of the image intensity in the x and y directions, respectively.

2. The method for optimizing the polishing path of the inner surface cladding layer of a cylindrical barrel based on machine learning according to claim 1 is characterized in that: The deep learning model uses convolutional neural networks and recurrent neural networks as the basic architecture to process and identify spatial hierarchical information in image data. The deep learning model includes input layer, convolution layer, pooling layer and fully connected layer. Among them, the input layer: adapts the network architecture by accepting the two-dimensional depth map converted from the three-dimensional scanning data or the processed point cloud data; Convolutional layer: Implement multiple convolutional layers, each layer uses multiple convolution kernels to extract local features of the data. The formula of the convolutional layer is as follows: （6）； in, It is Layer In the feature map The characteristic response, represents the activation function, and They are The bias and convolution kernel weight of the layer, * indicates the convolution operation; Pooling layer: The maximum pooling of the pooling layer is introduced to reduce the feature dimension and extract more abstract feature representation; Fully connected layer: A fully connected layer is applied at the back end of the network to map the distribution of the polished area into a form recognizable by the output layer. The formula of the fully connected layer is: （7）； in, is the output layer neurons, is the output of the previous layer, and represent weight and bias respectively, is the activation function; The deep learning model uses the cross entropy loss function to evaluate the model: （8）； in, is the number of categories for classification, is the actual label, is the probability predicted by the model.

3. The method for optimizing the polishing path of the inner surface cladding layer of a cylindrical barrel based on machine learning according to claim 1 is characterized in that: In step 3, the preprocessed dataset is divided into training set, validation set and test set. The training process is monitored and the model hyperparameters are adjusted by evaluating the model performance on the validation set. The model is trained using batch gradient descent, and the weights and biases are updated iteratively: （9）； in, represents the model parameters, is the learning rate, is the gradient of the loss function with respect to the parameters.

Images