A laser rust removal method based on machine learning

Abstract

The application discloses a laser rust removal method based on machine learning. First, collect rust removal experimental data of different materials of metal under different laser rust removal parameters, including pre- and post-rust removal metal surface images, rust degree, substrate damage and other data and preprocessing. Then, extract features related to rust removal effect and substrate damage, and screen key features through a feature selection algorithm. Select a suitable machine learning model, train the model with the key features as input and the rust removal effect evaluation index as output, and optimize through cross-validation. In actual rust removal, the sensor collects data and combines the model for online monitoring and prediction, and the parameters are dynamically adjusted by the optimization algorithm according to the results. After rust removal, the effect is evaluated, the actual and predicted results are compared, and the model is updated when the deviation is large. The method can accurately control the rust removal effect, protect the substrate, improve the efficiency and realize self-optimization, and has a wide application prospect.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of laser rust removal, and particularly to a laser rust removal method based on machine learning. BACKGROUND

[0002] In many fields such as industrial production, cultural relic restoration, and building maintenance, metal surface rust removal is a common and important work. Traditional rust removal methods, such as mechanical polishing and chemical rust removal, have many limitations. Mechanical polishing can easily damage the metal substrate and is difficult to handle workpieces with complex shapes and delicate structures; chemical rust removal can cause environmental pollution and requires subsequent waste liquid treatment procedures.

[0003] Laser rust removal, as an emerging rust removal technology, has the advantages of non-contact, no wear, high efficiency, and environmental protection, and has gradually attracted widespread attention. However, the laser rust removal process is affected by many factors, such as laser power, scanning speed, pulse frequency, spot diameter, and the rusting degree and material properties of the metal surface. How to accurately control these parameters to achieve the best rust removal effect while avoiding excessive damage to the substrate is a key problem currently faced by laser rust removal technology.

[0004] Machine learning technology has strong capabilities in handling complex multivariate relationships and optimizing decisions. Through learning a large amount of laser rust removal experimental data or actual rust removal case data, a machine learning model can establish an accurate mapping relationship between the rust removal effect and various influencing factors, thereby providing scientific basis and intelligent guidance for the optimization of laser rust removal parameters. SUMMARY

[0005] The present application aims to provide a laser rust removal method based on machine learning to solve the problems raised in the background art.

[0006] To achieve the above-mentioned purpose, the present application provides the following technical solution: a laser rust removal method based on machine learning, comprising the following steps:

[0007] Step 1, data collection and preprocessing: collect a large amount of rust removal experimental data of metal surfaces of different materials under different laser rust removal parameters, including laser power, scanning speed, pulse frequency, and spot diameter. The rust removal experimental data includes image data, rusting degree data, and substrate damage data of the metal surface before and after rust removal. Perform cleaning and normalization preprocessing operations on the collected data to improve the quality and usability of the data.

[0008] Step 2, feature extraction and selection: Extract features related to laser rust removal effect and substrate damage from pre-processed data, including laser parameter features, metal surface features, and post-rust surface quality features; use feature selection algorithms to filter out key features that significantly affect laser rust removal effect and substrate damage, reduce data dimensionality, and improve model training efficiency;

[0009] Step 3, model training: Select a neural network model, use the filtered features in Step 2 as input, and the rust removal effect evaluation index as output to train the model; use cross-validation method to adjust the model's hyperparameters and optimize the model's performance, so that the model can accurately predict the rust removal effect and substrate damage under different laser parameters;

[0010] Step 4, online monitoring and parameter optimization: In the actual laser rust removal process, use sensors to collect real-time data from the metal surface, and combine the pre-trained neural network model in Step 3 to monitor and predict the current rust removal effect and substrate damage; based on the prediction results, dynamically adjust the laser rust removal parameters through optimization algorithms to achieve the best rust removal effect and minimal substrate damage;

[0011] Step 5, effect evaluation and model update: After completing the laser rust removal operation, comprehensively evaluate the rust removal effect by detecting the microscopic morphology and residual rust amount on the metal surface after rust removal using optical microscopes and scanning electron microscopes; compare the actual evaluation results with the model prediction results, and if there is a large deviation between the two, add this rust removal data to the training data set and update and optimize the machine learning model to continuously improve the model's accuracy and adaptability.

[0012] Preferably, the specific implementation steps of data collection and preprocessing in Step 1 are as follows:

[0013] Step 11, experimental design and sample preparation: Select experimental samples for various common metal materials; for each experimental sample, prepare samples with different degrees of rust, and control and adjust the rust degree through natural exposure time and artificial accelerated rusting methods to cover various conditions from mild to severe rust;

[0014] Step 12, image data collection: Use high-resolution industrial cameras or optical microscopes to take images of the metal surface before and after rust removal to obtain image data;

[0015] Step 13, rust degree data collection: Quantify the rust degree using chemical analysis, specifically measure the mass loss per unit area of the rusted sample before rust removal to determine the rust amount, use a high-precision electronic balance to weigh the mass change of the sample before and after rust removal, and calculate the rust mass thickness based on the sample surface area;

[0016] Step 14, substrate damage data collection: Use a scanning electron microscope to observe the micro-morphology of the metal substrate surface after rust removal, check whether there are micro-cracks, fusion layer, heat affected zone substrate damage characteristics; For the existing damage, through the measurement of the depth, width of the damage area and the calculation of the area ratio of the damage area to the entire rust removal area, the damage is quantified; At the same time, the hardness test method is used to compare the change of the hardness of the metal substrate before and after rust removal, and the hardness test is used to evaluate the damage degree of the substrate indirectly, and the hardness test is measured by using a microhardness tester in the rust removal area and the untreated area respectively and taking the average value;

[0017] Step 15, outlier processing: For image data, check if there are image blur, missing or obviously wrong pixel area caused by camera failure, abnormal light, etc.; For rust degree data and substrate damage data, if the measured value deviates obviously from the normal range, further check the original measurement record or repeat the measurement to determine the accuracy of the data;

[0018] Step 16, image data normalization: For the image data collected in the above steps, the pixel value of the image is normalized, that is, the pixel value of the image is mapped to the interval [0, 1];

[0019] Step 17, numerical data normalization: Rust degree data, substrate damage data and laser rust removal parameter data are normalized by Min-Max normalization or Z-score standardization method.

[0020] Preferably, the specific steps of feature extraction and selection of step 2 are as follows:

[0021] Step 21, feature extraction:

[0022] a. Laser parameter feature: directly extract the laser rust removal parameters set in data collection as features;

[0023] b. Metal surface feature:

[0024] Roughness feature: use a roughness measuring instrument to measure the metal surface before rust removal to obtain the surface roughness parameter; Texture feature: extract the texture feature from the image data of the metal surface before rust removal;

[0025] c. Surface quality feature after rust removal:

[0026] Residual rust amount feature: determine the residual rust amount of the metal surface after rust removal by chemical analysis method, which is an improvement of the mass loss measurement method, and calculate the difference between the ideal rust removal mass and the mass after rust removal to determine the residual rust amount; Surface flatness feature: use a profilometer or an optical interferometer to measure the flatness change of the metal surface after rust removal;

[0027] Step 22, feature selection: principal component analysis, first, calculate the covariance matrix of the feature data in step 21; then, solve the eigenvalue and eigenvector of the covariance matrix, the eigenvalue represents the variance of the data in the corresponding eigenvector direction; finally, sort the eigenvectors according to the size of the eigenvalues, select the first k eigenvectors, and the eigenvectors constitute the principal component transformation matrix, that is, the k eigenvectors are the key features after screening.

[0028] Preferably, the specific implementation steps of model training in step 3 are as follows:

[0029] Step 31, neural network model structure selection and initialization: select a multi-layer feedforward neural network, including input layer, hidden layer and output layer; the number of input layer nodes is equal to the number of screened features, that is, after feature selection, it is determined that the four key features are laser power, scanning speed, metal surface roughness and residual rust amount, so the number of input layer nodes is 4, the number of output layer nodes is 2, which corresponds to the rust removal rate and substrate damage degree in the rust removal effect evaluation index respectively; the number and node number of the hidden layer need to be determined according to experience and experiment;

[0030] Step 32, forward propagation: let the input feature vector be X = [x1, x2, x3, x4], where x1 is laser power, x2 is scanning speed, x3 is metal surface roughness, and x4 is residual rust amount. For the calculation from the input layer to the hidden layer, the input of the jth node of the hidden layer is net j The calculation formula is:

[0031]

[0032] Where ω ij is the weight from the ith node of the input layer to the jth node of the hidden layer, b j is the bias of the jth node of the hidden layer;

[0033] The output h j of the hidden layer node is calculated by using the activation function, and the commonly used sigmoid activation function is selected here: For the calculation from the hidden layer to the output layer, the input of the kth node of the output layer is net′ k The calculation formula is:

[0034]

[0035] Where ω′ jk is the weight from the jth node of the hidden layer to the kth node of the output layer, b′ k is the bias of the kth node of the output layer;

[0036] The output y kLikewise, the activation function is used, and the rust removal rate is mapped to the interval [0, 1] by using the sigmoid function, representing the proportion of rust removal; the base damage degree is mapped to a numerical value by using the linear activation function, representing the quantitative value of the damage degree;

[0037] Step 33, loss function calculation: the mean square error is used as the loss function to measure the difference between the model prediction value and the true value, and the loss function L is calculated as where yk is the true value of the kth output, i.e., the rust removal rate and the base damage degree actually measured, and yk is the model prediction value.

[0038] Step 34, back propagation and weight update: the gradient of the loss function with respect to the weights and biases of each layer is calculated by the back propagation algorithm, and then the weights and biases are updated according to the gradient descent algorithm, wherein the weights ω' from the hidden layer to the output layer are updated as jk The update formula is:

[0039]

[0040] where η is the learning rate, which is a hyperparameter, controlling the step size of weight update, is the gradient of the loss function with respect to ω', which is calculated step by step from the output layer to the hidden layer by the chain rule; jk

[0041] Step 35, cross-validation adjustment of hyperparameters: the collected laser rust removal experiment data is divided into k subsets, and each time k-1 subsets are selected as the training set and the remaining 1 subset is selected as the validation set; the hyperparameter combination with the smallest loss function value on the validation set is selected as the final model hyperparameters, and then the model is retrained under the hyperparameter combination using the entire training data set to obtain the final trained neural network model.

[0042] Preferably, the specific implementation of the online monitoring and parameter optimization of step 4 is as follows:

[0043] A, sensor data acquisition and transmission

[0044] Laser parameter sensor: a high-precision laser power sensor is installed on the laser rust removal equipment to monitor the laser output power in real time; at the same time, a scanning speed sensor is installed to accurately measure the moving speed of the laser scanning head through encoder and other technical means;

[0045] ​Surface state sensor: use infrared temperature sensor to measure the real-time temperature change of metal surface under the action of laser; at the same time, use optical reflectivity sensor to detect the reflected light intensity after laser irradiation to the metal surface, the reflected light intensity will be different under different rust degree and metal substrate state, through the monitoring of reflectivity, auxiliary judge the progress of rust removal;

[0046] B, data preprocessing and feature extraction

[0047] Real-time preprocessing of collected sensor data, including removing noise interference;

[0048] Extract real-time features corresponding to model input features, directly obtain current laser power value from laser power sensor data, obtain scanning speed value from scanning speed sensor, extract metal surface real-time temperature features from infrared temperature sensor data, extract reflectivity features from optical reflectivity sensor data, etc. The above features will be used as input for subsequent model prediction;

[0049] C, prediction based on machine learning model: input the real-time feature data after preprocessing and extraction into the pre-trained neural network model in step 3, the model predicts the current rust removal rate and substrate damage degree according to the input feature data and the relationship between the learned features and rust removal effect and substrate damage in the training stage;

[0050] D, optimization algorithm and parameter adjustment

[0051] Optimization goal setting: set the optimization goal as improving the rust removal rate while ensuring the substrate damage degree is lower than the set threshold;

[0052] Optimization algorithm selection and application: genetic algorithm is used for parameter optimization, first, encode laser power, scanning speed, metal surface roughness and residual rust amount parameters into chromosome form, then initialize a population containing multiple such encoded parameter combination individuals, calculate the predicted rust removal rate and substrate damage degree corresponding to each individual, set fitness function according to optimization goal, generate new parameter combination individuals through selection, crossover and mutation genetic operations, and find the parameter combination with the highest fitness;

[0053] E, parameter adjustment implementation: send the optimized laser rust removal parameters to the controller of the laser rust removal equipment, the controller adjusts the power output of the laser generator and the rotating speed of the scanning motor according to the received parameter instructions, realizes the dynamic adjustment of the laser rust removal parameters, after adjusting the parameters, continue to collect data through the sensor, repeat the above monitoring, prediction and optimization process until the satisfactory rust removal effect and the substrate damage within the controllable range are achieved.

[0054] Preferably, the comparison and analysis of the actual evaluation results and the model prediction results in step 5 are as follows: compare the rust removal effect and substrate damage degree data obtained by evaluation with the corresponding results predicted by the machine learning model during the laser rust removal process, calculate the deviation value, for the rust removal rate deviation: ΔRR = |RR 实际 -RR 预测 |, for the substrate damage degree deviation ΔBD = |BD 实际 -BD 预测 |, set a reasonable deviation threshold, compare the calculated deviation value with the threshold, if the deviation is large, the machine learning model needs to be updated; if the deviation is within the threshold range, it is considered that the current model prediction effect is good, and the model can not be updated temporarily.

[0055] Compared with the prior art, the beneficial effects of the present application are: through the learning of a large amount of experimental data and actual case data by the machine learning model, the present application can accurately establish the complex relationship between the rust removal effect and the laser power, scanning speed, pulse frequency, spot diameter and many other parameters as well as the metal surface characteristics. In the actual rust removal process, based on the real-time data collected by the sensor, the model can accurately predict the current rust removal effect, and dynamically adjust the laser rust removal parameters using the optimization algorithm. This makes the rust removal process no longer rely on empirical parameter setting, but intelligently optimize according to the real-time state of the metal surface, thereby significantly improving the rust removal rate and ensuring that the metal surface can reach a higher cleanliness standard. For example, when dealing with metal parts with complex rusting conditions and different material mixtures, the model can quickly adapt and find the best parameter combination to effectively remove stubborn rust, while traditional methods often cannot achieve such precise parameter matching, resulting in incomplete rust removal or excessive processing; adapt to various working conditions: whether it is facing metal surfaces of different materials (such as carbon steel, stainless steel, aluminum alloy, etc.), or dealing with various rusting conditions (from mild surface oxidation to heavy rust layer accumulation), this method can flexibly adjust the parameter strategy through the self-adaptive learning ability of the machine learning model, achieving the best rust removal effect. This greatly expands the application range of laser rust removal technology, enabling it to be more widely and efficiently applied in many industrial fields (such as mechanical manufacturing, automobile repair, ship maintenance, etc.) as well as cultural relic restoration, building decoration and other special industries, solving the poor adaptability problem of traditional rust removal methods when dealing with different working conditions;

[0056] Meanwhile, the machine learning model can also predict and monitor the base body damage in real time while accurately controlling the rust removal parameters. Through comprehensive analysis of the metal surface features, laser parameters, and real-time monitoring data, the model can provide early warning for parameter settings that may cause excessive damage to the base body and make timely adjustments to control the base body damage to the minimum. For example, when detecting that the high metal surface temperature or the large laser energy density may cause base body melting or micro-cracks, the model will automatically adjust the scanning speed or reduce the laser power to avoid damage to the structural integrity and performance of the base body, while the traditional fixed parameter setting laser rust removal method cannot achieve such precise damage prevention and control. BRIEF DESCRIPTION OF DRAWINGS

[0057] Figure 1 The method flowchart of the present application is shown. DETAILED DESCRIPTION

[0058] The technical solutions in the embodiments of the present application will be described clearly and completely below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only a part of the embodiments of the present application, rather than all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor fall within the scope of protection of the present application.

[0059] Please refer to Figure 1 The present application provides a technical solution: a laser rust removal method based on machine learning, comprising the following steps:

[0060] Step 1, data collection and preprocessing: collecting a large amount of experimental data of different metal surface materials under different laser rust removal parameters, including laser power, scanning speed, pulse frequency, and spot diameter, and the experimental data includes image data, rust degree data, and base body damage data of the metal surface before and after rust removal; cleaning and normalizing the collected data to improve the quality and availability of the data;

[0061] The specific implementation steps of data collection and preprocessing are as follows:

[0062] Step 11, experimental design and sample preparation: selecting experimental samples for a variety of common metal materials; for each experimental sample, samples with different rust degrees are prepared, and the rust degree is controlled and adjusted by natural exposure time and artificial accelerated rusting method to cover various conditions from mild to severe rusting;

[0063] Step 12, image data collection: using a high-resolution industrial camera or an optical microscope to take images of the metal surface before and after rust removal to obtain image data;

[0064] Step 13, Corrosion Degree Data Collection: Quantify the corrosion degree by chemical analysis, specifically measure the mass loss per unit area of the corroded sample before cleaning to determine the amount of corrosion. Use a high-precision electronic balance to weigh the mass change of the sample before and after cleaning, and calculate the corrosion thickness by combining the sample surface area;

[0065] Step 14, Base Damage Data Collection: Use a scanning electron microscope to observe the microstructure of the metal base surface after cleaning, check for the presence of micro-cracks, fusion layers, and heat-affected zone base damage characteristics; For the existing damage, quantify by measuring the depth and width of the damage area and calculating the area ratio of the damage area to the entire cleaning area; At the same time, use hardness testing method to compare the change of metal base hardness before and after cleaning, indirectly evaluate the base damage degree, hardness testing uses microhardness tester to measure multiple points in cleaning area and untreated area respectively and takes average value;

[0066] Step 15, Outlier Processing: For image data, check if there are blurred, missing or obviously wrong pixel areas caused by camera failure, abnormal lighting, etc.; For corrosion degree data and base damage data, if the measured value deviates significantly from the normal range, further check the original measurement record or repeat the measurement to determine the accuracy of the data;

[0067] Step 16, Image Data Normalization: For the image data collected in the above steps, normalize the pixel values of the image, specifically map the pixel values of the image to the [0, 1] interval;

[0068] Step 17, Numerical Data Normalization: For corrosion degree data, base damage data and laser cleaning parameter data, use Min-Max normalization or Z-score standardization method for normalization processing; The normalization method depends on the distribution characteristics of the data and the needs of the subsequent model, the normalized numerical data can improve the efficiency and accuracy of machine learning model training, avoid the problem of numerical instability in model training process caused by large difference in data magnitude.

[0069] Through the above detailed data collection and preprocessing steps, high-quality, standardized data basis can be provided for laser cleaning method based on machine learning, thereby improving the reliability and effectiveness of subsequent model training and prediction.

[0070] Step 2, Feature Extraction and Selection: Extract features related to laser cleaning effect and base damage from preprocessed data, including laser parameter features, metal surface features, and surface quality features after cleaning; Use feature selection algorithm to filter out key features that have significant impact on laser cleaning effect and base damage, reduce data dimension, and improve model training efficiency;

[0071] The specific steps of feature extraction and selection are as follows:

[0072] Step 21, feature extraction:

[0073] a. Laser parameter features: directly extract the laser rust removal parameters set in the data collection as features; including laser power (unit: watt, W), scanning speed (unit: millimeter per second, mm / s), pulse frequency (unit: kilohertz, kHz), spot diameter (unit: millimeter, mm). These parameters directly affect the interaction process between laser and metal surface, and are the key factors to determine the rust removal effect and substrate damage. For example, higher laser power can provide more energy to remove rust, but it may also increase the risk of substrate damage; faster scanning speed may reduce the laser energy deposition per unit area, affecting the thoroughness of rust removal, etc.

[0074] b. Metal surface features:

[0075] Roughness features: use a roughness measuring instrument to measure the metal surface before rust removal to obtain surface roughness parameters such as arithmetic average roughness (Ra), root mean square roughness (Rq), etc. The roughness of the metal surface will affect the reflection and absorption characteristics of the laser, and then affect the rust removal effect. Rougher surface may cause laser scattering, reducing the rust removal efficiency, but in some cases it may also help the distribution of laser energy on the surface, reducing substrate damage caused by local overheating; texture features: extract texture features from the image data of the metal surface before rust removal, specifically using the gray level co-occurrence matrix (GLCM) method to calculate contrast, correlation, energy, entropy, etc. Texture parameters. For example, metal surfaces with different rust textures (such as uniform rust, local concentrated rust, mottled rust, etc.) may respond differently to laser, and these texture features help describe the relationship between the initial state of the metal surface and the rust removal effect;

[0076] c. Surface quality features after rust removal:

[0077] Residual rust amount features: determine the residual rust amount on the metal surface after rust removal by chemical analysis method, which is an improvement of the mass loss measurement method. The sample is weighed again after rust removal, and the difference from the ideal rust removal mass is calculated to determine the residual rust amount; surface flatness features: use a profilometer or optical interferometer to measure the flatness change of the metal surface after rust removal;

[0078] Step 22, feature selection: principal component analysis, first, calculate the covariance matrix of the feature data in step 21; then, solve the eigenvalues and eigenvectors of the covariance matrix, the eigenvalues represent the variance of the data in the corresponding eigenvector direction; finally, sort the eigenvectors according to the size of the eigenvalues, select the first k eigenvectors, and the eigenvectors constitute the principal component transformation matrix, that is, the k eigenvectors are the key features after screening, wherein the value of k is usually determined according to the cumulative contribution rate, when the cumulative contribution rate reaches a certain proportion (such as 85%-95%), the k principal components are considered to retain most of the information of the original data. For example, if there are 10 features originally, and after PCA analysis, it is found that the cumulative contribution rate of the first 5 principal components reaches 90%, then select the features corresponding to the 5 principal components as the key features, realize the dimensionality reduction of the data, and at the same time retain the main information, reduce the interference of redundant features on the subsequent machine learning model training.

[0079] Step 3, model training: select a neural network model, take the screened features in step 2 as input, and take the rust removal effect evaluation index as output to train the model; adjust the hyperparameters of the model by cross-validation method to optimize the performance of the model, so that the model can accurately predict the rust removal effect and substrate damage under different laser parameters;

[0080] The specific implementation steps of model training are as follows:

[0081] Step 31, neural network model structure selection and initialization: select a multi-layer feedforward neural network, which includes an input layer, a hidden layer and an output layer; the number of nodes in the input layer is equal to the number of screened features, that is, after feature selection, it is determined that the four key features are laser power, scanning speed, metal surface roughness and residual rust amount, so the number of nodes in the input layer is 4, the number of nodes in the output layer is 2, which corresponds to the rust removal rate and substrate damage degree in the rust removal effect evaluation index respectively; the number and node number of the hidden layer need to be determined according to experience and experiment;

[0082] Step 32, forward propagation: let the input feature vector be X = [x1, x2, x3, x4], where x1 is the laser power, x2 is the scanning speed, x3 is the metal surface roughness, and x4 is the residual rust amount. For the calculation from the input layer to the hidden layer, the input of the jth node of the hidden layer is net j The calculation formula is:

[0083]

[0084] Where ω ij is the weight from the ith node of the input layer to the jth node of the hidden layer, b j is the bias of the jth node of the hidden layer;

[0085] The output of the hidden layer node hj The calculation is performed using an activation function, and the commonly used sigmoid activation function is selected here: For the calculation from the hidden layer to the output layer, the input of the kth node of the output layer is net' k The calculation formula is:

[0086]

[0087] Where ω' jk is the weight from the jth node of the hidden layer to the kth node of the output layer, b' k is the bias of the kth node of the output layer;

[0088] The output y k of the output layer node is also calculated using an activation function. For the rust removal rate, a sigmoid function is used to map the output to the [0, 1] interval, representing the proportion of rust removal. For the base damage degree, a linear activation function is used, and the output is a numerical value representing the quantitative value of the damage degree.

[0089] Step 33, loss function calculation: the mean square error is used as the loss function to measure the difference between the model prediction value and the true value, set as the true value of the kth output, i.e. the rust removal rate and base damage degree actually measured, y k as the model prediction value, and the loss function L calculation formula is:

[0090] Step 34, backpropagation and weight update: the loss function is calculated by the backpropagation algorithm for the gradient of each layer weight and bias, and then the weight and bias are updated according to the gradient descent algorithm, where the weight ω' jk from the hidden layer to the output layer is updated according to the formula:

[0091]

[0092] Where η is the learning rate, which is a hyperparameter, controlling the step size of weight update, is the gradient of the loss function with respect to ω' jk , which is calculated step by step from the output layer to the hidden layer through the chain rule;

[0093] Step 35, cross-validation adjustment of hyperparameters: the collected laser rust removal experiment data is divided into k subsets, and each time k-1 subsets are selected as the training set and the remaining 1 subset is selected as the validation set; the hyperparameter combination with the smallest loss function value on the validation set is selected as the final model hyperparameters, and then the model is retrained under this hyperparameter combination using the entire training data set to obtain the final trained neural network model;

[0094] Through the above steps, the neural network model is used to learn and model the data in the laser rust removal method based on machine learning, to realize accurate prediction and analysis of the laser rust removal effect and substrate damage, and to provide reliable model support for subsequent online monitoring and parameter optimization.

[0095] Step 4, online monitoring and parameter optimization: in the actual laser rust removal process, the relevant data of the metal surface are collected in real time by the sensor, and the neural network model pre-trained in step 3 is combined to monitor and predict the current rust removal effect and substrate damage online; according to the prediction result, the laser rust removal parameters are dynamically adjusted through the optimization algorithm to realize the best rust removal effect and the minimum substrate damage;

[0096] The specific implementation content of online monitoring and parameter optimization is as follows:

[0097] A, sensor data acquisition and transmission

[0098] Laser parameter sensor: install a high-precision laser power sensor on the laser rust removal equipment to monitor the laser output power in real time; at the same time, install a scanning speed sensor to accurately measure the moving speed of the laser scanning head through encoder and other technical means;

[0099] Surface state sensor: use an infrared temperature sensor to measure the real-time temperature change of the metal surface under the action of laser; at the same time, use an optical reflectivity sensor to detect the reflected light intensity after the laser irradiates the metal surface, the reflected light intensity will be different under different rusting degrees and metal substrate states, and the rust removal progress can be judged by monitoring the reflectivity;

[0100] B, data preprocessing and feature extraction

[0101] The collected sensor data are preprocessed in real time, including removing noise interference;

[0102] Extract real-time features corresponding to model input features, directly obtain current laser power value from laser power sensor data, obtain scanning speed value from scanning speed sensor, extract metal surface real-time temperature features from infrared temperature sensor data, and extract reflectivity features from optical reflectivity sensor data, etc. The above features will be used as input for subsequent model prediction;

[0103] C, prediction based on machine learning model: input the preprocessed and extracted real-time feature data into the neural network model pre-trained in step 3, the model predicts the current rust removal rate and substrate damage degree according to the input feature data, combining the relationship between the learned features and the rust removal effect and substrate damage in the training stage;

[0104] D, optimization algorithm and parameter adjustment

[0105] Optimization target setting: set the optimization target as improving the rust removal rate while ensuring that the substrate damage level is below a set threshold;

[0106] Optimization algorithm selection and application: genetic algorithm is used for parameter optimization. First, the laser power, scanning speed, metal surface roughness, and residual rust amount parameters are encoded into chromosome form. Then, a population is initialized, containing multiple such encoded parameter combination individuals. The predicted rust removal rate and substrate damage level corresponding to each individual are calculated. According to the optimization target setting, the fitness function is set. Through selection, crossover, and mutation genetic operations, new parameter combination individuals are continuously generated, and the parameter combination with the highest fitness is found.

[0107] E. Parameter adjustment implementation: the optimized laser rust removal parameters are sent to the controller of the laser rust removal equipment. The controller adjusts the power output of the laser generator and the speed of the scanning motor according to the received parameter instructions, realizing dynamic adjustment of the laser rust removal parameters. After adjusting the parameters, the data is continuously collected through the sensor, and the above monitoring, prediction, and optimization process is repeated until the desired rust removal effect is achieved and the substrate damage is within a controllable range.

[0108] Through the above online monitoring and parameter optimization steps, the laser rust removal parameters can be intelligently adjusted in real time during the laser rust removal process according to the actual situation of the metal surface and the prediction results of the pre-trained model, improving the rust removal quality and efficiency, reducing substrate damage, and realizing automatic and precise laser rust removal operations.

[0109] Step 5, Effect Evaluation and Model Updating: After completing the laser rust removal operation, the rust removal effect is comprehensively evaluated. Specifically, the microscopic morphology and residual rust amount on the metal surface after rust removal are detected by optical microscopy and scanning electron microscopy equipment. Specifically, an image analysis method is used: high-resolution microscopes or industrial cameras are used to collect images of the metal surface after rust removal. The shooting area should cover the entire rusted surface or multiple areas according to certain sampling rules. Image processing software is used to analyze the collected images. The metal matrix and possible residual rust area are separated by image segmentation algorithm, and then the area proportion or pixel number proportion of residual rust area in the total detection area is calculated. For example, if the total detection area pixel number is 1 million, and the residual rust area pixel number obtained by image analysis is 50,000, then the residual rust area proportion is 5%. This method has the advantages of non-destructive and rapid detection, but the detection accuracy may be limited for some microscopic and similar color residual rust on the matrix; microscopic morphology observation: scanning electron microscopy (SEM) is used to observe the microscopic morphology of the metal surface after rust removal. SEM can provide high magnification surface images, clearly showing whether there are micro-cracks, fusion layers, heat-affected zones and other matrix damage features on the metal surface. The length, width and depth of the micro-cracks are measured, and the number and distribution density of the parameters are counted to quantify the matrix damage degree. For example, if 3 micro-cracks with a length of 10-20 μm are found in a 10 mm x 10 mm observation area, the relevant data can be recorded for subsequent evaluation; hardness test: a microhardness tester is used to test the hardness of the rust removal area and the untreated matrix area. In the rust removal area, multiple test points are selected (such as 9 points distributed according to a 5 mm x 5 mm grid), and the hardness values are measured and averaged. At the same time, the same number and distribution of test points are selected in the untreated area as a comparison. The difference between the hardness values of the rust removal area and the untreated area or the hardness change rate is calculated to evaluate the hardness change of the matrix caused by the laser rust removal process, and to infer the matrix damage degree;

[0110] The actual evaluation results are compared with the model prediction results. If there is a large deviation between the two, the rust removal data of this time is added to the training data set, and the machine learning model is updated and optimized to continuously improve the accuracy and adaptability of the model;

[0111] The specific content of the comparison and analysis of the actual evaluation results and the model prediction results is as follows: the rust removal effect and matrix damage degree data obtained by evaluation are compared with the corresponding results predicted by the machine learning model during the laser rust removal process. The deviation value is calculated. For rust removal rate deviation: ΔRR = |RR 实际 -RR 预测 |, for matrix damage degree deviation ΔBD = |BD 实际 -BD 预测|, set a reasonable deviation threshold, compare the calculated deviation value with the threshold, and determine that the deviation is large, and the machine learning model needs to be updated; if the deviation is within the threshold range, it is considered that the current model prediction effect is good, and the model can be temporarily not updated; when the deviation is large, the related data of this laser rust removal operation (including laser rust removal parameters, metal surface initial characteristics, sensor real-time acquisition data, and actual evaluation of this time rust removal effect and substrate damage data) is added to the training data set. The same preprocessing operations as the original data are performed on the newly added data, such as data cleaning, normalization, etc., to ensure the consistency and usability of the data; the expanded training data set is used to retrain the machine learning model (such as neural network). Reset the training parameters, such as the number of training rounds, learning rate, etc., which can be adjusted appropriately according to experience or experiment. For example, for a neural network model, increase the number of training rounds to 1.5 times the original, and appropriately reduce the learning rate to improve the convergence accuracy of the model. In the training process, the same loss function (such as mean square error function) and optimization algorithm (such as stochastic gradient descent algorithm) as the initial training are used to update and optimize the weights and biases of the model, so that the model can better learn the characteristics and rules in the new data, thereby improving the accuracy of the model in predicting the laser rust removal effect and substrate damage; after retraining is completed, an independent test data set is used to evaluate the performance of the updated model. Calculate the prediction accuracy, mean square error, and other evaluation indicators of the model on the test set, and compare them with the performance indicators of the model before updating to verify whether the model updating is effective. If the performance of the updated model is significantly improved, it can be applied to the subsequent laser rust removal online monitoring and parameter optimization process; if the performance improvement is not obvious or even decreases, further analysis is needed, such as checking data quality, adjusting model structure or training parameters, etc., to retrain and optimize the model until a satisfactory model performance is obtained.

[0112] The present application can accurately establish the complex relationship between the rust removal effect and the laser power, scanning speed, pulse frequency, spot diameter and other parameters, and the metal surface characteristics through the learning of a large amount of experimental data and actual case data by a machine learning model. In the actual rust removal process, based on the real-time data collected by the sensor, the model can accurately predict the current rust removal effect, and dynamically adjust the laser rust removal parameters using an optimization algorithm. This makes the rust removal process no longer rely on empirical parameter settings, but intelligently optimize according to the real-time state of the metal surface, thereby significantly improving the rust removal rate and ensuring that the metal surface can meet higher cleanliness standards. For example, when dealing with metal parts with complex rust conditions and different material mixtures, the model can quickly adapt and find the best parameter combination to effectively remove stubborn rust, while traditional methods often cannot achieve such precise parameter matching, resulting in incomplete rust removal or over-treatment. Adapt to various working conditions: whether it is facing metal surfaces of different materials (such as carbon steel, stainless steel, aluminum alloy, etc.), or dealing with various rust levels (from mild surface oxidation to heavy rust layer accumulation), this method can flexibly adjust the parameter strategy through the adaptive learning ability of the machine learning model, achieving the best rust removal effect. This greatly expands the application range of laser rust removal technology, enabling it to be more widely and efficiently applied in many industrial fields (such as mechanical manufacturing, automobile repair, ship maintenance, etc.), as well as in special industries such as cultural relic restoration and building decoration, solving the poor adaptability problem of traditional rust removal methods in different working conditions;

[0113] While accurately controlling the rust removal parameters, the machine learning model can also predict and monitor the substrate damage in real time. Through comprehensive analysis of the metal surface characteristics, laser parameters, and real-time monitoring data, the model can provide early warning of parameter settings that may cause excessive substrate damage and timely adjustments to control the substrate damage to the minimum. For example, when detecting that the metal surface temperature is too high or the laser energy density is too large, which may cause substrate melting or micro-cracks, the model will automatically adjust the scanning speed or reduce the laser power to avoid damaging the structural integrity and performance of the substrate, while the traditional fixed parameter setting laser rust removal method cannot achieve such precise damage prevention and control; prolong the service life of metal parts: as it can effectively reduce substrate damage, metal parts after rust removal by the present method can maintain better mechanical properties and surface quality, thereby significantly prolonging their service life. In some fields where metal parts reliability and durability are highly required, such as aerospace, high-end equipment manufacturing, etc., this advantage is particularly prominent. Compared with the potential substrate damage hidden dangers that may be caused by traditional rust removal methods, the present application provides a strong guarantee for the long-term stable operation of metal parts, reduces the risk of part failure caused by the rust removal process, and reduces maintenance and replacement costs;

[0114] The close combination of the data collected by the sensor in real time and the machine learning model during the entire laser rust removal process realizes online monitoring and instant feedback of the rust removal effect and substrate damage. This real-time nature enables the operator to understand the progress of the rust removal operation in a timely manner, and once an abnormality occurs (such as the rust removal effect not meeting the expected or the substrate damage increasing), the system can quickly respond and adjust the parameters, avoiding the time delay and inconvenience of manual parameter adjustment and periodic inspection in traditional rust removal methods. This not only improves the overall efficiency of the rust removal operation, but also reduces errors and uncertainties caused by human factors, ensuring the stability and consistency of the rust removal quality; the intelligent control based on machine learning enables the laser rust removal process to be highly automated. From data collection, model prediction, parameter optimization to the final rust removal effect evaluation and model updating, the entire process requires minimal human intervention, only the relevant parameters need to be set and the equipment started in the initial stage, and the system can automatically run and adaptively adjust according to the actual situation. This greatly reduces the labor intensity and technical requirements of the operator, while improving the production efficiency and the standardization of the rust removal operation. In large-scale industrial production or batch rust removal tasks, the advantages of automation are more obvious, which can significantly shorten the production cycle, improve the economic efficiency and market competitiveness of enterprises.

[0115] Although embodiments of the present application have been shown and described, it is to be understood that various modifications, substitutions, replacements and changes can be made to these embodiments without departing from the principles and spirit of the present application, and the scope of the present application is defined by the appended claims and their equivalents.

Claims

1. A machine learning based laser rust removal method, characterized by, Comprising the following steps: Step 1, data collection and preprocessing: Collect a large number of different material metal surface in different laser rust removal parameters, specific parameters including laser power, scanning speed, pulse frequency, spot diameter under the rust removal experimental data, and the rust removal experimental data include the image data of metal surface before and after rust removal, rust degree data, substrate damage data; The collected data is cleaned, normalized and preprocessed to improve the quality and availability of the data; The specific implementation steps are as follows: Step 11, experimental design and sample preparation: Select experimental samples for a variety of common metal materials; For each experimental sample, prepare samples with different corrosion levels, corrosion levels are controlled and adjusted by natural exposure time and artificial accelerated corrosion methods to cover various conditions from mild to severe corrosion; Step 12, image data collection: Use high-resolution industrial cameras or optical microscopes to take pictures of the metal surface before and after rust removal to obtain image data; Step 13, rust degree data collection: Quantify the corrosion degree by chemical analysis, specifically measure the mass loss per unit area of the rusted sample before rust removal to determine the amount of rust, use a high-precision electronic balance to weigh the mass change of the sample before and after rust removal, and calculate the mass thickness of the rust by combining the sample surface area; Step 14, substrate damage data collection: Use a scanning electron microscope to observe the microstructure of the metal substrate after rust removal, check for micro-cracks, fusion layers, and heat-affected zone substrate damage characteristics; For the damage, quantify it by measuring the depth and width of the damage area and calculating the area ratio of the damage area to the entire rust removal area; At the same time, use hardness testing method to compare the change of metal substrate hardness before and after rust removal, indirectly evaluate the substrate damage degree, hardness test uses microhardness tester to measure multiple points in rust removal area and untreated area and takes average value; Step 15, outlier processing: For image data, check if there are fuzzy, missing or obviously wrong pixel areas caused by camera failure or abnormal lighting; For rust degree data and substrate damage data, if the measured value deviates significantly from the normal range, further check the original measurement record or repeat the measurement to determine the accuracy of the data; Step 16, image data normalization: For the image data collected in the above steps, normalize the pixel values of the image, specifically map the pixel values of the image to the [0, 1] interval; Step 17, numerical data normalization: Normalize the rust degree data, substrate damage data and laser rust removal parameter data using Min-Max normalization or Z-score standardization method; Step 2, feature extraction and selection: Extract features related to laser rust removal effect and substrate damage from preprocessed data, including laser parameter features, metal surface features, and surface quality features after rust removal; Use feature selection algorithm to filter out key features that have significant impact on laser rust removal effect and substrate damage, reduce data dimension, and improve model training efficiency; Step 3, model training: select a neural network model, use the features screened in step 2 as input, and the rust removal effect evaluation index as output to train the model; use cross-validation method to adjust the hyperparameters of the model, optimize the performance of the model, so that the model can accurately predict the rust removal effect and substrate damage under different laser parameters; Step 4, online monitoring and parameter optimization: in the actual laser rust removal process, use sensors to collect real-time data of the metal surface, and combine the pre-trained neural network model in step 3 to monitor and predict the current rust removal effect and substrate damage; according to the prediction result, adjust the laser rust removal parameters dynamically through optimization algorithm to achieve the best rust removal effect and minimum substrate damage; Step 5, effect evaluation and model updating: after completing the laser rust removal operation, comprehensively evaluate the rust removal effect, specifically detect the micro morphology and residual rust amount of the metal surface after rust removal through optical microscope and scanning electron microscope equipment; compare and analyze the actual evaluation result with the model prediction result, if there is a large deviation between them, add this time's rust removal data to the training data set, and update and optimize the machine learning model to continuously improve the accuracy and adaptability of the model.

2. The method of claim 1, wherein: The specific steps of feature extraction and selection in step 2 are as follows: Step 21, feature extraction: a. Laser parameter features: directly extract the laser rust removal parameters set in data collection as features; b. Metal surface features: Roughness features: use a roughness measuring instrument to measure the metal surface before rust removal to obtain the surface roughness parameters; Texture features: extract texture features from the image data of the metal surface before rust removal; c. Surface quality features after rust removal: Residual rust amount features: determine the residual rust amount of the metal surface after rust removal through chemical analysis method, which is an improvement of mass loss measurement method, and calculate the difference between the actual mass after rust removal and the ideal mass after rust removal to determine the residual rust amount; Surface flatness features: use a profilometer or optical interferometer to measure the flatness change of the metal surface after rust removal; 3. The method of claim 1, wherein: Step 22, feature selection: principal component analysis, first, calculate the covariance matrix of the feature data in step 21; then, solve the eigenvalues and eigenvectors of the covariance matrix, the eigenvalues represent the variance of the data in the corresponding eigenvector direction; finally, sort the eigenvectors according to the eigenvalues, select the first k eigenvectors, and the eigenvectors constitute the principal component transformation matrix, i.e. the selected key features are the first k eigenvectors. The specific implementation steps of model training in step 3 are as follows: Step 31, neural network model structure selection and initialization: Select a multi-layer feedforward neural network, including input layer, hidden layer and output layer; the number of input layer nodes is equal to the number of selected features, that is, after feature selection, the four key features of laser power, scanning speed, metal surface roughness and residual rust amount are determined, so the number of input layer nodes is 4, the number of output layer nodes is 2, corresponding to the rust removal rate and substrate damage degree in the rust removal effect evaluation index respectively; the number and node number of hidden layer need to be determined according to experience and test; Step 32, forward propagation: let the input feature vector be where is the laser power, is the scanning speed, is the metal surface roughness, is the residual rust amount, for the calculation from the input layer to the hidden layer, the input of the hidden layer node The calculation formula is: ； wherein is the weight of the input layer's th node to the hidden layer's th node, is the bias of the hidden layer's th node; Output of hidden layer nodes The calculation is performed using an activation function, here the commonly used sigmoid activation function: For the calculation from the hidden layer to the output layer, the input to the first node of the output layer is calculated as follows: ； in It is the hidden layer. The node to the output layer The weight of each node, It is the bias of the i-th node in the output layer; Output of the output layer node Similarly, the activation function is also used. The sigmoid function is used to map the output of the rust removal rate to the interval [0, 1], which represents the proportion of rust removal; the linear activation function is used for the matrix damage degree, and the output is a numerical value, which represents the quantitative value of the damage degree; Step 33, loss function calculation: the mean square error is used as the loss function to measure the difference between the model prediction value and the true value, set The true value of the first The actual measured rust removal rate and substrate damage degree of the output, i.e. The model prediction value, the loss function The calculation formula is: ; Step 34, back propagation and weight update: the gradient of the loss function with respect to the weights and biases of each layer is calculated by the back propagation algorithm, and then the weights and biases are updated according to the gradient descent algorithm, wherein the weights from the hidden layer to the output layer The update formula is: ； ； where is the learning rate, is a hyperparameter that controls the step size of the weight updates, is the gradient of the loss function with respect to , which is computed step by step from the output layer to the hidden layers by the chain rule. Step 35, cross-validation adjustment of hyperparameters: the collected laser rust removal experimental data is divided into subsets, each time selecting subsets as the training set, and the remaining 1 subset as the validation set; select the hyperparameter combination with the minimum loss function value on the validation set as the final model hyperparameters, and then retrain the model under the hyperparameter combination using the entire training data set to obtain the final trained neural network model.

4. The method of claim 1, wherein: The specific implementation content of the step 4 online monitoring and parameter optimization is as follows: A, sensor data acquisition and transmission Laser parameter sensor: install high-precision laser power sensor on the laser rust removal equipment to monitor the laser output power in real time; at the same time, install scanning speed sensor to accurately measure the moving speed of laser scanning head through encoder technology; Surface state sensor: use infrared temperature sensor to measure the real-time temperature change of metal surface under the action of laser; at the same time, use optical reflectivity sensor to detect the reflected light intensity after laser irradiation to metal surface, the reflected light intensity will be different under different rust degree and metal substrate state, through monitoring the reflectivity, the rust removal progress can be judged; B, data preprocessing and feature extraction Real-time preprocessing of collected sensor data, including removing noise interference; Extract real-time features corresponding to model input features, directly get current laser power value from laser power sensor data, get scanning speed value from scanning speed sensor, extract metal surface real-time temperature features from infrared temperature sensor data, and extract reflectivity features from optical reflectivity sensor data, the above features will be used as input for subsequent model prediction; C, prediction based on machine learning model: input the real-time feature data after preprocessing and extraction into the pre-trained neural network model in step 3, the model can predict the current rust removal rate and substrate damage degree according to the input feature data and the relationship between the learned features and the rust removal effect and substrate damage in the training stage; D, optimization algorithm and parameter adjustment Optimization goal setting: set the optimization goal as improving the rust removal rate while ensuring that the substrate damage degree is lower than the set threshold; Optimization algorithm selection and application: genetic algorithm is used for parameter optimization, first, encode laser power, scanning speed, metal surface roughness and residual rust amount parameters into chromosome form, then initialize a population containing multiple such encoded parameter combination individuals, calculate the predicted rust removal rate and substrate damage degree corresponding to each individual, set fitness function according to optimization goal, and through selection, crossover and mutation genetic operations, constantly generate new parameter combination individuals to find the parameter combination with the highest fitness. E、Parameter adjustment implementation: send the optimized laser rust removal parameters to the controller of the laser rust removal equipment, the controller adjusts the power output of the laser generator and the rotating speed of the scanning motor according to the received parameter instructions, realizes the dynamic adjustment of the laser rust removal parameters, after adjusting the parameters, continue to collect data through the sensor, repeat the above monitoring, prediction and optimization process until the satisfactory rust removal effect and the substrate damage are within the controllable range.

5. The machine learning based laser rust removal method of claim 1, wherein: The specific content of the comparison and analysis between the actual evaluation result and the model prediction result in step 5 is as follows: the rust removal effect and the substrate damage degree data obtained by evaluation are compared with the corresponding results predicted by the machine learning model in the laser rust removal process, and the deviation value is specifically calculated. For the rust removal rate deviation: For the substrate damage degree deviation A reasonable deviation threshold is set, and the calculated deviation value is compared with the threshold to determine whether the deviation is large, and the machine learning model needs to be updated; if the deviation is within the threshold range, it is considered that the current model prediction effect is good, and the model is not updated temporarily.

Images