A recommendation method for steel structure welding process quality management based on big data processing

Abstract

The present application discloses a steel structure welding process quality management recommendation method based on big data processing, which relates to the field of welding process technology, including: marking image data to generate a labeled data set; preprocessing the labeled data set; extracting key features based on the preprocessed data set; clustering analysis of the feature data set to obtain the process mode under different welding parameters and environmental conditions; establishing a welding model through transfer learning based on the preprocessed data set and the feature data set to perform welding defect detection; building a user portrait based on the user's historical behavior and preferences; establishing a recommendation model based on the user portrait, process mode and welding defect detection results to recommend welding processes; optimizing the recommendation results based on the identified process mode and the recommended welding process through a multi-armed tiger strategy. In view of the problem that welding processes in the prior art lack personalized recommendations, the present application improves the personalization of welding processes.

Description

Technical Field

[0001] The present application relates to the field of welding process technology, and in particular to a steel structure welding process quality management recommendation method based on big data processing. Background Art

[0002] In recent years, the rise of new generation information technologies such as big data, the Internet of Things, and artificial intelligence has provided a new way for intelligent welding process quality management. By collecting and analyzing various sensor data during the welding process and using machine learning algorithms to establish quality prediction and defect diagnosis models, online monitoring and intelligent control of welding quality can be achieved, effectively improving welding quality and production efficiency.

[0003] However, the welding process is complex, involving multiple factors such as materials, processes, equipment, and environment. The heterogeneity and dynamic nature of data make modeling and analysis difficult. Different welding tasks have different requirements for process parameters, and the lack of an effective process knowledge base and reasoning mechanism makes it difficult to achieve rapid process optimization. Existing methods mainly focus on general process planning, lack personalized services for specific user needs, and cannot give full play to the advantages of intelligent technology.

[0004] In the related art, for example, Chinese patent document CN115858940A provides a method for recommending quality management of steel structure welding process based on big data processing, which relates to the technical field of quality management recommendation of steel structure welding process. The present invention generates quality characteristic values ​​by processing and analyzing big data, judges the stability of steel structure welding in extreme environment by the size of quality characteristic values, and generates corresponding evaluation quality text by single quality characteristic value, so as to remind staff or researchers of the advantages and disadvantages of corresponding welding process standard technology, and then processes and screens multiple quality characteristic values ​​and generates optimal welding process recommendation text in combination with evaluation quality text, so as to realize the quality test analysis of multiple welding process standard technologies in big data in an intuitive way, so as to assist staff or researchers to efficiently manage and accurately recommend welding process standard technology, so as to help staff or researchers to formulate unified rules in accordance with standardized production. However, the scheme may not be able to fully express the complexity and diversity of welding process by generating corresponding evaluation quality text by single quality characteristic value. This may lead to oversimplification of the evaluation of welding process, and fail to provide sufficient information to support staff or researchers to make personalized decisions. Summary of the invention

[0005] 1. Technical issues to be solved

[0006] In response to the problem of lack of personalized recommendation for welding processes in the prior art, the present application provides a steel structure welding process quality management recommendation method based on big data processing, which improves the personalization of welding processes through feature selection and construction of user portraits.

[0007] 2. Technical solution

[0008] The purpose of this application is achieved through the following technical solutions.

[0009] The embodiment of the present specification provides a steel structure welding process quality management recommendation method based on big data processing, including: acquiring image data and real-time data in the welding process, the real-time data including welding current, voltage and speed; marking the image data to mark the location and type of welding defects; synchronizing the marked image data with the corresponding real-time data to generate a labeled data set; preprocessing the labeled data set, the preprocessing including image enhancement to generate a preprocessed data set; extracting key features based on the preprocessed data set, the key features including material type, thickness and welding method to generate a feature data set; performing cluster analysis on the feature data set to obtain process modes under different welding parameters and environmental conditions; establishing a welding model through transfer learning based on the preprocessed data set and the feature data set to perform welding defect detection; building a user portrait based on the user's historical behavior and preferences; establishing a recommendation model based on the user portrait, process mode and welding defect detection results to recommend a welding process; optimizing the recommendation results through a multi-armed bandit strategy based on the identified process mode and the recommended welding process.

[0010] Among them, welding defects refer to various discontinuities, imperfections or inconsistencies in welding joints or welds that do not meet quality requirements, which will reduce the performance of welding joints and may cause structural failure in severe cases. Common types of welding defects include cracks, incomplete penetration, slag inclusions, pores, weld nodules, etc. Defect location refers to the spatial coordinates or area of ​​welding defects in weld images. Rectangular frames, polygons, etc. are usually used to mark the approximate location and range of defects. Defect type refers to the classification of detected welding defects according to characteristics such as morphology, size, and cause. The definition and naming of defect types in different standards may vary, but generally include common defects such as cracks, incomplete penetration, slag inclusions, pores, weld nodules, etc. In this application, the annotation link is that experienced welding engineers or quality inspectors observe and analyze the collected welding images frame by frame to determine whether there are defects in the image. If a defect is identified, the approximate location of the defect is marked with a rectangular frame, and the corresponding type label is selected from the predefined defect type list. The annotation results are usually stored in formats such as XML and JSON. Dataset construction: Match the annotated defect information with the original image data and the synchronously collected real-time welding parameters (such as current, voltage, and speed) to construct a labeled welding process dataset. Each sample in the dataset contains three parts: image, real-time parameters, and defect annotations, which can be used for subsequent machine learning model training and testing. Defect detection model training: Divide the constructed labeled dataset into training set and test set according to a certain ratio. Use the training set samples to train the defect detection model. The model can use target detection algorithms such as YOLO and SSD, input welding images, and output the position coordinates and types of each defect in the image. Through loss function optimization, the model can automatically identify and locate welding defects in the image. Defect detection model evaluation: Use the test set samples to evaluate the performance of the trained defect detection model, calculate the accuracy, precision, recall rate and other indicators of the model in defect location and classification, and evaluate the defect detection ability of the model. Online defect detection application: Deploy the trained defect detection model to the welding production site to perform automatic defect detection on the real-time collected welding images. The model infers the input image and outputs the location and type information of the detected defects, realizing automatic assessment of weld quality and defect warning.

[0011] Specifically, a welding model is established through transfer learning for welding defect detection. Select the source domain model: select a deep learning model pre-trained on a large-scale general dataset as the source domain model, such as ResNet, VGG Net, etc. trained on the Image Net dataset. These models have learned rich image feature representations and can serve as the basis of the welding defect detection model. Freeze the bottom layer of the source domain model: freeze the parameters of the bottom convolution layer of the source domain model to keep its feature extraction capability unchanged. This can reduce the amount of training parameters and reduce the risk of overfitting, while utilizing the general image features learned by the source domain model. Add a task adaptation layer: add several new convolutional layers and fully connected layers on top of the source domain model to build an adaptation layer for the welding defect detection task. These layers will learn the mapping from source domain features to welding defect features to adapt the model to the new task. Train the task adaptation layer: use the pre-processed welding process dataset to train the task adaptation layer, input the welding images in the training set, calculate the output defect location and type through forward propagation, then evaluate the difference between the predicted result and the labeled value through the loss function, and use the back propagation algorithm to update the parameters of the adaptation layer so that the model gradually fits the welding defect detection task. Commonly used loss functions include multi-task loss and focal loss, which can balance the learning of defect location and classification. Fine-tuning the source domain model: After the adaptation layer training reaches a certain level, some high-level parameters of the source domain model can be selectively unfrozen and fine-tuned with a smaller learning rate. In this way, while maintaining the source domain feature representation, the model can better adapt to the distribution of welding defect data and improve the detection accuracy. Model evaluation: During the training process, the validation set data is regularly used to evaluate the model performance and monitor the convergence of the model. By calculating the accuracy, precision, recall rate and other indicators of defect location and classification, the model parameters with the best performance are selected and saved. Defect detection application: The trained welding defect detection model is deployed to the production environment for online detection. After preprocessing the collected welding images, they are input into the model for forward reasoning, and the location coordinates and types of each welding defect in the image are output. According to the preset defect thresholds and rules, the weld is judged to be qualified, and the defect information is fed back to the operator or the automatic control system to achieve real-time quality monitoring.

[0012] Specifically, construct user portraits and determine user feature dimensions: According to the business characteristics of welding process quality management, select key feature dimensions that can characterize user behavior and preferences. Common feature dimensions include: User roles: such as welding engineers, quality inspectors, production managers, etc. Different roles may have different concerns and decision preferences for welding processes. User experience: Considering the number of years the user has been engaged in welding-related work, senior users and novice users may have differences in process selection and parameter settings. Welding process preferences: Statistics on the user's historical process parameter selection under different materials, plate thickness, welding methods and other conditions to characterize their process preferences. Defect attention: Analyze the frequency of users' attention to different types of defects and the way they deal with them, reflecting their emphasis and standards on welding quality. Interactive behavior: Statistics on the frequency, duration, operation sequence, etc. of users using the welding process quality management system, and explore their behavior patterns and habits. Collect user behavior data: Extract and collect users' historical behavior data from channels such as logs and databases of the welding process quality management system. Mainly includes: Basic information of users: such as roles, departments, years of work, etc. Process parameter selection records: including information such as materials, process parameters, equipment, etc. selected by users in different welding tasks. Defect feedback data: defect annotation, evaluation, feedback from users on welding results, etc. Interaction behavior log: operation records of users logging in, querying, modifying parameters, generating reports, etc. in the system. Data cleaning and preprocessing: clean and preprocess the collected user behavior data, remove noise data such as outliers and missing values, and convert the data into a format suitable for analysis. Data filling, data transformation and other technologies can be used to improve data quality. Feature engineering: extract and select features from the preprocessed user behavior data. Design appropriate feature representation methods for different feature dimensions. For example, for discrete features such as user roles and experience, one-hot encoding, serial number encoding and other methods can be used to represent them. For welding process preferences, the frequency of users selecting different materials and process parameters can be counted to generate preference vectors. For the degree of defect attention, the proportion of users annotating and handling different defect types can be calculated to generate attention weights. For interactive behavior, the user's operation sequence, click count, etc. can be extracted to construct behavioral features. User portrait construction: the extracted user features are combined according to the predefined portrait template to generate a structured user portrait. Portraits are usually represented in the form of dictionaries, vectors, graphs, etc., and contain the characteristic values ​​of users in various dimensions. You can consider using technologies such as user clustering and association rule mining to discover the common characteristics and behavior patterns of user groups and enrich the portrait content.

[0013] Among them, the Multi-Armed Bandit (MAB) strategy is a machine learning algorithm for making decisions in an uncertain environment. It is inspired by the scene of playing slot machines. Players face slot machines with multiple arms (i.e., rods), each arm has a different probability distribution of benefits, and players need to find the arm with the highest benefit through repeated attempts. In the field of machine learning, multi-armed bandit strategies are often used to solve online learning and decision-making problems, optimizing the decision-making process by balancing exploration and exploitation, with the goal of maximizing overall benefits. In this application, candidate processes are used as arms: the identified process modes and recommended welding process parameter combinations are regarded as different arms in a multi-armed bandit. Each process combination corresponds to a bandit arm, and its expected benefit is the historical performance score of the process combination, such as welding quality score, user acceptance, etc. Initialize the benefit estimate of the arm: For newly identified process modes or process combinations recommended for the first time, since there is no historical performance data, its initial benefit estimate can be set to a higher default value to encourage the algorithm to explore new process combinations in the early stages. Arm selection strategy: Each time a welding process needs to be recommended, the multi-armed bandit algorithm selects an arm (i.e., process combination) as the recommendation result based on the current estimated benefits of each process combination. Common arm selection strategies include: ε-greedy strategy: randomly select an arm for exploration with probability ε, and select the arm with the highest current estimated benefit for use with probability 1-ε. Upper confidence bound (UCB) strategy: select the arm with the largest weighted sum of the current benefit estimate and uncertainty, and encourage exploration of arms that have been selected less frequently. Thompson sampling: sample according to the posterior distribution of each arm, and select the arm with the largest sampling value as the recommendation result.

[0014] Specifically, by applying the multi-armed bandit strategy to the optimization of welding process recommendations, the quality of recommendations and user satisfaction can be further improved on the basis of process pattern recognition and initial recommendations. The algorithm dynamically adjusts the recommendation strategy by balancing the exploration of new process combinations and the use of historical optimal processes to adapt to the changing user needs and task environment. At the same time, the online learning characteristics of the multi-armed bandit strategy enable it to efficiently process large-scale process parameter combination spaces and achieve real-time recommendation optimization.

[0015] Furthermore, image enhancement includes: dividing the labeled data set into multiple local windows; calculating the cumulative distribution function CDF of the pixel grayscale value in each local window, and mapping the CDF to a specified dynamic range, where the grayscale value range after mapping is 0 to 255, and the mapping function is:

[0016] Where i is the original gray value, CDF(i) is the cumulative distribution function value corresponding to the gray value i, CDF_min and CDF_max are the minimum and maximum values ​​of CDF in the local window, respectively, and round is the rounding function. The bilinear interpolation method is used to smoothly transition the mapping results of adjacent local windows, and the interpolation coefficient is calculated according to the overlap rate of the local windows. The calculation formula is:

[0017] coefficient=1-overlap_ratio, where window_size is the size of the local window, stride is the step size between adjacent local windows, overlap_ratio is the overlap ratio of the local windows, and coefficient is the interpolation coefficient.

[0018] Among them, the Cumulative Distribution Function (CDF) is a function used to describe the probability distribution of a random variable. For a random variable X, its CDF is defined as F(x) = P(X≤x), which represents the probability that the variable X takes a value less than or equal to x. In the field of image processing, the grayscale value of an image can be regarded as a random variable, and its CDF describes the cumulative probability distribution of the grayscale value. In this application, the distribution of pixel grayscale values ​​in a local window is calculated: the image is divided into several local windows, and the grayscale values ​​of the pixels in each window are counted to obtain a probability distribution histogram of the grayscale values. Calculate the CDF in the local window: According to the probability distribution histogram of the grayscale value, calculate the CDF value corresponding to each grayscale value. The specific steps are as follows: Normalize the probability distribution histogram of the grayscale value to obtain the probability P(X=j) of each grayscale value. Accumulate the normalized histogram to obtain the CDF value of each grayscale value. For example, CDF(0)=P(X=0), CDF(1)=P(X=0)+P(X=1), and so on. A CDF mapping table corresponding to the grayscale value is obtained, indicating the cumulative probability of each grayscale value.

[0019] Specifically, by applying locally adaptive CDF mapping, the above scheme can adaptively adjust the dynamic range of pixel grayscale values ​​according to the grayscale value distribution characteristics in the local window, thereby enhancing the contrast and detail expression of the image. CDF mapping can normalize the grayscale value distribution in the local window to the specified dynamic range, making the grayscale value distribution in different windows more consistent while retaining the local contrast information. Compared with the globally unified grayscale value mapping, the locally adaptive CDF mapping can better adapt to the grayscale value distribution differences in different areas of the image, providing a more delicate and natural enhancement effect.

[0020] Among them, bilinear interpolation is a commonly used image interpolation algorithm used to estimate the value of an unknown point between two known points. It calculates the value of the unknown point by performing linear interpolation in two directions, thereby achieving smooth transition and scaling of the image. Bilinear interpolation considers the four nearest neighboring pixels around the unknown point and calculates the interpolation result of the unknown point based on their grayscale values ​​and relative positions. Assume that four pixels Q are known. 11 , Q 12 , Q 21 and Q 22 , their coordinates are (x1,y1), (x1,y2), (x2,y1) and (x2,y2), and their grayscale values ​​are f(Q 11 )、f(Q 12 )、f(Q 21 ) and f(Q 22 ). The coordinates of the unknown point P are (x, y), and its interpolation result f(P) can be calculated by the following steps: Linear interpolation is performed in the x direction to obtain the grayscale values ​​of points P1 and P2: Perform linear interpolation in the y direction to obtain the final grayscale value of point P: In this application, the overlapping area is determined by calculating the overlapping area between adjacent local windows based on the size of the local window (window_size) and the stride between adjacent windows. The overlap ratio (overlap_ratio) represents the proportion of the overlapping area in the local window, and the calculation formula is: Calculation of interpolation coefficient: According to the overlap ratio (overlap_ratio), calculate the interpolation coefficient (coefficient) of bilinear interpolation. The interpolation coefficient represents the weight of the mapping result of the current window in the overlapping area. The calculation formula is: coefficient = 1-overlap_ratio. The value range of the interpolation coefficient is [0, 1]. When the overlap ratio is 0, it means that the current window completely covers the overlapping area, and the interpolation coefficient is 1; when the overlap ratio is close to 1, it means that the current window and the adjacent window almost completely overlap, and the interpolation coefficient is close to 0. Interpolation of overlapping areas: For the overlapping area of ​​each local window, use the bilinear interpolation method to smoothly transition the mapping results of adjacent windows. The specific steps are as follows: Determine the coordinates (x, y) of the unknown pixel point P in the overlapping area. Find the four nearest neighboring pixel points Q around point P. 11 , Q 12 , Q 21 and Q 22, which are located in the mapping results of the current window and the adjacent window respectively. According to the bilinear interpolation formula, the interpolation results f(P1) and f(P2) of point P in the x direction and y direction are calculated. Multiply f(P1) and f(P2) by the corresponding interpolation coefficients coefficient respectively to obtain the weighted values ​​of point P in the mapping results of the current window and the adjacent window. Add the weighted values ​​to obtain the final interpolation result of point P as the smooth transition value in the overlapping area. Repeat interpolation: Repeat step 3 for the overlapping area of ​​each local window until the overlapping areas of all windows are interpolated and smoothed. By applying bilinear interpolation to the overlapping areas of adjacent local windows, the above scheme can achieve a smooth transition between local mapping results and eliminate the discontinuity of the window edges. The calculation of the interpolation coefficient takes into account the size of the overlapping area, so that the larger the overlap rate, the greater the weight of the mapping results of the adjacent windows, and the smooth transition is more natural. The bilinear interpolation method can interpolate in two directions at the same time to obtain a continuous and smooth grayscale value transition, thereby improving the visual quality of the enhanced image. Compared with simple linear interpolation, bilinear interpolation can better preserve local details and edge information and reduce aliasing and distortion during the interpolation process.

[0021] Furthermore, image enhancement also includes: calculating the contrast and entropy value of the pixel gray value in each local window, and the contrast calculation formula is: Among them, max_gray and min_gray are the maximum and minimum grayscale values ​​of pixels in the local window respectively; the entropy value calculation formula is:

[0022] entropy=-∑(p×log2(p)), where p is the probability of occurrence of each gray value in the local window; the size and overlap rate of the local window are adjusted according to the contrast and entropy value; when the contrast is less than a preset contrast threshold and the entropy value is less than a preset entropy threshold, the size of the local window is increased and the overlap rate is reduced; when the contrast is greater than the preset contrast threshold and the entropy value is greater than the preset entropy threshold, the size of the local window is reduced and the overlap rate is increased; the mapping results and interpolation results of all local windows are combined to obtain contrast-enhanced image data.

[0023] Specifically, based on the judgment result, dynamically adjust the size and overlap ratio of the local window. When the window size needs to be increased, the current window size can be multiplied by a coefficient greater than 1, such as 1.2 or 1.5 times. When the overlap ratio needs to be reduced, the overlap ratio of the current window can be multiplied by a coefficient less than 1, such as 0.8 or 0.5 times. When the window size needs to be reduced, the current window size can be multiplied by a coefficient less than 1, such as 0.8 or 0.5 times. When the overlap ratio needs to be increased, the overlap ratio of the current window can be multiplied by a coefficient greater than 1, such as 1.2 or 1.5 times. Recalculate the interpolation coefficient: recalculate the interpolation coefficient between adjacent windows based on the adjusted window size and overlap ratio. The calculation formula of the interpolation coefficient is: coefficient = 1-overlap_ratio, where, Apply the adjusted window and interpolation coefficient: Use the adjusted window size and overlap rate to perform adaptive histogram equalization on the local area, and use the updated interpolation coefficient to perform bilinear interpolation on the mapping results of adjacent windows to obtain a smooth transition enhancement result. Combine the results of all local windows: Splice the mapping results and interpolation results of all local windows together to obtain the final contrast-enhanced image. By dynamically adjusting the size and overlap rate of the local windows, the above scheme can adaptively select appropriate window and overlap settings according to the contrast and information characteristics of different regions of the image. Using a larger window and a smaller overlap rate in areas with low contrast and little information can enhance the local contrast and reduce the noise caused by over-enhancement; using a smaller window and a larger overlap rate in areas with high contrast and large information can retain more detail information and improve the accuracy of enhancement. At the same time, by recalculating the interpolation coefficient, the continuity of the smooth transition between adjacent windows is guaranteed. This adaptive window and overlap rate adjustment strategy can improve the overall contrast of the image while taking into account the enhancement effects of different regions, and obtain a more natural and detailed enhancement result.

[0024] Furthermore, cluster analysis is performed on the feature data set, including: normalizing the feature data set using the minimum-maximum normalization method, scaling the value range of each feature in the feature data set to the interval between 0 and 1, and obtaining a normalized feature data set; dividing the normalized feature data set into a training set and a validation set, wherein the training set is used for feature selection, and the validation set is used to evaluate the clustering performance of the feature set; initializing the candidate feature subset to an empty set, initializing the optimal feature subset to the full set, and using the normalized feature data set as the initial candidate feature subset; repeating the following steps until the size of the optimal feature subset reaches a preset feature quantity threshold k: initializing the candidate feature subset to a full set; and For each candidate feature in the set, add it to the current optimal feature subset to obtain a new feature subset; use the new feature subset to train the clustering model on the training set to obtain the clustering result; use the trained clustering model to perform clustering on the validation set to evaluate the clustering performance; select the feature that improves the clustering performance the most after adding it, remove it from the candidate feature subset, add it to the optimal feature subset, and obtain a new optimal feature subset and a new candidate feature subset; if the clustering performance of the current candidate feature subset is better than the optimal feature subset, update the current candidate feature subset to the optimal feature subset; output the optimal feature subset, which contains the top k features that have the greatest impact on the clustering results.

[0025] Furthermore, cluster analysis is performed on the feature data set, which also includes: calculating the kernel density estimation of the optimal feature subset, wherein the kernel density estimation adopts the Gaussian kernel function, and the bandwidth parameter of the kernel density estimation is determined by the Silverman's rule of thumb method; according to the result of the kernel density estimation, the data point with the highest density is selected as the first cluster center; and the distance between each data point in the optimal feature subset and the first cluster center is calculated, and the distance calculation formula is: Among them, x is the eigenvector of the data point, c is the eigenvector of the cluster center, and Σ is the covariance matrix of the feature data set; according to the calculated distance, the probability of each data point being selected as the next cluster center is calculated, and the probability calculation formula is: Where d(x) is the distance between the data point x and the selected cluster center. According to the calculated probability, a new data point is randomly selected as the next cluster center using the Roulette Wheel Selection method. Repeat the above steps until k cluster centers are selected.

[0026] Among them, Silverman's rule of thumb is a commonly used method for selecting bandwidth for kernel density estimation.

[0027] Wherein, h is the bandwidth, n is the number of data points, and σ is the standard deviation of the data. This formula assumes that the data follows a Gaussian distribution and has good performance in terms of mean square error. Silverman's rule of thumb can automatically adapt to the scale of the data, provide a reasonable initial value of the bandwidth, and reduce the computational overhead of bandwidth selection. In this application, a Gaussian kernel function is used when calculating the kernel density estimate of the optimal feature subset. In order to determine the bandwidth parameter of the kernel density estimation, the Silverman's rule of thumb method is used. Calculate the standard deviation σ of the optimal feature subset. According to the number of data points n and the standard deviation σ of the feature subset, the bandwidth h is calculated using the Silverman's rule of thumb formula. The calculated bandwidth h is used as the bandwidth parameter of the Gaussian kernel function for kernel density estimation. Automatically determining the appropriate bandwidth through Silverman's rule of thumb can reduce the workload of manual parameter adjustment and improve the efficiency and repeatability of kernel density estimation.

[0028] Among them, Roulette Wheel Selection is a commonly used probability-based random selection method. This method regards candidate objects as sectors of a roulette wheel, and the area of ​​each sector is proportional to the probability of selection of the object. When selecting, a point is randomly drawn on the roulette wheel, and the corresponding object is selected in which sector it falls. In this application, when selecting a cluster center, first select the data point with the highest density as the first cluster center based on the result of kernel density estimation. Then, calculate the distance between each data point and the selected cluster center, and calculate the probability of the data point being selected as the next cluster center based on the distance. Using the Roulette Wheel Selection method, randomly select the next cluster center based on the probability distribution of the data point. Calculate the distance d(x) between each data point and the selected cluster center. Based on the distance d(x), calculate the probability p(x) of each data point being selected as the next cluster center. Normalize the probability p(x) of the data point to construct the probability distribution of the roulette wheel. Generate a random number in the range of [0, 1], and select the next cluster center based on the random number using the Roulette Wheel Selection method. Repeat the above steps until k cluster centers are selected. Roulette Wheel Selection is used to select cluster centers, which can avoid the local optimal problem that may be caused by the greedy strategy. According to the distance distribution between the data points and the selected centers, the selection probability is dynamically adjusted, so that data points with farther distances have a higher chance of being selected as new cluster centers, which increases the diversity and globality of the clustering results.

[0029] Specifically, Silverman's rule of thumb is used to automatically determine the bandwidth parameter of kernel density estimation to improve the efficiency and repeatability of estimation; Roulette Wheel Selection is used to randomly select cluster centers according to the probability distribution of data points to avoid local optimality and improve the diversity of clustering results. The application of these two methods in the above scheme can adaptively adjust the parameters of kernel density estimation and cluster center selection, and improve the performance and robustness of clustering algorithms.

[0030] Furthermore, cluster analysis is performed on the feature data set, which also includes: calculating the membership degree of each data point in the optimal feature subset to each cluster center. The membership degree calculation formula is: Among them, c i is the i-th cluster center, m is the coefficient, which controls the hardness or softness of clustering; according to the degree of membership of the data point, the position of the cluster center is updated, and the formula for updating the position of the cluster center is: Among them, u(x,c) is the membership of data point x to cluster center c, x is the eigenvector of the data point, and c is the eigenvector of the cluster center. Repeat the above steps until the change of the cluster center is less than the preset threshold or the maximum number of iterations is reached.

[0031] Furthermore, the clustering performance is evaluated, including: using the following formula to evaluate the compactness and separation of the clustering results: Where a(i) is the average distance between data point i and other data points in the same category, b(i) is the average distance between data point i and the nearest data points in other categories; s(i) ranges from -1 to 1, and the larger the s(i) value, the better the compactness and separation of the clustering results; the following formula is used to evaluate the overall performance of the clustering results: Among them, SS_b is the sum of squares of deviations between categories, SS_w is the sum of squares of deviations within categories, k is the number of clustered categories, and n is the total number of data points. The larger the CH value, the greater the difference between categories of the clustering results, the smaller the difference within the category, and the better the clustering performance.

[0032] Among them, compactness measures the closeness between data points in the same category in the clustering results. The higher the compactness, the closer the data points in the same category are, and the smaller the difference within the category.

[0033] The ideal clustering result should make the data points in the same category highly similar, forming a compact cluster structure. Among them, the separation degree measures the degree of distinction between different categories in the clustering results. The higher the separation degree, the greater the distance between different categories and the clearer the boundaries between categories. The ideal clustering result should make different categories clearly distinguishable, with low overlap or confusion between them. In this application, for each data point i, the average distance a(i) between it and other data points in the same category is calculated. For each data point i, the average distance b(i) between it and the nearest data point in other categories is calculated. Use the formula Calculate the compactness and separation index of data point i. Average the s(i) values ​​of all data points to get the average compactness and separation of the entire clustering result. The value range of s(i) is [-1, 1]. The larger the value, the better the compactness and separation of the clustering result at this data point. By calculating the average s(i) value of all data points, the overall compactness and separation performance of the clustering result can be evaluated.

[0034] Among them, the overall performance comprehensively considers the compactness and separation of the clustering results to evaluate the overall effect of the clustering algorithm. A good clustering result should have high separation while maintaining high compactness. The overall performance index quantifies the pros and cons of the clustering results by comparing the differences between categories and the differences within categories. The higher the overall performance, the more effective the clustering algorithm can be in dividing the data points into compact and well-separated categories. In this application, the sum of squared deviations SS_b between categories is calculated. SS_b measures the distance between the center points of different categories and the global center point, reflecting the differences between categories. The sum of squared deviations SS_w within the category is calculated. SS_w measures the distance between the data points and the center point of the category in the same category, reflecting the differences within the category. Use the formula Calculate the Calinski-Harabasz index. Where k is the number of clustered categories and n is the total number of data points. The larger the CH value, the greater the difference between the categories of the clustering results, the smaller the difference within the category, and the better the overall performance. By comparing the CH values ​​under different clustering algorithms or different parameter settings, you can select the clustering result with the best performance.

[0035] Specifically, compactness and separation are used to evaluate the performance of clustering results at the level of a single data point, measuring the compactness within the same category and the degree of distinction between different categories; the overall performance uses the Calinski-Harabasz index to evaluate the overall effect of the clustering results, taking into account the differences between categories and the differences within categories. By calculating these indicators, the performance of the clustering algorithm can be quantitatively evaluated, the optimal clustering result can be selected, and the clustering algorithm can be optimized and improved. In practical applications, appropriate evaluation indicators can be selected to evaluate clustering performance based on the characteristics and requirements of specific problems, and multiple indicators can be combined for comprehensive analysis and judgment.

[0036] Furthermore, a recommendation model is established in combination with user portraits, process modes and welding defect detection results to make welding process recommendations, including: constructing a ternary relationship matrix R of user-process mode-defect detection results, where the element r(u,i,j) in the matrix R represents the defect detection result j corresponding to the welding operation of user u under process mode i, and r(u,i,j) takes a value of 1 to indicate that the operation result of the user under the corresponding process mode is j, and takes a value of 0 to indicate that it is unknown or has not occurred; decomposing the constructed ternary relationship matrix R to obtain the user latent factor matrix P, the process mode latent factor matrix Q and the defect detection result latent factor matrix Y, and the decomposition formula is: R≈μ+B u +B i +B j +P×Q T +Y, where R is an m×n×l-dimensional ternary relationship matrix, m is the number of users, n is the number of process modes, and l is the number of defect detection result categories; P is an m×k-dimensional user latent factor matrix, k is the latent factor dimension; Q is an n×k-dimensional process mode latent factor matrix; Y is an l×k-dimensional defect detection result latent factor matrix; B u is the m-dimensional user bias matrix, B i is the n-dimensional process mode bias matrix, B j is the l-dimensional defect detection result bias matrix; μ is the global bias term.

[0037] Specifically, the formula introduces the user latent factor matrix P, the process mode latent factor matrix Q and the defect detection result latent factor matrix Y, which represent the representation of the user, process mode and defect detection result in the latent factor space respectively. The latent factor representation maps the user, process mode and defect detection result into a low-dimensional latent semantic space to capture the potential relationship and mutual influence between them. Through the latent factor representation, the implicit patterns and commonalities between users, process modes and defect detection results can be mined, which is convenient for similarity measurement and association analysis. The user latent factor matrix P and the user bias matrix B in the formula uThe user's personalized information is introduced. The user latent factor matrix P represents the position of each user in the latent factor space, capturing the user's personal preferences and characteristics. The user bias matrix B u It represents the inherent bias of each user and characterizes the individual differences of users. Through personalized modeling, it is possible to make personalized process recommendations and defect predictions based on the user's historical behavior and personal characteristics, thereby improving the accuracy of recommendations and user satisfaction.

[0038] The process mode latent factor matrix Q and process mode bias matrix B in the formula i , defect detection result latent factor matrix Y and defect detection result bias matrix B j Represent the latent features and biases of process mode and defect detection results, respectively. The process mode latent factor matrix Q and defect detection result latent factor matrix Y map process mode and defect detection results into latent factor space to capture their latent attributes and relationships. The process mode bias matrix B i And the defect detection result bias matrix B j Characterizes the inherent biases in process patterns and defect detection results.

[0039] By performing representation learning on process modes and defect detection results, we can discover the association patterns between them, predict the types of defects that may occur under a given process mode, and optimize process parameters and improve welding quality based on historical defect detection results. The global bias term μ in the formula represents the global average or baseline level of the entire data set. The global bias term characterizes the overall trend and general rules of the ternary relationship between user, process mode, and defect detection results. It provides a benchmark reference that enables the model to better fit the data and provide reasonable default recommendations when personalized information is lacking.

[0040] The formula decomposes the ternary relationship matrix R into the product of the user latent factor matrix P, the process mode latent factor matrix Q and the defect detection result latent factor matrix Y, plus the bias matrix and the global bias.

[0041] Through matrix decomposition, the high-dimensional ternary relationship matrix is ​​mapped to a low-dimensional latent factor space to achieve low-rank approximation. This low-rank approximation can effectively filter noise, capture the main patterns and structures in the data, and improve the robustness and generalization ability of recommendations and predictions. The matrix decomposition in the formula can be optimized and solved by machine learning algorithms, such as alternating least squares (ALS) and gradient descent. By minimizing the reconstruction error or regularized loss function, the optimal user latent factor matrix P, process mode latent factor matrix Q, and defect detection result latent factor matrix Y can be learned, so that the model can fit the observed data well and infer and predict unknown user-process mode-defect detection result combinations.

[0042] In summary, this formula maps users, process patterns, and defect detection results into a latent semantic space through matrix decomposition and latent factor representation, capturing the potential relationships and personalized features between them. This representation learning and low-rank approximation can effectively mine patterns in data, improve the accuracy of recommendations and predictions, and support personalized process optimization and quality improvement. At the same time, through model learning and optimization, the recommendation model can be continuously updated and improved to adapt to new data and changing needs, providing continuous intelligent decision support.

[0043] Furthermore, a recommendation model is established by combining user portraits, process modes and welding defect detection results to make welding process recommendations, which also includes: correcting the decomposition formula by the following formula: r(u,i,j)≈μ+b u +b i +b j +p u T ×q i +y j , where μ is the global bias term, b u is the bias term of user u, b i is the bias term of process mode i, b j is the bias term of defect detection result j, p u is the latent factor vector of user u, q i is the latent factor vector of process mode i, y j is the latent factor vector of defect detection result j; the stochastic gradient descent method is used to minimize the sum of squared errors and optimize the latent factor matrix P, Q, Y and bias term b u , b i , b j , the objective function of the sum of squared errors is: Among them, λ is the regularization parameter, ||·|| represents the Frobenius norm of the matrix, represents the L2 regularization term of all user bias terms, represents the L2 regularization term of all process mode bias terms, The L2 regularization term represents the bias term of all defect detection results. The objective function is optimized and solved by the alternating least squares method. Some parameters are fixed alternately, and other parameters are optimized. The optimization is iterated until convergence, and the optimal latent factor matrix P, Q, Y and bias term b are obtained. u , b i , b j .

[0044] Specifically, the global bias term μ and the user bias term b are introduced into the formula u , process mode bias term b i and the defect detection result bias term b jThe introduction of the bias term can capture the inherent bias and individual differences of users, process modes, and defect detection results, and improve the expressiveness and fitting ability of the recommendation model. The global bias term μ describes the overall mean or baseline level of the entire data set, and the user bias term b u Characterizes the user's personal preference, process mode bias term b i Characterizes the inherent properties of the process mode, the defect detection result bias term b j The inherent risks of different defect types are characterized. By introducing the bias term, the recommendation model can better adapt to the personalized characteristics of different users, process modes and defect detection results, and improve the accuracy and personalization of recommendations. The regularization term is introduced in the objective function. For latent factor matrices P, Q, Y and bias term b u , b i , b jL2 regularization is performed. The introduction of regularization terms can prevent overfitting of the model and improve the generalization ability and robustness of the model. By constraining the size of the parameters, regularization can control the complexity of the model, avoid overfitting of the training data, and improve the prediction performance of the model on unknown data. The λ in the regularization term is the regularization parameter, which controls the strength of the regularization. By adjusting the value of λ, the fitting error and model complexity can be balanced to find the optimal model parameters. The formula uses the stochastic gradient descent method to optimize and solve the objective function. The stochastic gradient descent method is an efficient optimization algorithm that gradually minimizes the objective function by iteratively updating parameters. In each iteration, one or a batch of training samples are randomly selected, the gradient is calculated and the parameters are updated to gradually approach the optimal solution. The advantages of the stochastic gradient descent method are high computational efficiency, can handle large-scale data sets, and have a certain degree of robustness to noisy data. By adjusting the learning rate and the number of iterations, the speed and accuracy of the optimization can be controlled. The formula uses the alternating least squares method to optimize and solve the objective function. The alternating least squares method is a commonly used matrix decomposition optimization algorithm. It alternately fixes some parameters, optimizes other parameters, and iteratively solves the optimal solution. In each iteration, the other parameters are fixed, and the objective function is converted into a linear least squares problem to solve the optimal value of the current parameters. Through repeated iterations until convergence, the optimal latent factor matrix and bias term are obtained. The advantages of the alternating least squares method are simple calculation, easy implementation, and good convergence and stability in practice. By introducing bias terms and latent factor matrices, the recommendation model can provide more interpretable results. The bias term characterizes the inherent properties of users, process modes, and defect detection results, and the latent factor matrix captures the potential relationship and mutual influence between them. By analyzing the bias term and latent factor matrix, we can gain insight into the reasons and influencing factors behind the recommendation results, providing useful insights and guidance for process optimization and quality improvement. At the same time, the recommendation model has good scalability and can easily incorporate new users, process modes, and defect detection results to adapt to changing business needs and data environments.

[0045] Furthermore, combining the user portrait, process mode and welding defect detection results, a recommendation model is established to recommend welding processes, which also includes: using the optimal latent factor matrix P, Q, Y and bias term b obtained by training u , b i , b j , predict the probability distribution of defect detection results for a given user under a given process mode. The prediction formula is: (u,i,j) =μ+b u +b i +b j +p u T ×q i +y j , where r (u,i,j)represents the predicted value of defect detection result j of user u under process mode i; μ is the global bias term; b u is the bias item corresponding to user u, b i is the bias term corresponding to process mode i, b j is the bias term corresponding to the defect detection result j; p u is the latent factor vector corresponding to user u, q i is the latent factor vector corresponding to process mode i, y j is the latent factor vector corresponding to the defect detection result j; according to the predicted probability distribution of the defect detection results, the potential preference of user u for process mode i is calculated, and the preference value calculation formula is:

[0046] Preference(u,i)=vj×r (u,i,j) ×v j , where v j is the utility value of defect detection result j, which can be flexibly set according to business needs. For a given user, the potential preference values ​​of different process modes are sorted to obtain the process mode preference sorting result of the corresponding user, and the process mode combination with the highest potential preference value is recommended to the user.

[0047] Specifically, formula r (u,i,j) =μ+b u +b i +b j +p u T ×q i +y j Using latent factor matrices and bias terms, the probability distribution of defect detection results j for a given user u under process mode i is predicted. By combining the latent factor representations of users, process modes, and defect detection results, as well as the corresponding bias terms, the interaction and personalized characteristics between them can be captured to improve the accuracy of defect detection result prediction. The predicted probability distribution of defect detection results can provide an important reference for process optimization and quality control, help identify high-risk process mode combinations, and take preventive and improvement measures. Formula Preference(u,i)=∑j×r (u,i,j) ×v jAccording to the predicted probability distribution of defect detection results, the potential preference value of user u for process mode i is calculated. By considering the utility value v_j of different defect detection results, the weight and priority of preference calculation can be flexibly adjusted according to business needs and quality requirements. For example, a higher negative utility value can be assigned to severe defects to avoid recommending process modes that may cause major quality problems. The calculation of the user's potential preference value combines the predicted probability and utility value of the defect detection results, provides a method for quantitatively evaluating the applicability and pros and cons of process modes, and provides a basis for personalized process recommendation. According to the calculated user's potential preference value, different process modes are sorted to obtain the user's process mode preference sorting result. By recommending the process mode combination with the highest potential preference value, personalized process selection and optimization suggestions can be provided to users to improve welding quality and efficiency. The process mode recommendation takes into account the user's historical performance, the characteristics of the process mode, and the feedback of the defect detection results, and provides a data-driven intelligent decision support to help users quickly find the optimal process parameters and operation plans.

[0048] 3. Beneficial effects

[0049] Compared with the prior art, the advantages of this application are:

[0050] (1) By acquiring multi-source heterogeneous data of the welding process and performing annotation, preprocessing and feature extraction, a high-quality labeled welding dataset is constructed, which provides rich data support for intelligent welding process analysis and overcomes the problems of insufficient data and poor quality in traditional methods.

[0051] (2) The image enhancement method using adaptive windows, contrast, and entropy indicators can dynamically adjust processing parameters according to the image content, effectively improving the quality of welding image data, improving the accuracy of subsequent defect recognition, and improving the robustness of the system;

[0052] (3) The welding processes are clustered by combining feature selection and fuzzy C-means clustering. By optimizing feature subsets and cluster centers, the optimal process mode under different environments can be adaptively discovered, providing a data-driven basis for process optimization and overcoming the subjectivity and limitations of traditional empirical process selection.

[0053] (4) Constructing multi-dimensional user portraits to comprehensively characterize user characteristics provides accurate user understanding for personalized process recommendations, overcoming the problem of insufficient understanding of user needs in traditional methods;

[0054] (5) A user-process-defect ternary relationship recommendation model is constructed, and a regularization term is introduced to control the complexity of the model. This model can mine the implicit relationship between users, processes, and defects, achieve accurate personalized process parameter recommendations, meet the needs of diverse users, and improve user satisfaction. This overcomes the shortcomings of traditional methods, such as single recommendation ideas and lack of personalization.

[0055] (6) The multi-armed bandit strategy is used to dynamically adjust the recommendation results and optimize the recommendation strategy online. This can ensure the accuracy of the recommendations while improving the novelty and exploratory nature of the recommendations, actively adapt to changes in user interests, enhance the sustainability of the system and user stickiness, and overcome the limitations of traditional static recommendations. BRIEF DESCRIPTION OF THE DRAWINGS

[0056] This specification will be further described in the form of exemplary embodiments, which will be described in detail by the accompanying drawings. These embodiments are not restrictive, and in these embodiments, the same number represents the same structure, wherein:

[0057] Figure 1 is an exemplary flow chart of a steel structure welding process quality management recommendation method based on big data processing according to some embodiments of this specification;

[0058] Figure 2 is an exemplary flow chart of image enhancement according to some embodiments of this specification;

[0059] Figure 3 is an exemplary flow chart of obtaining k cluster centers according to some embodiments of this specification;

[0060] Figure 4 This is an exemplary flowchart of updating cluster centers according to some embodiments of the present specification. DETAILED DESCRIPTION

[0061] The method and system provided in the embodiments of this specification are described in detail below with reference to the accompanying drawings.

[0062] Figure 1 This is an exemplary flow chart of a recommended method for steel structure welding process quality management based on big data processing according to some embodiments of this specification. Data collection and preprocessing, during the welding process, various sensors are used to collect parameter data such as welding current, voltage, speed, etc. in real time, and image data of the welding surface is collected synchronously. The collected raw data is cleaned, denoised and normalized to obtain standardized welding process parameter data and image data.

[0063] Figure 2It is an exemplary flow chart of image enhancement according to some embodiments of this specification, image data annotation and enhancement, data annotation: using data annotation tools (such as Label Me, Label Img, etc.), professional welding process experts manually annotate the collected welding image data. Experts mark different types of welding defects (such as welding cracks, welding slag inclusions, welding pores, etc.) and their locations based on the defect characteristics in the welding image. The annotated image data is synchronized with the corresponding welding process parameter data (such as welding current, voltage, speed, etc.) to construct a labeled welding process data set. Data division and window calculation: The labeled welding process data set is divided into multiple local windows, and the window size can be set according to factors such as image resolution and defect size. The cumulative distribution function (CDF) is calculated for the pixel grayscale value in each local window, that is, the cumulative probability of each grayscale value is calculated. The minimum value (CDF_min) and maximum value (CDF_max) of the CDF in each local window are recorded for subsequent grayscale value mapping. Gray value mapping: Map the gray value of each pixel in each local window to a specified dynamic range (such as 0 to 255). Use the formula: Gray value mapping is performed. Where i is the original gray value, CDF(i) is the cumulative distribution function value corresponding to the gray value i, CDF_min and CDF_max are the minimum and maximum values ​​of CDF in the local window, respectively, and round is the rounding function. Through gray value mapping, the contrast of the image can be enhanced to make the defect area more obvious. Window smooth transition: In order to avoid sudden changes in the mapping results between adjacent local windows, the bilinear interpolation method is used to smoothly transition the mapping results of adjacent windows. To calculate the overlap rate between adjacent local windows, use the formula: Where window_size is the size of the local window, and stride is the step size between adjacent local windows. The interpolation coefficient is calculated based on the overlap ratio, using the formula: coefficient = 1-overlap_ratio. The mapping results of adjacent local windows are weighted averaged, with the weight being the interpolation coefficient, to obtain the grayscale value after smooth transition. Adaptive window adjustment: In order to further improve the effect of image enhancement, the window size and overlap ratio are adaptively adjusted according to the contrast and entropy value within the local window. The contrast of the pixel grayscale value in each local window is calculated using the formula: Where max_gray and min_gray are the maximum and minimum values ​​of the pixel grayscale value in the local window, respectively. Calculate the entropy value of the pixel grayscale value in each local window using the formula: entropy = -∑(p×log2(p)), where p is the probability of each grayscale value in the local window. Set the contrast threshold and entropy threshold. When the contrast is less than the threshold and the entropy is less than the threshold, increase the size of the local window and reduce the overlap rate; when the contrast is greater than the threshold and the entropy is greater than the threshold, reduce the size of the local window and increase the overlap rate. By adaptively adjusting the window size and overlap rate, image enhancement can be performed in a targeted manner according to the local features of the image. Result combination: Combine the grayscale value mapping results and interpolation results of all local windows to obtain the complete image data after contrast enhancement. Resynchronize the enhanced image data with the original welding process parameter data to obtain the preprocessed welding process data set.

[0064] Feature extraction and selection, process feature extraction: Extract process features related to the welding process from the preprocessed welding process data set, including welding material type, plate thickness, welding method, etc. The welding material type can be represented by encoding the material number or name or one-hot encoding. The plate thickness can be directly represented by a numerical value, or binned or discretized according to the thickness range. The welding method can be represented by encoding different methods or one-hot encoding. Visual feature extraction: Extract visual features from the preprocessed welding image data, including image texture, shape, color and other features. Texture features can be extracted using methods such as gray level co-occurrence matrix (GLCM) and local binary pattern (LBP) to obtain feature vectors that reflect the distribution and change of image texture. Shape features can be obtained using methods such as edge detection and contour extraction to extract geometric features such as weld shape and defect shape. Color features can be obtained by calculating the color histogram and color moment of the image, reflecting the color distribution and change of the image. Feature selection: Use filtering and wrapping feature selection methods to evaluate the impact of each feature on welding quality and select the optimal feature subset. Filtering feature selection methods, such as variance threshold method and correlation coefficient method, select the most discriminative and relevant features based on the statistical characteristics of the features, such as variance and correlation. Wrapping feature selection methods, such as recursive feature elimination (RFE) and model-based feature importance evaluation, build a prediction model, evaluate the importance of features, and select the feature subset that contributes the most to model performance. Through feature selection, a refined welding process feature data set is obtained to reduce redundant and irrelevant features and improve the efficiency and accuracy of subsequent analysis and modeling. Data normalization: The minimum-maximum normalization method is used to normalize the refined feature data set, and the value range of each feature is scaled to the range of 0 to 1. For each feature, calculate its minimum and maximum values ​​in the data set, and then use the formula: Normalize the feature values. Normalization can eliminate the dimensional differences between features and make the value ranges of different features similar, which is convenient for subsequent clustering analysis and model training. Dataset partitioning: Divide the normalized feature dataset into a training set and a validation set, where the training set is used for feature selection and the validation set is used to evaluate the clustering performance of the feature set. Random partitioning can be used to divide the dataset into two parts according to a certain ratio (such as 70% training set and 30% validation set). Ensure that the distribution of the training set and the validation set is as similar as possible to avoid bias introduced by dataset partitioning. Feature subset search: Initialize the candidate feature subset to an empty set, initialize the optimal feature subset to the full set, and use the normalized feature dataset as the initial candidate feature subset. Repeat the following steps until the size of the optimal feature subset reaches the preset feature quantity threshold k: Add each candidate feature in the candidate feature subset to the current optimal feature subset to obtain a new feature subset. For each new feature subset, perform cluster analysis on the training set, such as using the K-means clustering algorithm, to evaluate clustering performance, such as the silhouette coefficient, Calinski-Harabasz index, etc. Select the feature subset with the best clustering performance as the new optimal feature subset. Remove the features that have been added to the optimal feature subset from the candidate feature subset and update the candidate feature subset. Through the above iterative process, gradually search and evaluate different feature combinations to find the feature subset with the best clustering performance.

[0065] Figure 3This is an exemplary flow chart for obtaining k cluster centers according to some embodiments of this specification. Cluster model training: a new feature subset is used to train the cluster model on the training set. Common clustering algorithms include K-means, Gaussian mixture model (GMM), hierarchical clustering, etc. Taking K-means as an example, the number of target categories k for clustering is set, k cluster centers are randomly initialized, and the following steps are iterated until convergence: for each sample in the training set, its distance from the k cluster centers is calculated, and the sample is assigned to the category to which the nearest cluster center belongs. For each category, the mean of all samples in the category is recalculated, and the mean is updated to the new cluster center. After the iterative process is completed, the clustering result on the training set is obtained, and each sample is assigned to the corresponding category. Clustering performance evaluation: the clustering model obtained by training is used for clustering on the validation set, and each sample in the validation set is assigned to the category to which the nearest cluster center belongs. To evaluate clustering performance, commonly used indicators include silhouette coefficient, Calinski-Harabasz index, Davies-Bouldin index, etc. Taking the silhouette coefficient as an example, for each sample in the validation set, the following two distances are calculated: a: the average distance between the sample and other samples in the same category. b: the average distance between the sample and the nearest sample in another category. Calculate the silhouette coefficient of the sample: The value range of the silhouette coefficient is [-1, 1]. The larger the value, the better the clustering effect. The mean of the silhouette coefficients of all samples in the validation set is calculated as an evaluation index of clustering performance. Feature selection iteration: Select the feature that improves the clustering performance the most after adding it, remove it from the candidate feature subset, add it to the optimal feature subset, and obtain a new optimal feature subset and a new candidate feature subset. The specific steps are as follows: For each feature in the candidate feature subset, add it to the current optimal feature subset in turn to form a new feature subset. Use the new feature subset to train the clustering model on the training set, evaluate the clustering performance on the validation set, and calculate the performance improvement value (such as the improvement of the silhouette coefficient). Select the feature with the largest performance improvement, remove it from the candidate feature subset, and add it to the optimal feature subset. The updated optimal feature subset contains the features that contribute the most to the clustering performance, and the updated candidate feature subset excludes the selected features. Feature subset update: If the clustering performance of the current candidate feature subset is better than the optimal feature subset, the current candidate feature subset is updated to the optimal feature subset. The specific steps are as follows: Use the current candidate feature subset to train the clustering model on the training set, and evaluate the clustering performance on the validation set. If the clustering performance (such as the silhouette coefficient) of the current candidate feature subset is higher than that of the optimal feature subset, the current candidate feature subset is updated to the optimal feature subset. Through this step, the optimal feature subset can be dynamically updated during the feature selection process to ensure that the selected feature combination can achieve the optimal clustering effect. Output the optimal feature subset: Output the optimal feature subset, which contains the top k features that have the greatest impact on the clustering results. The size of the optimal feature subset is determined by the preset feature number threshold k, which can be set according to actual needs and data scale. The features in the optimal feature subset have the greatest impact on the clustering results, can effectively distinguish different clustering categories, and improve the accuracy and interpretability of clustering.

[0066] Figure 4 This is an exemplary flow chart for updating cluster centers according to some embodiments of this specification, process cluster analysis, kernel density estimation: kernel density estimation is performed on the optimal feature subset, using a Gaussian kernel function. The expression of the Gaussian kernel function is: Where x is the distance between the data point and the kernel center. The bandwidth parameter of the kernel density estimation is determined using the Silverman's rule of thumb method. The calculation formula for the bandwidth parameter h is: Where σ is the standard deviation of the feature data set, and n is the number of data points. For each data point in the optimal feature subset, calculate its distance to other data points, and use the Gaussian kernel function to calculate the kernel density estimate. The calculation formula for the kernel density estimate is: Where x is the target data point, x iis the other data point, h is the bandwidth parameter, and K is the Gaussian kernel function. Select the first cluster center: According to the result of the kernel density estimation, select the data point with the highest density as the first cluster center. The data point with the highest density corresponds to the point with the largest kernel density estimate, indicating that the local density of this point in the feature space is the highest and it is suitable as the starting center of the cluster. Calculate the distance between the data point and the cluster center: Calculate the distance between each data point in the optimal feature subset and the first cluster center using the Mahalanobis distance. The calculation formula for the Mahalanobis distance is:

[0067] Where x is the eigenvector of the data point, c is the eigenvector of the cluster center, and Σ is the covariance matrix of the feature data set. The covariance matrix Σ reflects the correlation and scale difference between features. Using the Mahalanobis distance can consider the correlation between features and improve the accuracy of clustering. Calculate the probability of a data point being selected as the next cluster center: Based on the calculated distance, calculate the probability of each data point being selected as the next cluster center. The probability calculation formula is: is the distance between the data point x and the selected cluster center. The data point with a larger distance has a higher probability of being selected as the next cluster center, which can make the distribution of cluster centers more uniform and cover the entire feature space. Select the next cluster center: According to the calculated probability, use the Roulette Wheel Selection method to randomly select a new data point as the next cluster center. The Roulette Wheel Selection method randomly selects data points according to the probability. The data point with a larger probability is more likely to be selected. The specific steps are: calculate the sum of the selection probabilities of all data points, generate a random number between 0 and 1, compare the random number with the cumulative probability, and select the first data point with a cumulative probability greater than or equal to the random number as the next cluster center. Iteratively select cluster centers: repeat until k cluster centers are selected. After each new cluster center is selected, update the distance and selection probability of the data point to the cluster center until the required number of cluster centers is obtained. Calculate the membership of the data point: calculate the membership of each data point in the optimal feature subset to each cluster center, and use the membership calculation formula of Fuzzy C-means Clustering. The membership calculation formula is: Where c_i is the i-th cluster center, and m is the coefficient, which controls the hardness or softness of the cluster. The membership degree reflects the degree to which the data point belongs to each cluster center. The higher the membership degree, the higher the similarity between the data point and the cluster center. Update the location of the cluster center: Update the location of the cluster center according to the membership degree of the data point, and use the center update formula of fuzzy C-means clustering. The formula for updating the location of the cluster center is: Where u(x,c) is the degree of membership of data point x to cluster center c, x is the eigenvector of the data point, and c is the eigenvector of the cluster center. The position of the cluster center is updated by weighted averaging, so that the cluster center is closer to the data point with high membership, thereby improving the accuracy of clustering. Iteratively optimize the clustering results: repeat until the change of the cluster center is less than the preset threshold or the maximum number of iterations is reached. After each update of the cluster center, the degree of membership of the data point is recalculated, and the position of the cluster center is updated according to the degree of membership, and the clustering results are iteratively optimized. When the change of the cluster center is small or the maximum number of iterations is reached, the clustering results are considered to have converged and the iteration process is stopped.

[0068] Evaluate clustering performance, including: Evaluate the compactness and separation of clustering results: Use the Silhouette Coefficient to evaluate the compactness and separation of clustering results. For each data point i, calculate the following two distances: a(i) the average distance between data point i and other data points in the same category. b(i) the average distance between data point i and the nearest data point in other categories. Use the following formula to calculate the silhouette coefficient s(i) of data point i: The value range of the silhouette coefficient s(i) is [-1, 1]: When s(i) is close to 1, it means that the data point i is very close to other data points in the same category, but very far away from data points in other categories, and the compactness and separation of the clustering results are very good. When s(i) is close to 0, it means that the data point i is about the same distance from the data points in the same category and other categories, and the compactness and separation of the clustering results are average. When s(i) is close to -1, it means that the data point i is closer to data points in other categories, but farther away from data points in the same category, and the compactness and separation of the clustering results are poor. Calculate the average value of the silhouette coefficients of all data points as an indicator of the overall compactness and separation of the clustering results. Evaluate the overall performance of the clustering results: Use the Calinski-Harabasz Index to evaluate the overall performance of the clustering results. The calculation formula of the Calinski-Harabasz Index is: Among them, SS_b is the sum of squared deviations between categories, SS_w is the sum of squared deviations within a category, k is the number of categories in the cluster, and n is the total number of data points. The calculation steps of SS_b are: calculate the center point of each category (that is, the mean vector of all data points in the category). Calculate the square of the distance between the center point of each category and the center point of the entire data set, and multiply it by the number of data points in that category. Sum the above distance square products of all categories to obtain SS_b. The calculation steps of SS_w are: for each category, calculate the square of the distance between each data point in the category and the center point of the category. Sum the squared distances within each category to obtain the SS_w of the category. Sum the SS_w of all categories to obtain the total SS_w. The larger the value of the Calinski-Harabasz index, the greater the difference between categories in the clustering results, the smaller the difference within the category, and the better the clustering performance.

[0069] Intelligent defect detection uses transfer learning technology to quickly build a defect detection model under the current process parameters by referring to welding defect detection models under other materials and process conditions when there is less labeled data. The preprocessed welding image is input into the model for online defect recognition, judging the welding surface quality, and providing real-time quality feedback. User portrait construction collects the historical process parameter selection records of welding engineers, explores their preference rules for factors such as materials, thickness, and welding methods, and summarizes the personalized needs of engineers. At the same time, analyze the engineers' attention to different defect types and quality standards, and build a multi-dimensional user feature portrait.

[0070] Construction of process recommendation model and ternary relationship matrix R: Based on user portrait, process mode and welding defect detection results, a ternary relationship matrix R of user-process mode-defect detection results is constructed. The dimension of matrix R is m×n×l, where m is the number of users, n is the number of process modes, and l is the number of defect detection result categories. The element r(u,i,j) in matrix R represents the defect detection result j corresponding to the welding operation of user u under process mode i. When the operation result of user u under process mode i is defect detection result j, r(u,i,j) takes the value of 1; otherwise, r(u,i,j) takes the value of 0, indicating unknown or not occurring. Matrix decomposition: Decompose the constructed ternary relationship matrix R to obtain the user latent factor matrix P, process mode latent factor matrix Q and defect detection result latent factor matrix Y. The decomposition formula is: R≈μ+B u +B i +B j +P×Q T +Y, where R is the m×n×l dimensional ternary relationship matrix, m×n×l is the global bias term, and B u is the m-dimensional user bias matrix, B i is the n-dimensional process mode bias matrix, B jis the l-dimensional defect detection result bias matrix. P is the m×k-dimensional user latent factor matrix, k is the latent factor dimension, which represents the user's latent characteristics; Q is the n×k-dimensional process mode latent factor matrix, which represents the latent characteristics of the process mode; Y is the l×k-dimensional defect detection result latent factor matrix, which represents the latent characteristics of the defect detection result. Loss function and optimization: Define the loss function to measure the error and regularization term of matrix decomposition. Where L is the loss function, is the error term of matrix decomposition, λ(||P|| 2 +||Q|| 2 +||Y|| 2 ) is a regularization term used to prevent overfitting. Use an optimization algorithm, such as stochastic gradient descent (SGD) or alternating least squares (ALS), to minimize the loss function and solve the latent factor matrices P, Q, and Y. Welding process recommendation: Use the latent factor matrices P, Q, and Y obtained through training to recommend welding processes for new users and process modes. For new user u new and process mode i new , predict its defect detection result j new The value r(u new ,i new ,j new ): According to the size of the predicted value, the optimal combination of welding process parameters is recommended so that the defect detection results of the welding operation under this process mode are optimal. A threshold can be set to select process modes with predicted values ​​higher than the threshold as recommended candidates for user reference and selection. Model evaluation and optimization: Use methods such as cross-validation or leave-one-out to divide the data set into training and test sets to evaluate the performance of the recommended model. Calculate evaluation indicators such as root mean square error (RMSE), mean absolute error (MAE), accuracy, etc. to measure the accuracy and reliability of the recommendation results. According to the evaluation results, adjust the model's hyperparameters, such as latent factor dimension k, regularization coefficient λ, etc., to optimize the recommendation performance. Continue to iterate and optimize until a satisfactory recommendation effect is achieved.

[0071] Modified decomposition formula: Modify the original decomposition formula to obtain a new decomposition formula: r(u,i,j)≈μ+b u +b i +b j +p u T ×q i +y j , where μ is the global bias term, b u is the bias term of user u, b i is the bias term of process mode i, b j is the bias term of defect detection result j. uis the latent factor vector of user u, q i is the latent factor vector of process mode i, y j is the latent factor vector of defect detection result j. Objective function and regularization: Define the objective function of the sum of squared errors to optimize the latent factor matrices P, Q, Y and the bias term b u , b i , b j : in, is the error sum of squares term, which measures the difference between the predicted value and the true value. is a regularization term used to prevent overfitting. λ is a regularization parameter that controls the strength of the regularization term. ||·|| represents the Frobenius norm of the matrix. represents the L2 regularization term of all user bias terms, represents the L2 regularization term of all process mode bias terms, Represents the L2 regularization term of the bias term of all defect detection results. Stochastic gradient descent optimization: Stochastic gradient descent (SGD) is used to optimize the objective function. Randomly select a training sample (u,i,j) and calculate the error between the predicted value and the true value: According to the error value, calculate the gradient of each parameter: According to the gradient value, update each parameter: Where η is the learning rate, which controls the step size of parameter update. Repeat the above steps until convergence or the maximum number of iterations is reached. Alternating least squares optimization: Use alternating least squares (ALS) to optimize the objective function. Fix other parameters and optimize the user latent factor matrix P: P' = (Q T ×Q+λ×I) -1 ×(R1-uB u -B i -B j -Y), fix other parameters, and optimize the process mode latent factor matrix Q: Q = (P T ×P+λ×I) -1 ×(R2-uB u -B i -B j -Y). Fix other parameters and optimize the latent factor matrix Y of defect detection results: Fix other parameters and optimize user bias term b u : Fix other parameters and optimize process mode bias item b i : Fix other parameters and optimize the bias term b of defect detection results j : Where I is the identity matrix, R1, R2, and R3 are slices of the matrix R in different dimensions, and P u , Q i , Y j are the row vectors of matrices P, Q, and Y respectively, n u 、n i 、n j are the number of times user u, process mode i, and defect detection result j appear respectively. Repeat the above steps and optimize each parameter alternately until convergence or the maximum number of iterations is reached. Model training and recommendation: Use the training data set to optimize the objective function through stochastic gradient descent or alternating least squares method to obtain the optimal latent factor matrix P, Q, Y and bias term b u , b i , b j For new users new and process mode i new , predict its defect detection result j new The value on: According to the predicted value, the optimal welding process parameter combination is recommended.

[0072] Predict the probability distribution of defect detection results: Use the optimal latent factor matrix P, Q, Y and bias term b obtained through training u , b i , b j , predict the probability distribution of defect detection results for a given user under a given process mode. The prediction formula is: Among them, r (u,i,j) represents the predicted value of defect detection result j of user u under process mode i. μ is the global bias term, b u is the bias item corresponding to user u, b i is the bias term corresponding to process mode i, b j is the bias term corresponding to the defect detection result j. u is the latent factor vector corresponding to user u, q i is the latent factor vector corresponding to process mode i, y j is the latent factor vector corresponding to the defect detection result j. For a given user u and process mode i, calculate the predicted value r of all possible defect detection results j (u,i,j) , and obtain the probability distribution of the defect detection results of the user in this process mode. Calculate the user's potential preference for the process mode: According to the predicted probability distribution of defect detection results, calculate the potential preference of user u for process mode i. The preference value calculation formula is: Preference(u,i)=∑j×r (u,i,j) ×v j , where v jis the utility value of defect detection result j, which can be flexibly set according to business needs. The utility value v_j can be determined according to factors such as the severity of the defect detection result and the impact on welding quality. For example, different utility values ​​can be set for different types of defect detection results, such as 1 for no defect, 0.8 for minor defects, and 0.2 for severe defects. Calculate the potential preference value Preference(u,i) of user u for process mode i, that is, weighted sum the predicted probability distribution of defect detection results and the corresponding utility value. Generate process mode recommendation results: For a given user u, calculate its potential preference value Preference(u,i) for all process modes. Sort the process modes according to the potential preference value to obtain the process mode preference sorting result of the corresponding user. Recommend the process mode combination with the highest potential preference value to the user, that is, the process modes in the first few places in the sorting result. The number of recommended process modes can be set according to business needs, for example, recommending the first three process modes with the highest potential preference values. User u1, process modes i1, i2, i3, defect detection results j1 (no defect), j2 (minor defect), j3 (serious defect), the corresponding utility values ​​are v j1 =1, v j2 =0.8, v j3 =0.2. Using the latent factor matrix and bias term obtained through training, the probability distribution of defect detection results of user u1 under different process modes is predicted: Calculate the potential preference values ​​of user u1 for different process modes: Preference(u1,i1)=0.7+0.2×0.8+0.1×0.2=0.88,Preference(u1,i2)=0.5+0.3×0.8+0.2×0.2=0.78,Preference(u1,i3)=0.6+0.3×0.8+0.1×0.2=0.86。Sort the process modes according to the potential preference values: i1>i3>i2, and recommend the process mode combination with the highest potential preference value, i.e., i1 and i3, to user u1.

[0073] Dynamic recommendation optimization, each process mode is regarded as a slot machine arm, and the user's acceptance feedback on the recommended process is regarded as the slot machine reward. The Epsilon-Greedy algorithm is used to balance the utilization and exploration of recommendations: Exploitation: Select the process mode with the highest current reward estimate for recommendation with probability 1-ε. Exploration: Randomly select process modes for recommendation with probability ε to discover potential high-quality process combinations. Dynamically adjust the exploration probability ε, set a higher exploration probability in the initial stage, and gradually reduce the exploration probability as the number of recommendations increases to balance utilization and exploration. Reward estimation and update: For each process mode i, maintain its reward estimate Q(i) and the number of times it is selected N(i). Initially, the reward estimate Q(i) of all process modes is set to 0, and the number of times it is selected N(i) is set to 0. When process mode i is recommended to the user and receives feedback reward r, update its reward estimate and the number of times it is selected: N(i)'=N(i)+1, the reward r can be set according to the user's acceptance of the recommended process, for example, the reward for accepting the recommendation is 1, and the reward for not accepting the recommendation is 0. Initialize the estimated reward values ​​Q(i) and the number of selections N(i) of all process modes. Set the initial exploration probability ε, for example, ε=0.2. For each recommendation request: select the process mode with the highest current estimated reward value for recommendation (utilization) with probability 1-ε. Randomly select a process mode for recommendation (exploration) with probability ε. Obtain the user's feedback reward r for the recommended process. Update the estimated reward value Q(i) and the number of selections N(i) of the selected process mode. Dynamically adjust the exploration probability ε, for example, reduce the value of ε every certain number of recommendation requests. There are process modes i1, i2, and i3, the initial reward estimate Q(i1)=Q(i2)=Q(i3)=0, and the number of selections N(i1)=N(i2)=N(i3)=0. Set the initial exploration probability ε=0.2. First recommendation: randomly select process mode i2 for recommendation (exploration). The user accepts the recommendation and receives a reward of r=1. Update the estimated reward value and the number of times i2 is selected: Q(i2)=1, N(i2)=1. Second recommendation: select process mode i2 with the highest current estimated reward value for recommendation (utilization) with a probability of 0.8. The user does not accept the recommendation and receives a reward of r=0. Update the estimated reward value and the number of times i2 is selected: Q(i2)=0.5, N(i2)=2. Third recommendation: randomly select process mode i3 for recommendation (exploration) with a probability of 0.2. The user accepts the recommendation and receives a reward of r=1. Update the estimated reward value and the number of times i3 is selected: Q(i3)=1, N(i3)=1.

[0074] Process recommendation and application: When a new welding task arrives, the key process parameters are extracted according to the task requirements and input into the trained recommendation model to generate a series of candidate process parameter combinations. Combined with the feature profile of the current user, the applicability and potential defect risk of each candidate process are predicted, and the process parameter combination with the highest user preference and the lowest defect risk is selected to form the final recommendation result. Defect feedback and update: Apply the recommended process parameters to actual welding production, and evaluate the welding surface quality through the defect detection model. If quality defects are found, the association between the recommended process and the defect is recorded and fed back to the process recommendation model. The recommendation model is continuously trained and optimized using feedback data to continuously improve the accuracy and effectiveness of process recommendations. At the same time, new process defect samples are added to the training data set, and the defect detection model is updated to achieve iterative updates of the entire system.

Claims

1. A steel structure welding process quality management recommendation method based on big data processing, comprising: Acquire image data and real-time data during welding, including welding current, voltage and speed; Mark the image data to mark the location and type of welding defects; Synchronize the labeled image data with the corresponding real-time data to generate a labeled data set; Preprocess the labeled data set, including image enhancement, to generate a preprocessed data set; Extract key features based on the preprocessed data set, the key features include material type, thickness and welding method, and generate a feature data set; Perform cluster analysis on the characteristic data set to obtain the process mode under different welding parameters and environmental conditions; Based on the preprocessed data set and feature data set, a welding model is established through transfer learning to perform welding defect detection; Build user profiles based on their historical behaviors and preferences; Combine user portraits, process modes and welding defect detection results to establish a recommendation model and make welding process recommendations; According to the identified process mode and recommended welding process, the recommended results are optimized through the multi-armed bandit strategy.

2. The steel structure welding process quality management recommendation method based on big data processing according to claim 1 is characterized by: Image enhancement, including: Divide the labeled dataset into multiple local windows; Calculate the cumulative distribution function CDF of the pixel grayscale value in each local window, and map the CDF to the specified dynamic range. The grayscale value range after mapping is 0 to 255. The mapping function is: Where i is the original gray value, CDF(i) is the cumulative distribution function value corresponding to gray value i, CDF_min and CDF_max are the minimum and maximum values ​​of CDF in the local window, respectively, and round is the rounding function; The bilinear interpolation method is used to smoothly transition the mapping results of adjacent local windows. The interpolation coefficient is calculated according to the overlap rate of the local windows. The calculation formula is: coefficient = 1-overlap_ratio Among them, window_size is the size of the local window, stride is the step size between adjacent local windows, overlap_ratio is the overlap ratio of the local window, and coefficient is the interpolation coefficient.

3. The steel structure welding process quality management recommendation method based on big data processing according to claim 2 is characterized by: Image enhancement, also includes: Calculate the contrast and entropy value of the pixel grayscale value in each local window. The contrast calculation formula is: Among them, max_gray and min_gray are the maximum and minimum grayscale values ​​of pixels in the local window respectively; The entropy calculation formula is: entropy = -∑(p×log2(p)) Among them, p is the probability of each gray value appearing in the local window; Adjust the size and overlap rate of the local window according to the contrast and entropy values; Among them, when the contrast is less than a preset contrast threshold and the entropy value is less than a preset entropy threshold, the size of the local window is increased and the overlap rate is reduced; when the contrast is greater than the preset contrast threshold and the entropy value is greater than the preset entropy threshold, the size of the local window is reduced and the overlap rate is increased; The mapping results and interpolation results of all local windows are combined to obtain contrast-enhanced image data.

4. The steel structure welding process quality management recommendation method based on big data processing according to claim 3 is characterized by: Perform cluster analysis on feature data sets, including: The feature data set is normalized using the minimum-maximum normalization method, and the value range of each feature in the feature data set is scaled to the interval of 0 to 1 to obtain a normalized feature data set; The normalized feature data set is divided into a training set and a validation set, where the training set is used for feature selection and the validation set is used to evaluate the clustering performance of the feature set; Initialize the candidate feature subset to an empty set, initialize the optimal feature subset to the full set, and use the normalized feature data set as the initial candidate feature subset; Repeat the following steps until the size of the optimal feature subset reaches the preset feature quantity threshold k: Each candidate feature in the candidate feature subset is added to the current optimal feature subset to obtain a new feature subset; Use the new feature subset on the training set to train the clustering model and obtain the clustering results; Use the trained clustering model to perform clustering on the validation set and evaluate the clustering performance; Select the feature that improves the clustering performance the most after adding it, remove it from the candidate feature subset, add it to the optimal feature subset, and obtain a new optimal feature subset and a new candidate feature subset; If the clustering performance of the current candidate feature subset is better than that of the optimal feature subset, the current candidate feature subset is updated to the optimal feature subset; Output the optimal feature subset, which contains the top k features that have the greatest impact on the clustering results.

5. The steel structure welding process quality management recommendation method based on big data processing according to claim 4 is characterized by: Cluster analysis of feature data sets also includes: Calculate the kernel density estimation of the optimal feature subset, where the kernel density estimation uses the Gaussian kernel function and the bandwidth parameter of the kernel density estimation is determined using the Silverman's rule of thumb method; According to the results of kernel density estimation, the data point with the highest density is selected as the first cluster center; Calculate the distance between each data point in the optimal feature subset and the first cluster center. The distance calculation formula is: Among them, x is the eigenvector of the data point, c is the eigenvector of the cluster center, and Σ is the covariance matrix of the feature data set; According to the calculated distance, the probability of each data point being selected as the next cluster center is calculated. The probability calculation formula is: Where d(x) is the distance between the data point x and the selected cluster center; According to the calculated probability, a new data point is randomly selected as the next cluster center using the Roulette Wheel Selection method; Repeat the above steps until k cluster centers are selected.

6. The steel structure welding process quality management recommendation method based on big data processing according to claim 5 is characterized by: Cluster analysis of feature data sets also includes: Calculate the membership of each data point in the optimal feature subset to each cluster center. The membership calculation formula is: Among them, c i is the i-th cluster center, m is the coefficient, which controls the hardness or softness of clustering; According to the degree of membership of the data point, the position of the cluster center is updated. The formula for updating the position of the cluster center is: Among them, u(x,c) is the membership of data point x to cluster center c, x is the eigenvector of the data point, and c is the eigenvector of the cluster center; Repeat the above steps until the change in cluster center is less than the preset threshold or the maximum number of iterations is reached.

7. The steel structure welding process quality management recommendation method based on big data processing according to claim 6 is characterized by: Evaluate clustering performance, including: The following formula is used to evaluate the compactness and separation of clustering results: Where a(i) is the average distance between data point i and other data points in the same category, b(i) is the average distance between data point i and the nearest data points in other categories; s(i) ranges from -1 to 1, and the larger the s(i) value, the better the compactness and separation of the clustering results; The following formula is used to evaluate the overall performance of the clustering results: Among them, SS_b is the sum of squares of deviations between categories, SS_w is the sum of squares of deviations within categories, k is the number of clustered categories, and n is the total number of data points. The larger the CH value, the greater the difference between categories of the clustering results, the smaller the difference within the category, and the better the clustering performance.

8. The steel structure welding process quality management recommendation method based on big data processing according to any one of claims 1 to 7, characterized in that: Combined with user portraits, process modes and welding defect detection results, a recommendation model is established to make welding process recommendations, including: Construct a ternary relationship matrix R of user-process mode-defect detection result. The element r(u,i,j) in the matrix R represents the defect detection result j corresponding to the welding operation of user u under process mode i. The value of r(u,i,j) is 1, which means the operation result of the user under the corresponding process mode is j, and the value of r(u,i,j) is 0, which means it is unknown or has not occurred. The constructed ternary relationship matrix R is decomposed to obtain the user latent factor matrix P, the process mode latent factor matrix Q and the defect detection result latent factor matrix Y. The decomposition formula is: R≈μ+B u +B i +B j +P×Q T +Y Among them, R is an m×n×l-dimensional ternary relationship matrix, m is the number of users, n is the number of process modes, and l is the number of defect detection result categories; P is an m×k-dimensional user latent factor matrix, k is the latent factor dimension; Q is an n×k-dimensional process mode latent factor matrix; Y is an l×k-dimensional defect detection result latent factor matrix; B u is the m-dimensional user bias matrix, B i is the n-dimensional process mode bias matrix, B j is the l-dimensional defect detection result bias matrix; μ is the global bias term.

9. The steel structure welding process quality management recommendation method based on big data processing according to claim 8 is characterized by: Combining user portraits, process modes and welding defect detection results, a recommendation model is established to make welding process recommendations, including: Correct the decomposition formula by the following formula: r(u,i,j)≈μ+b u +b i +b j +p u T ×q i +y j Among them, μ is the global bias term, b u is the bias term of user u, b i is the bias term of process mode i, b j is the bias term of defect detection result j, p u is the latent factor vector of user u, q i is the latent factor vector of process mode i, y j is the latent factor vector of defect detection result j; The stochastic gradient descent method is used to minimize the sum of squared errors and optimize the latent factor matrices P, Q, Y and the bias term b. u , b i , b j , the objective function of the sum of squared errors is: Among them, λ is the regularization parameter, ||·|| represents the Frobenius norm of the matrix, represents the L2 regularization term of all user bias terms, represents the L2 regularization term of all process mode bias terms, Represents the L2 regularization term of the bias term of all defect detection results; The alternating least squares method is used to optimize the objective function, alternately fix some parameters, optimize other parameters, and iterate until convergence to obtain the optimal latent factor matrix P, Q, Y and bias term b u , b i , b j .

10. The steel structure welding process quality management recommendation method based on big data processing according to claim 9 is characterized in that: Combining user portraits, process modes and welding defect detection results, a recommendation model is established to make welding process recommendations, including: Using the optimal latent factor matrix P, Q, Y and bias term b obtained through training u , b i , b j , predict the probability distribution of defect detection results for a given user under a given process mode. The prediction formula is: r (u,i,j) =μ+b u +b i +b j +p u T ×q i +y j Among them, r (u,i,j) represents the predicted value of defect detection result j of user u under process mode i; μ is the global bias term; b u is the bias item corresponding to user u, b i is the bias term corresponding to process mode i, b j is the bias term corresponding to the defect detection result j; p u is the latent factor vector corresponding to user u, q i is the latent factor vector corresponding to process mode i, y j is the latent factor vector corresponding to the defect detection result j; According to the predicted probability distribution of defect detection results, the potential preference of user u for process mode i is calculated. The preference value calculation formula is: Preference(u,i)=∑j×r (u,i,j) ×v j Among them, v j is the utility value of defect detection result j, which can be flexibly set according to business needs; For a given user, the potential preference values ​​of different process modes are sorted to obtain the process mode preference sorting result of the corresponding user, and the process mode combination with the highest potential preference value is recommended to the user.

Images