Welding quality intelligent monitoring method and system

Abstract

The invention provides an intelligent welding quality monitoring method and system, and relates to the technical field of welding quality monitoring, and the method comprises the steps: building a database in which a first feature vector composed of historical welding data with the same timestamp is stored, training a welding quality evaluation model by using the historical welding data, and the real-time welding data is integrated into a second feature vector, the similarity between the second feature vector and each first feature vector is calculated, whether welding quality evaluation is directly carried out or a feedback signal is sent out is determined according to whether the similarity is larger than a preset similarity threshold value or not, if the evaluation result meets the expectation, data collection continues, and otherwise, the feedback signal is sent out. According to the method, the historical data can be fully utilized in the welding quality monitoring process, so that the welding quality can be accurately analyzed.

Description

Technical Field

[0001] The present application relates to the technical field of welding quality monitoring, and in particular to a welding quality intelligent monitoring method and system. Background Art

[0002] In the welding process, welding quality is a key factor in improving product reliability and safety. However, traditional welding quality monitoring methods often rely on manual inspection and empirical judgment, which is not only inefficient but also easily affected by human factors, resulting in limited accuracy and reliability of monitoring results.

[0003] With the rapid development of industrial manufacturing, the demand for real-time monitoring and intelligent evaluation of welding quality is becoming increasingly urgent. In order to achieve this goal, it is necessary to provide an intelligent monitoring method that can automatically collect welding data, intelligently analyze welding quality, and provide real-time feedback of monitoring results.

[0004] At present, although there are some welding quality monitoring systems based on sensors and data acquisition technology, most of these systems can only realize simple data recording and storage functions, and cannot make full use of the information in historical welding data, which leads to false alarms and missed alarms during the monitoring process. Summary of the invention

[0005] In order to make full use of historical data in the process of welding quality monitoring and realize accurate analysis of welding quality, the present application provides a welding quality intelligent monitoring method and system.

[0006] In the first aspect, the present application provides a method for intelligent monitoring of welding quality, which adopts the following technical solution: A welding quality intelligent monitoring method comprises the following steps: Building a database: Building a database storing a plurality of first feature vectors, wherein the first feature vectors are composed of a plurality of historical welding data with the same timestamp; Modeling: Establish a welding quality assessment model, use historical welding data to train the welding quality assessment model, and obtain the trained welding quality assessment model; Data collection: collect various real-time welding data and integrate all the real-time welding data into the second feature vector; First calculation: respectively calculating the similarity between the second eigenvector and each first eigenvector to obtain a plurality of similarity values; First judgment: judging whether there is a similarity value greater than a preset similarity threshold, if so, executing the first evaluation step; if not, executing the feedback step; First evaluation: input the real-time welding data into the trained welding quality evaluation model to obtain the first evaluation result; Second judgment: judging whether the first evaluation result meets expectations, if so, executing the data collection step after a preset time interval; if not, executing the feedback step; Feedback: sending a feedback signal, wherein the feedback signal includes: the similarity between the first feature vector and the second feature vector and / or the first evaluation result.

[0007] The present application can effectively integrate and utilize historical welding data by establishing a database storing multiple first feature vectors. Then, the welding quality assessment model is trained using the historical welding data. The trained welding quality model can capture the key features in the welding process, thereby accurately assessing the welding quality. Then, various real-time welding data are collected and integrated into a second feature vector. By calculating the similarity between the second feature vector and each first feature vector, the current welding state is intelligently matched with the known state in history. Then, based on the comparison result of the similarity with the preset similarity threshold, it is intelligently determined whether there is a potential risk in the current welding process. Based on the comparison result of the similarity with the preset threshold, an evaluation or feedback mechanism is triggered to achieve timely monitoring of the welding quality. Then, the real-time welding data is input into the trained welding quality assessment model to obtain an accurate assessment result (i.e., the first assessment result) of the welding quality. Then, based on whether the first assessment result meets the expected judgment result, it is determined whether a feedback signal needs to be issued. If the assessment result does not meet expectations, the feedback mechanism is triggered so that the operator can understand the welding quality problem in a timely manner. The feedback signal includes the similarity between the first eigenvector and the second eigenvector and / or the first evaluation result, providing detailed and accurate information to the operator. The present application realizes comprehensive, accurate and real-time monitoring of welding quality by integrating historical data, real-time data collection, automated evaluation and feedback mechanisms. This not only improves the stability and reliability of welding quality, but also reduces errors and realizes accurate analysis of welding quality.

[0008] Optionally, after performing the second judgment step and before performing the feedback step, the method further includes: Second evaluation: including feature extraction, second calculation, third judgment and output; Feature extraction: Based on the first evaluation step, a feature vector of the first evaluation result is obtained, which is recorded as a third feature vector. Each feature value in the third feature vector represents the probability of real-time welding data being divided into different categories. Second calculation: using a weighted average algorithm to calculate the evaluation value of the real-time welding data based on the eigenvalues ​​in the third eigenvector; Third judgment: judging whether the evaluation value is greater than a preset evaluation threshold, if so, executing the output step; if not, executing the feedback step; Output: Output the signal indicating normal welding; The feedback signal also includes: an evaluation value of real-time welding data.

[0009] The present application further extracts the characteristic vector (i.e., the third characteristic vector) of the result of the first evaluation, and the third characteristic vector reflects the probability when the real-time welding data is divided into different quality categories. Then, a weighted average algorithm is used to combine the characteristic values ​​in the third characteristic vector to calculate the evaluation value of the real-time welding data. Through quantitative evaluation, the present application obtains a numerical indicator (i.e., evaluation value) that can reflect the welding quality, so as to facilitate subsequent decision-making. Then, the evaluation value is compared with the preset evaluation threshold to decide whether to output a signal indicating that the welding is normal or to perform a feedback step. When the evaluation value is greater than the preset evaluation threshold, it indicates that the welding quality under the real-time welding data is excellent, and then a signal indicating that the welding is normal is output. The evaluation value of the real-time welding data is added to the feedback signal, providing the operator with more comprehensive quality information. The present application not only improves the accuracy and reliability of the monitoring results, but also provides the operator with more comprehensive feedback information, which helps to promote the continuous improvement of welding quality.

[0010] Optionally, after executing the third judgment step and before executing the output step, the method further includes: Incremental learning: Obtain the loss function of the trained welding quality assessment model, use the online gradient descent algorithm, and use real-time welding data to calculate the gradient of the loss function with respect to the parameters of the trained welding quality assessment model; Update parameters: update the parameters of the trained welding quality assessment model using the gradient to obtain a new welding quality assessment model; Data storage: Update the real-time welding data to the database and execute the data collection steps after a preset time interval.

[0011] The present application obtains the loss function of the trained welding quality assessment model, and uses an online gradient descent algorithm to calculate the gradient of the loss function with respect to the model parameters using real-time welding data. Incremental learning allows the welding quality assessment model to gradually update itself when receiving new data without retraining the entire model, which saves time and improves the adaptability and accuracy of the model, so that the trained welding quality assessment model can adapt to changes in the welding process more quickly and improve the processing ability of welding data of new welding materials. Then the calculated gradient is used to update the parameters of the trained welding quality assessment model. By continuously adjusting the model parameters, the trained welding quality assessment model can better fit the real-time welding data, thereby improving the accuracy of the assessment. The continuous optimization of the model parameters makes the welding quality assessment model more accurate and can more accurately reflect the welding quality status. Then the real-time welding data is updated to the database. Provide rich historical data support for subsequent incremental learning and model training. The continuous enrichment of the database further enhances the ability of the welding quality assessment model to utilize historical data, which helps to improve the generalization performance of the model. Through incremental learning and parameter updating, the model can continuously adapt to changes in the welding process, improve the processing ability of new data, and thus improve the overall performance.

[0012] Optionally, after executing the step of performing the third judgment and before executing the step of providing feedback, the method further includes: First setting: normalize the first eigenvector to obtain the fourth eigenvector, and sequentially concatenate the fourth eigenvectors into a first matrix, where the element in the i-th row and j-th column of the first matrix represents the j-th eigenvalue in the i-th fourth eigenvector; First acquisition: sequentially concatenate the maximum values ​​of each column in the first matrix into a fifth eigenvector; sequentially concatenate the minimum values ​​of each column in the first matrix into a sixth eigenvector; The third calculation: calculate the index value of the ith fourth eigenvector and the fifth eigenvector in the first matrix. The calculation model is as follows: ; in, is the index value of the i-th fourth eigenvector and the fifth eigenvector in the first matrix; is the eigenvalue of the jth column in the i-th fourth eigenvector; is the jth eigenvalue in the fifth eigenvector; n is the total number of eigenvalues ​​in the fifth eigenvector; m is the number of fourth eigenvectors in the first matrix; Fourth calculation: Calculate the index value of the i-th fourth eigenvector and the sixth eigenvector in the first matrix. The calculation model is as follows: ; in, is the index value of the i-th fourth eigenvector and the sixth eigenvector in the first matrix; is the jth eigenvalue in the sixth eigenvector; Fifth calculation: Calculate the ideal value of the i-th fourth eigenvector. The calculation model is as follows: ; in, is the ideal value of the ith fourth eigenvector; Second setting: setting the historical welding data corresponding to the maximum ideal value as the target data; First adjustment: adjust the real-time welding data according to the target data.

[0013] This application can eliminate the dimensional differences between different features by standardizing the first eigenvector, making the data more comparable. Then, the standardized fourth eigenvector is spliced ​​into the first matrix, and the fifth eigenvector and the sixth eigenvector are respectively formed by extracting the maximum and minimum values ​​of each column in the first matrix. Then, by calculating the index values ​​of the fourth eigenvector and the fifth eigenvector and the sixth eigenvector, the distance between each fourth eigenvector and the extreme eigenvector can be quantified, thereby identifying abnormal values ​​or deviations from the normal range in the data. Then, the ideal value of the fourth eigenvector is calculated, and the ideal value reflects the optimal or expected state in the data. Then, the historical welding data corresponding to the maximum ideal value is set as the target data, which helps to provide a reference standard or target for the real-time welding data. Then, the real-time welding data is adjusted according to the target data, which can make the real-time data closer to the ideal state or target data, thereby improving the welding quality.

[0014] Optionally, after performing the second setting step and before performing the first adjustment step, the method further includes: Third evaluation: input the target data into the trained welding quality evaluation model to obtain the third evaluation result; Fourth judgment: judge whether the third evaluation result meets expectations, if so, execute the first adjustment step; if not, execute the feedback step.

[0015] This application inputs the target data into the trained welding quality assessment model to obtain the third assessment result to further verify the quality of the target data and determine whether the target data does represent the expected state in the welding process. The third assessment result is used to determine whether it meets expectations. If the assessment result meets expectations, it means that the target data is reliable and the real-time welding data can be adjusted according to the target data. Through the third assessment and the fourth judgment, the reliability of the target data can be further verified before adjusting the real-time welding data, which helps to reduce adjustment errors caused by inaccurate target data and enhances the accuracy of decision-making.

[0016] Optionally, after executing the fourth judgment step and before executing the feedback step, the method further includes: Initialization: setting relevant parameters of the gray wolf population, randomly initializing the position of the gray wolf population according to the second eigenvector, each gray wolf in the gray wolf population represents an adjustment scheme for welding data, and the relevant parameters include the population size, the maximum number of iterations and the convergence factor; Calculate fitness: recalculate the fitness value of the second eigenvector for each gray wolf, and take the gray wolf with the smallest fitness value as the leader; Position update: Update the position of the alpha wolf according to the social hierarchy of the gray wolf, and execute the fitness calculation steps until the stopping condition is met; Acquisition plan: Acquire the adjustment plan of the welding data corresponding to the leader as the final adjustment plan; Second adjustment: adjust the real-time welding data according to the final adjustment plan; The feedback signal also includes: a final adjustment plan.

[0017] This application sets the relevant parameters of the gray wolf population (such as population size, maximum number of iterations and convergence factor), which provides a basis for the subsequent gray wolf optimization algorithm. Then, the position of the gray wolf population is randomly initialized according to the second eigenvector, and each gray wolf in the gray wolf population represents a welding data adjustment scheme. Then, by recalculating the fitness value of each gray wolf, the effect of each adjustment scheme is quantified. The smaller the fitness value, the closer the scheme is to the optimal solution. Then, the gray wolf with the smallest fitness value is used as the alpha wolf, that is, the current optimal welding data adjustment scheme. The gray wolf optimization algorithm uses the social hierarchy of gray wolves to update the position, which helps the algorithm to quickly converge to the optimal solution in the solution space. Then, by continuously iteratively updating the position of the alpha wolf and recalculating the fitness value, the optimal welding data adjustment scheme can be gradually approached. When the stop condition is met (such as reaching the maximum number of iterations or converging to a certain threshold), the welding data adjustment scheme corresponding to the current alpha wolf is obtained as the final adjustment scheme. Then, the real-time welding data is adjusted according to the final adjustment scheme in order to achieve better welding quality. The final adjustment plan is included in the feedback signal, which helps operation and maintenance personnel or users understand the current optimal adjustment plan and provide a basis for subsequent decision-making.

[0018] Optionally, after executing the step of acquiring the solution and before executing the step of second adjustment, the method further includes: Fourth evaluation: inputting the welding data corresponding to the final adjustment scheme into the trained welding quality evaluation model to obtain a fourth evaluation result; Fifth judgment: judging whether the fourth evaluation result meets expectations, if yes, executing the second adjustment step; if no, executing the alarm step; Alarm: Send an alarm signal.

[0019] After obtaining the final adjustment plan, the corresponding welding data is input into the trained welding quality assessment model again for verification to determine the effectiveness of the adjustment plan. Through the evaluation of the model, quantitative indicators of the effect of the adjustment plan can be obtained, which helps to improve the accuracy and reliability of the adjustment. The fourth evaluation result is used to determine whether the adjustment plan meets expectations. This is a key step in determining subsequent operations (adjustment or alarm). If the evaluation result does not meet expectations, it means that there may be problems with the adjustment plan or deviations in the model prediction. At this time, executing the alarm step can detect problems in time and take measures.

[0020] Optionally, after executing the first judgment step and before executing the first evaluation step, the method further includes: Sixth calculation: recording the historical welding data corresponding to the similarity value being greater than the preset similarity threshold as the first data; Sixth judgment: judging whether the quantity of the first data is one, if so, executing the clustering step; if not, executing the first evaluation step; Clustering: K-Means clustering algorithm is used to cluster historical welding data to obtain normal welding data sets and abnormal welding data sets; Obtaining the center: Obtain the cluster center of the normal welding data set, recorded as the second data; obtain the cluster center of the abnormal welding data set, recorded as the third data; Seventh calculation: calculating the Euclidean distance between the first data and the second data and recording it as the first Euclidean distance; calculating the Euclidean distance between the first data and the third data and recording it as the second Euclidean distance; Seventh judgment: respectively judge whether the first Euclidean distance and the second Euclidean distance are less than a preset distance threshold, if so, execute the first evaluation step; if not, execute the data collection step.

[0021] The present application determines whether a clustering step needs to be performed by counting the number of similarity values ​​greater than a preset similarity threshold value, and judging whether the number of first data is one. If the number of first data is one, it means that there is only one sample in the historical welding data that is highly similar to the target data. At this time, clustering is needed to further analyze the data set, and the first data can be judged whether it can fall into the corresponding category by Euclidean distance. If the number of first data is not one, that is, there are multiple similar samples, the evaluation can be performed directly, which improves efficiency. The K-Means clustering algorithm is used to divide the historical welding data, which helps to divide the data set into two categories: normal welding data set and abnormal welding data set, providing a basis for subsequent analysis. Afterwards, clustering can be used to obtain the cluster centers of the normal welding data set and the abnormal welding data set, which represent the typical characteristics of each data set. Afterwards, obtaining the cluster centers (second data and third data) is the basis for the subsequent calculation of Euclidean distance, which helps to quantify the distance between the first data and the normal and abnormal welding data sets. Afterwards, by calculating the Euclidean distance between the first data and the second and third data, the similarity between the target data and the normal and abnormal data sets can be quantified. Whether the data collection step needs to be performed is determined based on whether the Euclidean distance is less than the preset distance threshold, and then whether the current real-time welding data has errors. If the target data is close to the cluster center of the normal or abnormal data set, it means that the real-time welding data can be mapped to the corresponding data set. When the Euclidean distance is greater than the preset threshold, it means that there is an error in the real-time welding data collection, and then the data collection step is performed.

[0022] Optionally, after executing the seventh determination step and before executing the data collection step, the method further includes: Reclassification: Use the big law to redetermine the value of classifying the historical welding data, and reclassify the historical welding data into a normal data set and an abnormal data set according to the value; Reset the center: calculate the Euclidean distance between the i-th sample in the normal data set and the remaining samples, record it as the fourth data, take the sample corresponding to the minimum sum of the fourth data as the new second data, calculate the Euclidean distance between the i-th sample in the abnormal data set and the remaining samples, record it as the fifth data, take the sample corresponding to the minimum sum of the fifth data as the new third data, and re-execute the seventh calculation step.

[0023] This application adopts the Otsu's Method to redetermine the value of dividing the historical welding data (i.e., the Otsu's Method is used to determine which historical welding data is used as the threshold for dividing the normal data set and the abnormal data set). This is an adaptive threshold determination method that can automatically select the optimal threshold according to the statistical characteristics of the data, thereby optimizing the division of the data set. By re-dividing, the division of the normal data set and the abnormal data set can be made more accurate, the misclassification and omission can be reduced, and the data quality can be improved. By calculating the Euclidean distance between each sample in the normal data set and the abnormal data set and the remaining samples, and selecting the sample with the smallest sum of distances as the new clustering center (the second data and the third data), the position of the clustering center can be updated so that it more accurately represents the typical characteristics of each data set. Resetting the center helps to enhance the robustness of the clustering results and reduce the deviation of the clustering results caused by improper selection of the initial clustering center. After re-dividing and resetting the center, the Euclidean distance between the target data and the new clustering center needs to be recalculated to obtain the latest distance information, so that the intelligent monitoring method of welding quality can perform better when facing complex and variable data sets, and improve reliability and accuracy.

[0024] In the second aspect, the present application provides a welding quality intelligent monitoring system, which adopts the following technical solutions: A welding quality intelligent monitoring system, comprising: a processor and a memory, The memory stores program code; The processor executes the steps of the method when calling the program code in the memory.

[0025] In summary, the present application includes at least one of the following beneficial technical effects: 1. This application realizes comprehensive, accurate and real-time monitoring of welding quality by integrating historical data, real-time data collection, automated evaluation and feedback mechanisms. This not only improves the stability and reliability of welding quality, but also reduces human intervention and realizes accurate analysis of welding quality.

[0026] 2. This application can eliminate the dimensional differences between different features by standardizing the first eigenvector, making the data more comparable. Then, the standardized fourth eigenvector is spliced ​​into a first matrix, and the fifth eigenvector and the sixth eigenvector are respectively formed by extracting the maximum and minimum values ​​of each column in the first matrix. Then, by calculating the index values ​​of the fourth eigenvector and the fifth and sixth eigenvectors, the distance between each fourth eigenvector and the extreme eigenvector can be quantified, thereby identifying abnormal values ​​or deviations from the normal range in the data. Then, the ideal value of the fourth eigenvector is calculated, and the ideal value reflects the optimal or expected state in the data. Then, the historical welding data corresponding to the maximum ideal value is set as the target data, which helps to provide a reference standard or target for the real-time welding data. Then, the real-time welding data is adjusted according to the target data, which can make the real-time data closer to the ideal state or target data, thereby improving the welding quality. BRIEF DESCRIPTION OF THE DRAWINGS

[0027] Figure 1 is a flow chart of Example 1 of the present application; Figure 2 is a flow chart of Example 2 of the present application; Figure 3 This is a flow chart of Example 3 of the present application. DETAILED DESCRIPTION

[0028] The following combination Figures 1 to 3 This application is described in further detail.

[0029] Embodiment 1: This embodiment discloses a welding quality intelligent monitoring method, referring to Figure 1 , the method includes: S1 database building, S2 modeling, S3 data collection, S4 first calculation, S5 first judgment, S6 first evaluation, S7 second judgment and S8 feedback. First, a database storing a first feature vector composed of historical welding data is established, and a welding quality evaluation model is trained. After that, real-time welding data is collected and integrated to obtain a second feature vector. The similarity between the second feature vector and each first feature vector is calculated to determine whether further evaluation of welding quality is needed. If the similarity meets the standard, the evaluation model is used to make a quality judgment; if it does not meet the standard or the evaluation result does not meet expectations, a feedback signal containing the similarity or evaluation result is issued. The process of this embodiment is as follows: S1: Building a database storing a plurality of first feature vectors, wherein the first feature vectors are composed of a plurality of historical welding data with the same timestamp.

[0030] The historical welding data contains multiple parameters such as current, voltage, welding speed, gas flow, etc. These data are recorded at the same timestamp to form a complete description of the welding state. Through feature extraction technology, key information can be extracted from these historical welding data to form the first feature vector.

[0031] S2 modeling, establish a welding quality assessment model, use historical welding data to train the welding quality assessment model, and obtain a trained welding quality assessment model. The function of the trained welding quality assessment model is to evaluate the quality of the input welding data and determine whether it meets the preset quality standards.

[0032] The welding quality assessment model can be any one of the models such as CNN model, RNN model, LSTM model, BI-LSTM model, etc. As long as the model can be trained with historical welding data and has the ability to determine whether the input data meets the preset quality standards.

[0033] S3 data acquisition, in the actual welding process, various welding parameters will be recorded in real time to form real-time welding data. These data may come from sensors, controllers or other data acquisition devices. In order to compare with the first feature vector in the database, these real-time data need to be integrated into a second feature vector. The composition of this vector is consistent with the first feature vector to improve the accuracy of calculating similarity.

[0034] For example, the real-time welding data is: current 200A, voltage 24V, welding speed 30cm / min, then the second eigenvector is expressed as: [200, 24, 30].

[0035] S4 is a first calculation, in which the similarity between the second eigenvector and each first eigenvector is calculated respectively to obtain a plurality of similarity values.

[0036] Similarity is an indicator that measures the degree of proximity between two vectors, and can be calculated by a variety of methods, such as cosine similarity, Euclidean distance, etc. In this step, multiple similarity values ​​are calculated, each of which corresponds to the degree of similarity between a first eigenvector and a second eigenvector.

[0037] S5 is a first judgment, judging whether there is a similarity value greater than a preset similarity threshold, if so, executing S6 a first evaluation; if not, executing S8 feedback.

[0038] S6 first evaluation, inputting the real-time welding data into the trained welding quality evaluation model to obtain a first evaluation result.

[0039] The first evaluation result is the welding quality grade or score predicted by the trained welding quality evaluation model according to the input welding parameters.

[0040] S7 second judgment, by comparing the first evaluation result with the preset quality standard, judge whether the first evaluation result meets the expectation, if so, execute S3 data collection after a preset time interval; if not, execute S8 feedback.

[0041] The preset quality standard can be different scores corresponding to different welding qualities, for example, when the output score is 80 points (inclusive) to 100 points (inclusive), it is excellent, when it is 60 points (inclusive) to 80 points (exclusive), it is qualified, and when it is less than 60 points (exclusive), it is unqualified. Different score thresholds can also be set according to needs.

[0042] The preset quality standard may also be a tolerance range for the number, size or distribution of welding defects. For example, in a weld of a preset length, the number of pores shall not exceed a certain number (e.g., the number of pores per meter of weld shall not exceed 3), and the maximum diameter of the pores shall not exceed 10% of the weld thickness; the number of slag inclusions in the weld shall be controlled within a certain range, and the size of the slag inclusions shall not be too large to avoid affecting the mechanical properties of the weld, etc.

[0043] S8 feedback: If the similarity does not meet the standard or the evaluation result does not meet expectations, a feedback signal needs to be issued. The feedback signal may include the similarity between the first eigenvector and the second eigenvector and / or the first evaluation result, so that the operator can understand the actual situation of the welding quality and take corresponding measures. The feedback signal can be in the form of sound and light alarm, screen display, data recording, etc., depending on the actual application scenario and requirements.

[0044] This embodiment includes establishing a database for storing a first feature vector composed of historical welding data; establishing and training a welding quality assessment model; collecting and integrating real-time welding data; similarity; judging whether the similarity is greater than a preset similarity threshold, and deciding whether to evaluate or feedback; and then judging whether the first assessment result meets expectations, and if so, executing the data collection step, and if not, sending a feedback signal. This application realizes real-time monitoring and early warning of the welding process by collecting welding data in real time and comparing it with historical data, and combining the trained welding quality assessment model for quality prediction.

[0045] Example 2: Reference Figure 2 The difference between this embodiment and embodiment 1 is that after executing the second judgment in S7 and before executing the feedback in S8, the following is further included: S701 second evaluation, including: feature extraction, second calculation, third judgment and output.

[0046] Feature extraction: Feature extraction is performed based on the result of the first evaluation of S6. Key information is extracted from the first evaluation result to obtain the feature vector of the first evaluation result, which is recorded as the third feature vector. Each eigenvalue in the third feature vector represents the probability of real-time welding data being divided into different categories. Each eigenvalue in the third feature vector represents the probability of real-time welding data being divided into a specific category.

[0047] Second calculation: A weighted average algorithm is used to calculate the evaluation value of the real-time welding data. This evaluation value is calculated based on the eigenvalues ​​in the third eigenvector (i.e., the probability values ​​of each category). The weighted average algorithm takes into account the weight of each eigenvalue, making the evaluation result more accurate and reliable.

[0048] A weight can be assigned to each characteristic value, and then the weighted average is calculated as the evaluation value. The evaluation value reflects the performance of real-time welding data in the overall welding quality and is used to judge whether the welding quality meets expectations.

[0049] The third judgment: determine whether the evaluation value is greater than the preset evaluation threshold. If so, it means that the quality of the real-time welding data is high and meets or exceeds the expected standard. The reason why the first evaluation result in the second judgment of S7 does not meet the expectations is that the real-time welding data is a new evaluation result caused by new welding materials. It is necessary to execute S702 incremental learning to enhance the generalization ability of the trained welding quality evaluation model; if not, it means that the quality of the real-time welding data is low and does not meet the expected standard. The real-time welding data needs to be adjusted, and then the first setting of S705 is executed.

[0050] S702 incremental learning obtains the loss function of the trained welding quality assessment model, adopts an online gradient descent algorithm, and uses real-time welding data to calculate the gradient of the loss function with respect to the parameters of the trained welding quality assessment model.

[0051] The online gradient descent algorithm will first initialize the parameters of the trained welding quality assessment model and set a learning rate. Then, for each real-time welding data sample, the algorithm will calculate the corresponding loss function value and the gradient of the loss function with respect to the model parameters. The model parameters are then updated based on the gradient and learning rate. This process will be repeated until the model's performance on new data reaches the preset stopping condition (such as the loss function value no longer decreases significantly).

[0052] S703 updates the parameters, using the gradient to update the parameters of the trained welding quality assessment model to obtain a new welding quality assessment model.

[0053] Adopting gradient to update parameters is an existing technology. This step adopts ordinary gradient descent algorithm for calculation. The calculation model is as follows: ; in, represents the updated model parameters, represents the model parameters before updating, represents the learning rate, Represents the gradient of the loss function L with respect to the model parameters.

[0054] For another example, the mini-batch gradient descent algorithm is used to update the model parameters. The process is as follows: ; Represents the gradient of the loss function L on the mini-batch data with respect to the parameters.

[0055] S704 data storage, updates the real-time welding data to the database, and executes S3 data collection after a preset time interval.

[0056] Output: Output the signal that welding is normal.

[0057] S705: First setting: performing standardization processing on the first eigenvector, and recording the first eigenvector after the standardization processing as the fourth eigenvector.

[0058] Then the fourth eigenvectors are concatenated into the first matrix in sequence (the concatenation process is performed in the order of the fourth eigenvectors, and each fourth eigenvector becomes a row of the matrix. The eigenvalues ​​of each fourth eigenvector are placed in the corresponding columns of the matrix according to their positions in the vector), and the element in the i-th row and j-th column of the first matrix represents the j-th eigenvalue in the i-th fourth eigenvector.

[0059] S706 first obtains, traverses each column of the first matrix, finds the maximum value of the column, and sequentially concatenates these maximum values ​​into a fifth eigenvector, which contains the maximum value of each column of the first matrix and is used to represent the upper limit of the eigenvalue represented by the column.

[0060] Similarly, each column of the first matrix is ​​traversed to find the minimum value of the column, and these minimum values ​​are sequentially concatenated into a sixth eigenvector. The sixth eigenvector contains the minimum value of each column of the first matrix and is used to represent the lower limit of the eigenvalue represented by the column.

[0061] S707: The third calculation is to calculate the index values ​​of the i-th fourth eigenvector and the fifth eigenvector in the first matrix. The calculation model is as follows: ; in, is the index value of the i-th fourth eigenvector and the fifth eigenvector in the first matrix; is the eigenvalue of the jth column in the i-th fourth eigenvector; is the jth eigenvalue in the fifth eigenvector; n is the total number of eigenvalues ​​in the fifth eigenvector; m is the number of fourth eigenvectors in the first matrix.

[0062] S708 is a fourth calculation, which is to calculate the index values ​​of the i-th fourth eigenvector and the sixth eigenvector in the first matrix. The calculation model is as follows: ; in, is the index value of the i-th fourth eigenvector and the sixth eigenvector in the first matrix; is the jth eigenvalue in the sixth eigenvector.

[0063] S709 fifth calculation, calculating the ideal value of the i-th fourth eigenvector, the calculation model is as follows: ; in, is the ideal value of the i-th fourth eigenvector.

[0064] S710 second setting, set the historical welding data corresponding to the maximum ideal value as the target data, that is, traverse all fourth eigenvectors (that is, all historical welding data), calculate their ideal values, and find the historical welding data corresponding to the fourth eigenvector with the maximum ideal value. This historical welding data is considered to be the best or closest to the expected state, and is therefore set as the target data.

[0065] S711: the third evaluation, inputting the target data into the trained welding quality evaluation model to obtain the third evaluation result.

[0066] S712 is the fourth judgment, judging whether the third evaluation result meets expectations. If so, it means that the welding quality represented by the target data meets the standard or expectation, and then S721 is executed for the first adjustment; if not, S713 is executed for initialization.

[0067] S713 initialization, setting relevant parameters of the gray wolf population, randomly initializing the position of the gray wolf population according to the second eigenvector, each gray wolf in the gray wolf population represents an adjustment scheme for welding data, and the relevant parameters include population size, maximum number of iterations and convergence factor.

[0068] Gray wolf population: Each gray wolf in the gray wolf population represents an adjustment plan for welding data, that is, each gray wolf has a unique set of welding parameter settings, which together constitute an adjustment plan.

[0069] Related parameters: The related parameters that need to be set include the population size (i.e. the number of gray wolves), the maximum number of iterations (the maximum number of cycles the algorithm runs) and the convergence factor (a parameter that affects the convergence speed of the algorithm and the search accuracy).

[0070] Initialization position: Randomly initialize the position of the gray wolf population according to the second eigenvector (i.e., the vector formed by the real-time welding data). The initial parameters of each gray wolf (i.e., the adjustment scheme of each welding data) are randomly selected within a certain range according to the second eigenvector.

[0071] S714 calculates the fitness, recalculates the fitness value of the second eigenvector for each gray wolf, and takes the gray wolf corresponding to the minimum fitness value as the leader wolf.

[0072] Fitness value: For each gray wolf (i.e., each welding data adjustment scheme), the fitness value of its second eigenvector needs to be recalculated. This fitness value reflects the performance or effect of the welding data in the welding process after the real-time welding data is adjusted according to the adjustment scheme.

[0073] Alpha wolf selection: The gray wolf with the smallest fitness value is selected as the alpha wolf, that is, in the current population, the gray wolf with the best performance (that is, the smallest fitness value) is regarded as the leader (optimal solution).

[0074] S715 position update, update the position of the alpha wolf according to the social hierarchy of the gray wolf, and execute S714 to calculate the fitness until the stop condition is met, and then execute S716 to obtain the solution. The stop condition includes: reaching the maximum number of iterations or converging to a predetermined fitness value.

[0075] S716 obtains a plan, and obtains the adjustment plan of the welding data corresponding to the leader as the final adjustment plan.

[0076] S717 fourth evaluation, inputting the welding data corresponding to the final adjustment solution into the trained welding quality evaluation model to obtain a fourth evaluation result.

[0077] S718 is a fifth judgment, determining whether the fourth evaluation result meets expectations. If so, executing S720 is a second adjustment; if not, executing S719 is an alarm.

[0078] S719 alarm, sends out an alarm signal.

[0079] S720 second adjustment, adjusting the real-time welding data according to the final adjustment plan.

[0080] S721 is the first adjustment, which is to adjust the real-time welding data according to the target data.

[0081] In S8 feedback, the feedback signal also includes: a final adjustment plan and an evaluation value of real-time welding data.

[0082] This embodiment extracts key information from the first evaluation result, obtains a third eigenvector representing the classification probability of real-time welding data, and then uses a weighted average algorithm to calculate the evaluation value of the real-time welding data based on the eigenvalues ​​in the third eigenvector. Then, based on the comparison between the evaluation value and the preset threshold, it is determined whether to perform incremental learning or adjust the real-time welding data. If the evaluation value does not meet expectations and is caused by new welding materials, the incremental learning step is performed (using the online gradient descent algorithm to update the welding quality evaluation model parameters). If the evaluation value does not meet expectations but is not caused by new materials, the real-time welding data is adjusted.

[0083] The process of adjusting the real-time welding data in this embodiment is: through the first setting step to the fourth judgment step, the target data similar to the current working condition is searched in the historical welding data, and the real-time welding data is adjusted according to the target data. If there is no welding data in the historical welding data that makes the evaluation result (i.e., the third evaluation result) meet the expectations, a new adjustment scheme is searched through the initialization step to the second adjustment step, and the real-time welding data is adjusted according to the adjustment scheme, so that the welding quality evaluation model can adapt to different welding materials as much as possible. By searching for target data similar to the current working condition in the historical welding data and adjusting the real-time welding data according to the target data, the value of the historical data is fully utilized, and the accuracy and efficiency of the adjustment are improved. When there is no welding data that meets the conditions in the historical data, a new adjustment scheme is searched through the initialization step, so that the welding quality evaluation model can adapt to different welding materials and working conditions. The implementation of the adjustment scheme enables the welding quality evaluation model to adapt to different welding materials as much as possible, and improves the practicality and reliability of the model.

[0084] Example 3: Reference Figure 3 The difference between this embodiment and embodiment 1 is that after executing S5 first judgment and before executing S6 first evaluation, it also includes: S91 is the sixth calculation, recording the historical welding data corresponding to the similarity value being greater than the preset similarity threshold as the first data.

[0085] S92 is a sixth judgment, judging whether the number of the first data is one, if so, executing S93 clustering; if not, executing S6 first evaluation.

[0086] S93 clustering, uses K-Means clustering algorithm to cluster historical welding data to obtain normal welding data sets and abnormal welding data sets.

[0087] S94 obtains the center. After clustering is completed, obtain the cluster center of the normal welding data set (recorded as the second data) and the cluster center of the abnormal welding data set (recorded as the third data).

[0088] The cluster centers of these data sets are actually welding data. The welding data corresponding to the cluster center of the normal welding data set is recorded as the second data, and the welding data corresponding to the cluster center of the abnormal welding data set is recorded as the third data.

[0089] S95 seventh calculation: calculating the Euclidean distance between the first data and the second data, recording it as the first Euclidean distance; calculating the Euclidean distance between the first data and the third data, recording it as the second Euclidean distance.

[0090] S96 is the seventh judgment, judging whether the first Euclidean distance and the second Euclidean distance are less than the preset distance threshold respectively. If so, it means that the first data is close enough to a certain cluster center, and it can be considered that it belongs to the cluster represented by the cluster center, and then S6 is executed for the first evaluation; if not, it means that the current data division may not be accurate enough or the selection of the cluster center is not reasonable enough, and S97 is executed for re-division.

[0091] S97 reclassification uses the Great Law Method (also known as the Otsu Method) to re-determine the numerical values ​​for classifying the historical welding data, and reclassifies the historical welding data into a normal data set and an abnormal data set based on the numerical values.

[0092] S98 resets the center, calculates the Euclidean distance between the i-th sample in the normal data set and the remaining samples, records them as the fourth data, takes the sample corresponding to the minimum sum of the fourth data as the new second data, calculates the Euclidean distance between the i-th sample in the abnormal data set and the remaining samples, records them as the fifth data, takes the sample corresponding to the minimum sum of the fifth data as the new third data, and re-executes the seventh calculation of S95 to update the value of the Euclidean distance.

[0093] This embodiment includes counting the number of similarity values ​​greater than a preset threshold, and judging whether K-Means clustering is required to obtain normal and abnormal welding data sets based on this, and then calculating the Euclidean distance between these similar data and the cluster center, and judging the attribution of the data based on the distance or using the big law to re-divide the data set and reset the cluster center when necessary, and executing S95 seventh calculation again to verify whether the classification error is caused by the problem of the cluster center.

[0094] Embodiment 4: This embodiment discloses a welding quality intelligent monitoring system, the system comprising: processor and memory, The memory stores program code; The processor executes the steps of the method when calling the program code in the memory.

[0095] The above are all preferred embodiments of the present application, and the protection scope of the present application is not limited thereto. Therefore, any equivalent changes made according to the structure, shape, and principle of the present application should be included in the protection scope of the present application.

Claims

1. A welding quality intelligent monitoring method, characterized in that: include: Building a database: Building a database storing a plurality of first feature vectors, wherein the first feature vectors are composed of a plurality of historical welding data with the same timestamp; Modeling: Establish a welding quality assessment model, use historical welding data to train the welding quality assessment model, and obtain the trained welding quality assessment model; Data collection: collect various real-time welding data and integrate all the real-time welding data into the second feature vector; First calculation: respectively calculating the similarity between the second eigenvector and each first eigenvector to obtain a plurality of similarity values; First judgment: judging whether there is a similarity value greater than a preset similarity threshold, if so, executing the first evaluation step; if not, executing the feedback step; First evaluation: input the real-time welding data into the trained welding quality evaluation model to obtain the first evaluation result; Second judgment: judging whether the first evaluation result meets expectations, if so, executing the data collection step after a preset time interval; If not, the feedback step is executed; Feedback: sending a feedback signal, wherein the feedback signal includes: the similarity between the first feature vector and the second feature vector and / or the first evaluation result.

2. The welding quality intelligent monitoring method according to claim 1 is characterized in that: After executing the second judgment step and before executing the feedback step, the method further includes: Second evaluation: including feature extraction, second calculation, third judgment and output; Feature extraction: Based on the first evaluation step, a feature vector of the first evaluation result is obtained, which is recorded as a third feature vector. Each feature value in the third feature vector represents the probability of real-time welding data being divided into different categories. Second calculation: using a weighted average algorithm to calculate the evaluation value of the real-time welding data based on the eigenvalues ​​in the third eigenvector; Third judgment: judging whether the evaluation value is greater than a preset evaluation threshold, if so, executing the output step; if not, executing the feedback step; Output: Output the signal indicating normal welding; The feedback signal also includes: an evaluation value of real-time welding data.

3. The welding quality intelligent monitoring method according to claim 2 is characterized in that: After executing the third judgment step and before executing the output step, the method further includes: Incremental learning: Obtain the loss function of the trained welding quality assessment model, use the online gradient descent algorithm, and use real-time welding data to calculate the gradient of the loss function with respect to the parameters of the trained welding quality assessment model; Update parameters: update the parameters of the trained welding quality assessment model using the gradient to obtain a new welding quality assessment model; Data storage: Update the real-time welding data to the database and execute the data collection steps after a preset time interval.

4. The welding quality intelligent monitoring method according to claim 2 or 3, characterized in that: After executing the third judgment step and before executing the feedback step, the method further includes: First setting: normalize the first eigenvector to obtain the fourth eigenvector, and sequentially concatenate the fourth eigenvectors into a first matrix, where the element in the i-th row and j-th column of the first matrix represents the j-th eigenvalue in the i-th fourth eigenvector; First acquisition: sequentially concatenate the maximum values ​​of each column in the first matrix into a fifth eigenvector; sequentially concatenate the minimum values ​​of each column in the first matrix into a sixth eigenvector; The third calculation: calculate the index value of the ith fourth eigenvector and the fifth eigenvector in the first matrix. The calculation model is as follows: ； in, is the index value of the i-th fourth eigenvector and the fifth eigenvector in the first matrix; is the eigenvalue of the jth column in the i-th fourth eigenvector; is the jth eigenvalue in the fifth eigenvector; n is the total number of eigenvalues ​​in the fifth eigenvector; m is the number of fourth eigenvectors in the first matrix; Fourth calculation: Calculate the index value of the i-th fourth eigenvector and the sixth eigenvector in the first matrix. The calculation model is as follows: ； in, is the index value of the i-th fourth eigenvector and the sixth eigenvector in the first matrix; is the jth eigenvalue in the sixth eigenvector; Fifth calculation: Calculate the ideal value of the i-th fourth eigenvector. The calculation model is as follows: ； in, is the ideal value of the i-th fourth eigenvector; Second setting: setting the historical welding data corresponding to the maximum ideal value as the target data; First adjustment: adjust the real-time welding data according to the target data.

5. The welding quality intelligent monitoring method according to claim 4 is characterized in that: After executing the second setting step and before executing the first adjustment step, the method further includes: Third evaluation: input the target data into the trained welding quality evaluation model to obtain the third evaluation result; Fourth judgment: judge whether the third evaluation result meets expectations, if so, execute the first adjustment step; if not, execute the feedback step.

6. The welding quality intelligent monitoring method according to claim 5 is characterized in that: After executing the fourth determination step and before executing the feedback step, the method further includes: Initialization: setting relevant parameters of the gray wolf population, randomly initializing the position of the gray wolf population according to the second eigenvector, each gray wolf in the gray wolf population represents an adjustment scheme for welding data, and the relevant parameters include the population size, the maximum number of iterations and the convergence factor; Calculate fitness: recalculate the fitness value of the second eigenvector for each gray wolf, and take the gray wolf with the smallest fitness value as the leader; Position update: Update the position of the alpha wolf according to the social hierarchy of the gray wolf, and execute the fitness calculation steps until the stopping condition is met; Acquisition plan: Acquire the adjustment plan of the welding data corresponding to the leader as the final adjustment plan; Second adjustment: adjust the real-time welding data according to the final adjustment plan; The feedback signal also includes: a final adjustment plan.

7. The welding quality intelligent monitoring method according to claim 6 is characterized in that: After executing the step of acquiring the solution and before executing the step of second adjustment, the method further includes: Fourth evaluation: inputting the welding data corresponding to the final adjustment scheme into the trained welding quality evaluation model to obtain a fourth evaluation result; Fifth judgment: judging whether the fourth evaluation result meets expectations, if yes, executing the second adjustment step; if no, executing the alarm step; Alarm: Send an alarm signal.

8. The welding quality intelligent monitoring method according to any one of claims 1 to 3, characterized in that: After executing the first judgment step and before executing the first evaluation step, the method further includes: Sixth calculation: recording the historical welding data corresponding to the similarity value being greater than the preset similarity threshold as the first data; Sixth judgment: judging whether the quantity of the first data is one, if so, executing the clustering step; if not, executing the first evaluation step; Clustering: K-Means clustering algorithm is used to cluster historical welding data to obtain normal welding data sets and abnormal welding data sets; Obtaining the center: Obtain the cluster center of the normal welding data set, recorded as the second data; obtain the cluster center of the abnormal welding data set, recorded as the third data; Seventh calculation: calculating the Euclidean distance between the first data and the second data and recording it as the first Euclidean distance; calculating the Euclidean distance between the first data and the third data and recording it as the second Euclidean distance; Seventh judgment: respectively judge whether the first Euclidean distance and the second Euclidean distance are less than a preset distance threshold, if so, execute the first evaluation step; if not, execute the data collection step.

9. The welding quality intelligent monitoring method according to claim 8, characterized in that: After executing the seventh determination step and before executing the data collection step, the method further includes: Reclassification: Use the big law to redetermine the value of classifying the historical welding data, and reclassify the historical welding data into a normal data set and an abnormal data set according to the value; Reset the center: calculate the Euclidean distance between the i-th sample in the normal data set and the remaining samples, record it as the fourth data, take the sample corresponding to the minimum sum of the fourth data as the new second data, calculate the Euclidean distance between the i-th sample in the abnormal data set and the remaining samples, record it as the fifth data, take the sample corresponding to the minimum sum of the fifth data as the new third data, and re-execute the seventh calculation step.

10. An intelligent welding quality monitoring system, characterized in that: include: processor and memory, The memory stores program code; When the processor calls the program code in the memory, the steps of the method according to any one of claims 1 to 9 are executed.

Images