Welding quality monitoring method and device and excavator
Through welding status data processing and multi-model integration, welding quality evaluation is achieved, solving the problem of low quality monitoring efficiency during welding process, and improving the stability and monitoring accuracy of the welding process.
Patent Information
- Application Number
- CN202411045041.2
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-07-31
- Publication Date
- 2024-11-12
AI Technical Summary
The welding quality monitoring efficiency and accuracy during welding process are low, and the lack of understanding of the welding mechanism leads to insufficient stability and adaptability of real-time welding monitoring.
By obtaining welding status data and field data during welding, processing it as feature vectors, combining heat source models and multiple defect classification models to conduct welding quality evaluation, including deep neural networks, support vector machines, K nearest neighbor algorithms and extreme gradient enhancement trees Integration of models, welding defect detection and molten pool state analysis are carried out.
It significantly improves the accuracy and stability of welding quality monitoring, enhances the generalization performance of the welding process and the robustness of real-time monitoring, and can effectively monitor faults and defects in the welding process.
Smart Images

Figure CN118926775A_ABST
Abstract
Description
Technical Field
[0001] The present application relates to the field of welding technology, and in particular to a welding quality monitoring method, device and excavator. Background Art
[0002] In modern industrial manufacturing, welding is the main means of connecting metal materials. Welding is a process method that uses heating, pressurization, or both to produce atomic bonding between two or more metal workpieces to form a permanent connection. During the welding process, the workpieces are heated to a molten state and then cooled and solidified to form a solid connection. Therefore, the quality of welding is directly related to the structural strength, sealing, service life and safety of the product. However, the complex environment of the welding process and the diverse changes in the welding state lead to low efficiency and low accuracy in monitoring the welding quality. At present, in the actual industrial production process, more empirical models are used for welding state monitoring, and there is a lack of understanding of the welding mechanism. This brings unstable risks to the generalization performance of real-time welding monitoring, and the adaptability to complex welding processes needs to be further improved. Summary of the invention
[0003] In order to solve the above technical problems, the present application is proposed. The embodiments of the present application provide a welding quality monitoring method, device and excavator, which can improve the accuracy of welding quality monitoring.
[0004] According to one aspect of the present application, a welding quality monitoring method is provided, comprising: acquiring welding status data and welding site data during a welding process; processing the welding status data to obtain a feature vector; using the feature vector as an input of a defect detection model to obtain a quality evaluation of an output of the defect detection model; using the welding site data and historical welding data as an input of a heat source model to obtain a molten pool state output by the heat source model; coupling the molten pool state and the quality evaluation to obtain a welding quality evaluation result.
[0005] In one embodiment, the welding site data includes: initial plate temperature, welding time, density, specific heat capacity, thermal conductivity, thermal diffusion coefficient, heat source efficiency, plate parameters, moving speed, calculation time interval, heat input and heat source parameters; wherein, the welding site data and historical welding data are used as inputs of the heat source model to obtain the molten pool state output by the heat source model, including: inputting the initial plate temperature, welding time, density, specific heat capacity, thermal conductivity, thermal diffusion coefficient, heat source efficiency, plate parameters, moving speed, calculation time interval, heat input and heat source parameters into the temperature field analytical model to obtain the temperature field data calculated by the temperature field analytical model; calculating the melt width and melt depth according to the temperature field data to obtain theoretical calculation results; using the theoretical calculation results, historical welding data, workpiece parameters, voltage, current and sound data, weld shape, process parameters, image data and heat source parameters as inputs of the heat source model to analyze the welding penetration and obtain the molten pool state.
[0006] In one embodiment, the welding quality monitoring method also includes: integrating multiple welding defect classification models as a defect detection model; wherein the welding defect classification model includes a deep neural network, a support vector machine, a K-nearest neighbor algorithm, a decision tree and an extreme gradient boosting tree; wherein, using the feature vector as the input of the defect detection model to obtain a quality evaluation of the output of the defect detection model, including: using the feature vector as the input of multiple welding defect classification models to obtain multiple quality evaluations; based on the evaluation accuracy of the welding defect classification model, weighting the multiple quality evaluations; summing the prediction possibilities of the prediction types of the multiple welding defect classification models after weighted processing; using the prediction type with the greatest prediction possibility as the output of the defect detection model to obtain the final quality evaluation.
[0007] In one embodiment, the welding state data is processed to obtain a feature vector, including: classifying the welding state data according to a preset flaw detection method to obtain a plurality of welding types; selecting signal segments with corresponding features from the welding types and adding labels; preprocessing the labeled signal segments to obtain filtered valid data; calibrating the valid data on the time axis to obtain valid data calibrated within a preset time period; aligning and packaging the valid data calibrated within the preset time period to form a data packet; extracting features from the data packet to obtain extracted feature parameters; performing importance evaluation and uniqueness evaluation on the feature parameters; and performing feature merging on the feature parameters that pass the importance evaluation and the uniqueness evaluation to obtain a feature vector.
[0008] In one embodiment, the welding status data includes visual signals, acoustic signals and electrical signals; wherein, feature extraction is performed from the data packet to obtain the extracted feature parameters, including: using the autoencoder network as a feature extraction means to extract image feature information in the visual signal to obtain the autoencoder feature parameters; extracting the spectrum of the acoustic signal based on Fourier transform, and summing, averaging and calculating the variance of the acoustic signal according to the interval of the spectrum to obtain the frequency domain feature parameters of the sound; extracting the time domain statistical features of the acoustic signal and the electrical signal to obtain the time domain feature parameters; wherein, the feature parameters are evaluated for importance and uniqueness, including: evaluating the importance and uniqueness of the autoencoder feature parameters, the frequency domain feature parameters and the time domain feature parameters.
[0009] In one embodiment, an importance evaluation is performed on the self-encoding feature parameter, the frequency domain feature parameter, and the time domain feature parameter, including: respectively inputting the self-encoding feature parameter, the frequency domain feature parameter, and the time domain feature parameter into a welding defect classification model to obtain multiple quality prediction accuracy rates; inputting all of the self-encoding feature parameter, the frequency domain feature parameter, and the time domain feature parameter into the welding defect classification model to obtain an overall quality prediction accuracy rate; taking any one of the multiple quality prediction accuracy rates as a target quality prediction accuracy rate, and when the ratio of the target quality prediction accuracy rate to the overall quality prediction accuracy rate is greater than or equal to a preset ratio, determining that the feature parameter corresponding to the target quality prediction accuracy rate passes the importance evaluation.
[0010] In one embodiment, the self-encoding feature parameters, the frequency domain feature parameters and the time domain feature parameters are evaluated for uniqueness, including: the self-encoding feature parameters, the frequency domain feature parameters and the time domain feature parameters form a feature set; the feature set is input into a welding defect classification model to obtain an overall quality prediction accuracy; any one of the self-encoding feature parameters, the frequency domain feature parameters and the time domain feature parameters is used as a target feature parameter, and the target feature parameter is deleted from the feature set to obtain a non-target feature parameter set; the non-target feature parameter set is input into a welding defect classification model to obtain a non-target feature parameter prediction accuracy; the decrease value of the non-target feature parameter prediction accuracy relative to the overall quality prediction accuracy is calculated; when the percentage of the decrease value in the overall quality prediction accuracy is greater than or equal to a preset percentage, it is determined that the target feature parameter passes the uniqueness evaluation.
[0011] In one embodiment, the welding status data includes visual signals, acoustic signals and electrical signals, and the signal segments after adding labels are preprocessed to obtain filtered valid data, including: filtering and denoising the visual signal segments after adding labels to obtain filtered visual valid data; performing spectrum analysis on the acoustic signal segments after adding labels to remove high-frequency noise irrelevant to the welding quality to obtain denoised acoustic valid data; performing zero drift compensation on the acoustic signal segments and electrical signal segments after adding labels to obtain compensated acoustic valid data, current valid data and voltage valid data; wherein, the valid data is calibrated on the time axis to obtain valid data calibrated within a preset time period, including: calibrating the visual valid data, acoustic valid data, current valid data and voltage valid data on the time axis according to the collection time.
[0012] According to another aspect of the present application, a welding quality monitoring device is provided, including: an acquisition module, used to acquire welding state data and welding site data during the welding process; a processing module, used to process the welding state data to obtain a feature vector; a first evaluation module, used to use the feature vector as an input of a defect detection model to obtain a quality evaluation of an output of the defect detection model; a second evaluation module, used to use the welding site data and historical welding data as inputs of a heat source model to obtain a molten pool state output by the heat source model; and a coupling evaluation module, used to couple the molten pool state and the quality evaluation to obtain a welding quality evaluation result.
[0013] According to another aspect of the present application, an excavator is provided, comprising: a sensor, the sensor being mounted on the boom of the excavator, the sensor being used to collect welding status data and welding site data during the boom welding process; a quality monitoring device, the quality monitoring device being communicatively connected to the sensor, the quality monitoring device being used to execute the welding quality monitoring method described in any one of the above embodiments, so as to monitor the welding quality of the boom of the excavator.
[0014] The welding quality monitoring method, device and excavator provided in the present application can effectively predict and monitor the welding status based on the welding mechanism. On the basis of quality evaluation based on the welding status data, a heat source model is added to highly simulate the temperature field of the welding process. The welding quality is finally evaluated in combination with the molten pool state and the corresponding quality evaluation. Faults in the welding process are effectively monitored, which greatly improves the stability and generalization performance of the welding process and makes remarkable improvements to the robustness of real-time monitoring of welding. BRIEF DESCRIPTION OF THE DRAWINGS
[0015] By describing the embodiments of the present application in more detail in conjunction with the accompanying drawings, the above and other purposes, features and advantages of the present application will become more apparent. The accompanying drawings are used to provide a further understanding of the embodiments of the present application and constitute a part of the specification. Together with the embodiments of the present application, they are used to explain the present application and do not constitute a limitation of the present application. In the accompanying drawings, the same reference numerals generally represent the same components or steps.
[0016] Figure 1 A schematic structural diagram of the distribution relationship between the overall sensor and the robot, workpiece, industrial computer and control cabinet provided by an exemplary embodiment.
[0017] Figure 2 It is a flow chart of a welding quality monitoring method provided by an exemplary embodiment of the present application.
[0018] Figure 3 It is a flow chart of a welding quality monitoring method provided by another exemplary embodiment of the present application.
[0019] Figure 4 It is a flow chart of a welding quality monitoring method provided by another exemplary embodiment of the present application.
[0020] Figure 5 It is a flow chart of a welding quality monitoring method provided by another exemplary embodiment of the present application.
[0021] Figure 6 It is an overall flow chart of a welding quality monitoring method provided by an exemplary embodiment of the present application.
[0022] Figure 7 It is a structural schematic diagram of a welding quality monitoring device provided by an exemplary embodiment of the present application.
[0023] Figure 8 It is a structural diagram of an electronic device provided by an exemplary embodiment of the present application.
[0024] Explanation of the accompanying reference numerals: 10. gas cylinder; 11. welding machine; 12. Hall sensor; 13. data acquisition card; 14. control cabinet; 15. robot; 16. workpiece; 17. sound sensor; 18. CCD camera; 19. industrial computer; 80. electronic device; 81. processor; 82. memory; 83. input device; 84. output device. DETAILED DESCRIPTION
[0025] Below, the exemplary embodiments according to the present application will be described in detail with reference to the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present application, rather than all the embodiments of the present application, and it should be understood that the present application is not limited to the exemplary embodiments described here.
[0026] Exemplary excavator
[0027] This exemplary embodiment provides an excavator, comprising: a sensor, which is installed on the boom of the excavator, and is used to collect welding status data and welding site data during the boom welding process; a quality monitoring device, which is communicatively connected to the sensor, and is used to execute the welding quality monitoring method provided in this application to monitor the welding quality of the boom of the excavator.
[0028] The present application can be applied to excavators, with the automated welding of the excavator's boom as the target welding real-time monitoring object, thereby replanning the robot's operating path, and installing sound sensors, voltage sensors, current sensors and visual sensors to collect sound, voltage, current and image information during the welding process while keeping the process unchanged. On the basis of these signals, the construction of an online monitoring model for welding defects based on ensemble learning is completed, and a welding heat source neural network adaptive intelligent matching generalization model is added to highly simulate the temperature field of the welding process, and effectively monitor burn-through, incomplete penetration, weld deviation, wire jamming and porosity in the welding process to achieve effective monitoring of the welding process.
[0029] Figure 1 The distribution relationship and connection relationship of the overall sensor provided by an exemplary embodiment and the robot, workpiece, industrial computer and control cabinet are as follows: Figure 1As shown, it mainly includes a gas cylinder 10, a welding machine 11 (MIG welding method), a control cabinet 14 (controls the welding operation trajectory of the robot 15 and the start and stop of welding), a robot 15 (the main support for implementing welding operations), a workpiece 16 (the object to be welded), a data acquisition card 13 (for example, a data acquisition card USB4711A, used to input data into an industrial computer 19) and an industrial computer 19 (a data processing and implementation platform for a real-time welding defect monitoring system), and sensors include a Hall sensor 12 (collecting voltage signals and current signals), a sound sensor 17 and a CCD camera 18. The gas cylinder 10 is connected to the welding machine 11, the Hall sensor 12 is connected to the welding machine 11, the welding machine 11 is connected to the robot 15, the robot 15 is connected to the control cabinet 14, the robot 15 is used to implement welding operations on the workpiece 16, the sound sensor 17 and the CCD camera 18 are connected to the robot 15, the Hall sensor 12, the sound sensor 17 and the CCD camera 18 are connected to the data acquisition card 13, and the data acquisition card 13 is connected to the industrial computer 19. In order to protect the sound sensor 17 without blocking the sound propagation path or reducing the amplitude of the sound, a honeycomb sponge can be used to block the physical propagation path of the sound sensor 17 and the welding spatter and smoke. This method effectively prevents spatter and smoke from damaging the sound sensor 17. In order to avoid the influence of strong arc light and collect a clear molten pool image, a 660nm filter and a dimming film can be used to filter out and reduce the strong light, thereby collecting visible and clear molten pool image information. It can be understood that the sensors are not limited to the above examples, and other sensors that can collect welding status data and welding site data can also be used according to needs.
[0030] If the welding quality monitoring method provided by the present application is applied to the boom of an excavator in practice, the model of the boom of a specific engineering machinery excavator can be fixedly monitored, so that the changes in the production process are reduced, and the freedom of the type of workpiece, the posture of the robot, the welding angle, the geometric dimensions of the occupied space and the path passed can be reduced, which provides the possibility for the layout construction and installation of the hardware, and thus provides the physical conditions for real-time quality monitoring of welding. It can focus on a specific excavator boom welding work station, and realize the simplification and adaptation of variable working conditions. After the welding station is limited, the workpiece, the weld position, the welding process, the robot working posture, the welding angle and the working path are simplified from the more complex and changeable scenes, and through further adjustment, it can adapt to the changing working conditions, which enables the sensor entity of the welding monitoring equipment to smoothly collect information. In addition, the welding process has strong noise, smoke, and strong light due to high current and high voltage, which brings great uncertainty to the smooth information collection of the welding monitoring system. By constructing appropriate protective devices, it is ensured that the quality monitoring device can work stably in the near-welding position with high current, high voltage, strong sound, smoke splashing, strong light and serious heat accumulation. The shielding device is used to protect the internal current of the sensor and the internal hard disk and other fine electronic equipment of the acquisition system, so that the entire set of acquisition and quality monitoring devices can work stably for a long time.
[0031] Exemplary Methods
[0032] Figure 2 is a flow chart of a welding quality monitoring method provided by an exemplary embodiment of the present application, such as Figure 2 As shown, the welding quality monitoring method includes:
[0033] Step 100: Acquire welding status data and welding site data during the welding process.
[0034] While keeping the process unchanged, install sound sensors, voltage sensors (such as Hall sensors), current sensors (such as Hall sensors) and visual sensors (such as CCD cameras) to collect welding state data and welding site data during the welding process, wherein the welding state data may include acoustic signals, electrical signals and visual signals, and the welding site data may include parameters such as initial plate temperature, welding time, density, specific heat capacity, thermal conductivity, thermal diffusion coefficient, heat source efficiency, plate parameters, moving speed, calculation time interval, heat input and heat source parameters. These data can provide references for subsequent model construction, for example, they can be used to construct training sets and as inputs to models.
[0035] For example, the boom of the key component of the excavator can be welded by MIG (metal inert gas welding). MIG welding is a complex and important process, which directly affects the structural strength of the boom and the performance of the whole machine. During the MIG welding process, CCD cameras, sound sensors (40kHz), voltage sensors, and current sensors (currently commonly used Hall sensors can be measured at the same time) are used to collect various effective information of key visual, acoustic, and electrical signals in the process.
[0036] Step 200: Process the welding state data to obtain a feature vector.
[0037] After obtaining the welding state data, it is necessary to perform preprocessing operations such as cleaning, labeling, and normalization on the welding state data to ensure the quality and consistency of the data, and convert it into a form that the model can understand and process. In addition, it is necessary to perform feature extraction on the welding state data to extract features that are useful to the model. These feature vectors will be input into the subsequent defect detection model.
[0038] Step 300: Using the feature vector as the input of the defect detection model, obtaining a quality evaluation of the output of the defect detection model.
[0039] When the model input has been established, in order to adapt to large changes in welding conditions, ensemble learning can be used to build the final defect detection model, such as integrating deep neural networks, support vector machines, K-nearest neighbor algorithms, decision trees, and extreme gradient boosting trees. Through these welding defect classification models based on different mechanisms, the features of welding situations in different states can be effectively processed based on different prior assumptions, the welding process can be described from multiple different angles, and the defects can be identified and analyzed in multiple dimensions, which increases the overall model accuracy and avoids the problem that it is difficult to fully and accurately describe the welding process from a single angle.
[0040] Step 400: Using the welding site data and historical welding data as inputs of the heat source model, the molten pool state output by the heat source model is obtained.
[0041] During the welding process, when the workpiece is partially heated to a molten state, the liquid metal area formed is called a molten pool. The molten pool is the result of concentrated heat during welding. The temperature of the molten pool has a direct impact on the welding quality. Different welding site data are collected at the welding site, and the corresponding welding penetration analysis is performed through the heat source model. The welding parameters such as the weld depth and weld width are simulated. The final output result has the ability to adapt to a variety of complex situations and has a good practical application effect on site. Adding a heat source model is equivalent to adding a mechanism model to it, providing information supplement when making judgments without sufficient information.
[0042] Step 500: coupling the molten pool state and quality evaluation to obtain a welding quality evaluation result.
[0043] There is a close relationship between the state of the molten pool and the welding quality. The molten pool is a liquid metal part with a certain geometric shape formed by the molten electrode metal and the partially melted base material during the welding process. Its state directly affects the final quality and performance of the welding. For example, when the temperature of the molten pool is too high, the molten iron is easy to flow down, and the back of the single-sided welding and double-sided forming is easy to burn through, forming a weld nodule. At the same time, the gas in the molten pool may not have time to escape at high temperature, forming pores. When the temperature of the molten pool is low, the molten pool is small, the molten iron has poor fluidity, and it is easy to produce defects such as incomplete penetration and incomplete fusion. Therefore, combined with the state of the molten pool, the welding state can be predicted and monitored from the welding mechanism, and the burn-through, incomplete penetration, welding deviation, wire jamming and pores in the welding process can be effectively monitored to achieve effective monitoring of the welding process. The stability and generalization performance of the welding process have been greatly improved, and the robustness of real-time monitoring of welding has been significantly improved.
[0044] In one embodiment, the welding site data may include: initial plate temperature, welding time, density, specific heat capacity, thermal conductivity, thermal diffusion coefficient, heat source efficiency, plate parameters, moving speed, calculation time interval, heat input and heat source parameters; wherein, step 400 (taking the welding site data and historical welding data as inputs of the heat source model to obtain the molten pool state output by the heat source model) includes: inputting the initial plate temperature, welding time, density, specific heat capacity, thermal conductivity, thermal diffusion coefficient, heat source efficiency, plate parameters, moving speed, calculation time interval, heat input and heat source parameters into the temperature field analytical model to obtain temperature field data calculated by the temperature field analytical model; calculating the weld width and the weld depth according to the temperature field data to obtain theoretical calculation results; taking the theoretical calculation results, historical welding data, workpiece parameters, voltage, current, sound data, weld bead shape, process parameters, image data and heat source parameters as inputs of the heat source model to analyze the welding penetration and obtain the molten pool state.
[0045] The temperature field analytical model is an important tool for describing and predicting temperature distribution in the field of physics and engineering. It is usually based on mathematical and physical principles and simulates and analyzes the changes in the temperature field by establishing a mathematical model. The temperature field refers to the collection of temperatures at various points in a material system. It is a function of time and space coordinates and reflects the distribution of temperature in space and time. In industries such as metallurgy and material processing, temperature field analysis is crucial for controlling product quality and process. Through the temperature field analytical model, the temperature distribution of the material during heating and cooling can be predicted, thereby optimizing the process parameters. Therefore, the temperature field analytical model can assist in improving the welding heat source neural network adaptive intelligent matching generalization model (heat source model). First, different data are collected at the welding site, including initial plate temperature, welding time, density, specific heat capacity, thermal conductivity, thermal diffusion, heat source efficiency, plate parameters, moving speed, calculation time interval, heat input and heat source parameters. These data are used to calculate the temperature field analytical model, and the corresponding melt width and melt depth are derived according to the temperature field data obtained by the theoretical analytical temperature field analytical model. The results of these theoretical calculations are combined with historical data, workpiece parameters, voltage, current, sound data, weld shape, process parameters, image data and heat source parameters as the input of the heat source model to perform the corresponding welding penetration analysis. The final output results have the ability to adapt to a variety of complex situations and have good practical application effects on site. This improvement enables the welding heat source neural network adaptive intelligent matching generalization model to be applied in actual welding, while avoiding the defect detection model from becoming an empirical model that is difficult to face unknown welding situations, and improves the generalization ability and robustness of the model. The heat source model can use a double ellipsoid heat source or a single ellipsoid heat source.
[0046] Figure 3 is a flow chart of a welding quality monitoring method provided by another exemplary embodiment of the present application, such as Figure 3 As shown, the welding quality monitoring method may also include:
[0047] Step 600: Integrate multiple welding defect classification models as a defect detection model.
[0048] Among them, the welding defect classification models include deep neural networks, support vector machines, K nearest neighbor algorithms, decision trees, and extreme gradient boosting trees.
[0049] The welding defect classification model is used to monitor and evaluate the welding quality in real time during the welding process. It is a method to divide the welding quality into different categories or grades. It is based on a series of quality standards and evaluation indicators to conduct a comprehensive and systematic evaluation of the welded joint. This model helps to identify possible quality problems in the welding process and provide guidance for subsequent quality improvement. The welding defect classification model can provide evaluation criteria and reference basis for the defect detection model. The defect detection model can provide data support for the update and improvement of the welding defect classification model through real-time data monitoring and analysis, and continuously optimize and adjust the evaluation criteria and indicators of the welding defect classification model. In addition, integrating multiple welding defect classification models (also known as model fusion or model integration) is an effective method to improve model performance and robustness. For example, the model integration method can adopt ordinary voting, each welding defect classification model independently predicts, and the prediction result of the final defect detection model adopts the principle of minority obeys majority, or adopts a weighted voting method. On the basis of ordinary voting, different weights are assigned to the prediction results of each welding defect classification model. The weight can be determined according to the performance (such as accuracy) of the welding defect classification model.
[0050] Wherein, step 300 (taking the feature vector as the input of the defect detection model to obtain the quality evaluation of the defect detection model output) includes:
[0051] Step 310: Using the feature vector as input to multiple welding defect classification models to obtain multiple quality evaluations.
[0052] After obtaining multiple quality evaluations, the output results of multiple welding defect classification models can be evaluated first. If the overall performance of some welding defect classification models is too poor, they are first excluded, and then a set of welding defect classification models with better effects are screened out to evaluate the accuracy of welding defect classification models on training and validation sets. In other words, the output results of multiple welding defect classification models can be screened first to improve the reliability of the final evaluation results.
[0053] Step 320: Based on the evaluation accuracy of the welding defect classification model, multiple quality evaluations are weighted.
[0054] After evaluating the accuracy of the welding defect classification model on the validation set, the likelihood of the prediction results of the welding defect classification model is weighted according to this accuracy. For example, a higher degree of confidence is given to the prediction of the model with better performance (higher accuracy), and therefore a higher weight is given to it to improve the accuracy of the final quality evaluation.
[0055] Step 330: summing the predicted types and predicted probabilities of the weighted multiple welding defect classification models.
[0056] Each model will give a prediction result, for example, the result is: "This data is likely to be category A, the possibility is 60%; it may also be category B, the possibility is 40%". Because different weights are added to different welding defect classification models, these prediction results and their corresponding weights will be combined. Specifically, the prediction probability of each welding defect classification model for each category is multiplied by its weight, and then the prediction probability of all welding defect classification models for the same category is added together, which is the sum of the prediction results.
[0057] Step 340: Take the prediction type with the greatest prediction possibility as the output of the defect detection model to obtain the final quality evaluation.
[0058] The category with the highest total probability of prediction is used as the final prediction result of the defect detection model. This is a weighted voting method for model integration, which is suitable for classification problems. In this way, the accuracy of the validation set based on the prior is used as the final prediction possibility estimate, and this estimate is used as a reference to achieve the probability accuracy compensation optimization of the posterior decision. The final quality evaluation can effectively process the features of welding situations in different states based on different prior assumptions through welding defect classification models based on different mechanisms, describe the welding process from multiple different angles, and identify and analyze the defects in multiple dimensions, which increases the accuracy of the defect detection model and avoids the problem that it is difficult to fully and accurately describe the welding process from a single angle.
[0059] Figure 4 is a flow chart of a welding quality monitoring method provided by another exemplary embodiment of the present application, such as Figure 4 As shown, step 200 (processing welding state data to obtain a feature vector) includes:
[0060] Step 210: Classify the welding status data according to a preset flaw detection method to obtain multiple welding types.
[0061] For example, magnetic particle inspection can be used. The basic principle of magnetic particle inspection is based on the magnetization characteristics of ferromagnetic materials and the response of magnetic powder to magnetic fields. When ferromagnetic materials are magnetized, defects on their surface and near the surface (such as cracks, slag inclusions, hairline, etc.) will cause local distortion of magnetic lines of force and form leakage magnetic fields. These leakage magnetic fields will adsorb and accumulate on the magnetic powder applied to the surface of the workpiece, forming visible magnetic powder accumulation (magnetic traces), thereby indicating the location and shape of the defects. Under appropriate lighting conditions, these magnetic traces can clearly show the outline of the defects, thereby achieving qualitative and quantitative detection of defects. According to the weld quality results given by methods such as magnetic particle inspection of workpiece welds, the welding data can be classified to obtain six welding conditions: normal, burn-through, incomplete penetration, weld deviation, wire jamming, and porosity, and then manually select signal segments with corresponding characteristics to add labels.
[0062] Step 220: Select a signal segment with corresponding characteristics in the welding type and add a label.
[0063] Select signal segments with corresponding characteristics from the six welding situations of normal, burn-through, incomplete penetration, weld deviation, wire jam, and porosity to add labels. After adding labels, the preliminary construction of the six welding data sets is completed. The welding types are not limited to the six welding situations in the above examples, and can also be added, deleted or modified as needed.
[0064] Step 230: pre-process the labeled signal segments to obtain filtered valid data.
[0065] The directly collected signals often contain more environmental noise, which may interfere with the final monitoring results. For this reason, after the data set is labeled, it needs to be processed in advance to remove the interference of environmental white noise. The data set contains image information, sound information, and current and voltage information. Therefore, different filtering processes are performed for each type of information. For example, various filtering methods represented by mean filtering are used to filter and reduce noise for image information. The filtering methods include but are not limited to median filtering and Gaussian filtering; for sound information, voltage information and current information, zero point compensation may be performed on them because of zero drift in the circuit. The implementation method is to calculate the average value in the frequency band without input, and then perform the above-mentioned mean negative vector translation on the overall signal, that is, to obtain the data after zero drift compensation; sound information generally also contains high-frequency noise, so it is also necessary to use Fourier transform to obtain the sound spectrum, and then perform high-frequency truncation processing, and then perform median filtering with a window size of 3, and then perform inverse Fourier transform of the spectrum, that is, to obtain the sound signal after removing high-frequency noise irrelevant to welding quality and performing noise reduction. After filtering the acoustic, electrical and optical signal segments, effective data after filtering the acoustic, electrical and optical signal segments are obtained respectively.
[0066] Step 240: Calibrate the valid data on the time axis to obtain the calibrated valid data within a preset time period.
[0067] After preprocessing, the signal needs to be made into a batch processing format that can be input into the quality detection model, which requires time calibration of the visual signal, sound signal, voltage signal and current signal on the time axis.
[0068] Step 250: align and package the valid data calibrated within the preset time period to form a data packet.
[0069] After calibration, the visual signals, sound signals, voltage signals, and current signals of the corresponding time period are aligned as required, and the aligned data can be packaged to form a data packet. For example, a data packet of a defect detection model includes a visual signal, two thousand sound signals, voltage signals, and current signals. This packaged data is the training data set for the selected time node, and the category to which the data belongs is the label used in the training process.
[0070] Step 260: Extract features from the data packet to obtain extracted feature parameters.
[0071] After the data packet is prepared, features can be extracted from it to input into the possible welding defect classification model. Different extraction methods can be used for visual signals, sound signals, voltage signals and current signals to extract feature parameters in a targeted manner.
[0072] Step 270: Perform importance evaluation and uniqueness evaluation on the feature parameters.
[0073] The above-mentioned feature parameters extracted from visual signals, sound signals, voltage signals and current signals have the value and possibility of welding quality monitoring, but the excessive number of features has the risk of overfitting and reduced generalization ability. Therefore, two dimensions can be used to screen the features, namely importance evaluation and uniqueness evaluation. Before the evaluation, the features from different sources must be standardized to reduce the unexpected impact of different distribution intervals of the features. Importance evaluation refers to the importance of the current feature, that is, when there is only one feature parameter, the generalization performance of the model accounts for the proportion of the generalization performance when all features exist; uniqueness evaluation refers to the loss of generalization performance of the model obtained after deleting this feature from all features and then training the model compared to the loss when it is not deleted. Through these two specific evaluation methods, the most important and indispensable features can be effectively selected.
[0074] Step 280: Perform feature merging on the feature parameters that have passed the importance evaluation and the uniqueness evaluation to obtain a feature vector.
[0075] After completing the importance and uniqueness evaluation of the features, you can select the features with higher importance and uniqueness, simplify the entire feature set, merge the filtered features, and then construct the feature vector, which can be used as the input of the defect detection model.
[0076] Figure 5 is a flow chart of a welding quality monitoring method provided by another exemplary embodiment of the present application, such as Figure 5 As shown, the welding state data includes visual signals, acoustic signals and electrical signals; wherein step 260 (extracting features from the data packet to obtain extracted feature parameters) includes:
[0077] Step 261: extract image feature information from the visual signal using the autoencoder network as a feature extraction method to obtain autoencoder feature parameters.
[0078] For visual signals, i.e. image information, a convolutional autoencoder network can be used as a means of encoding and decoding features. After training a fully convolution-based autoencoder and autodecoder network, the autoencoder network is used as a means of extracting features to extract feature information from the denoised image, i.e., autoencoder feature parameters. The feature parameters extracted by the autoencoder can be used for data dimensionality reduction, helping to better understand the intrinsic structure of the data, and can also achieve effective data compression, which has a lower dimension than the original data, thereby reducing the cost of storage and transmission.
[0079] Step 262: extract the spectrum of the acoustic signal based on Fourier transform, and sum, average and variance the acoustic signal according to the interval of the spectrum to obtain the frequency domain feature parameters of the sound.
[0080] For the extraction of acoustic signals, that is, the extraction of frequency domain features of sound information, the frequency domain information of sound is obtained by Fourier transform. The frequency domain features are mainly extracted from two aspects, one is the energy of the sound, and the other is the variance of the sound. These extracted features are summed, averaged, and varianceed according to the frequency spectrum interval to serve as the frequency domain feature parameters of the sound.
[0081] Step 263: extract the time domain statistical features of the acoustic signal and the electrical signal to obtain the time domain feature parameters.
[0082] For acoustic signals and electrical signals (voltage signals and current signals), statistical features in the time domain are used to extract features. These time domain feature parameters include but are not limited to mean, peak, standard deviation, root mean square, peak factor, skewness, kurtosis, zero crossing rate, waveform factor, variance, pulse index, energy, tolerance index, square root amplitude, peak and average amplitude, etc.
[0083] Correspondingly, step 270 (evaluating the importance and uniqueness of the feature parameters) includes:
[0084] Step 271: perform importance evaluation and uniqueness evaluation on the self-encoding feature parameters, frequency domain feature parameters and time domain feature parameters.
[0085] The autoencoder, time domain and frequency domain information extracted from the visual signal, voltage signal, current signal and sound signal mentioned above have the value and possibility of welding quality monitoring, but the excessive number of features has the risk of overfitting and reduced generalization ability. Therefore, two dimensions are used here to screen the features, namely importance evaluation and uniqueness evaluation. Only the features that pass the importance evaluation and uniqueness evaluation are used as the input of the defect detection model.
[0086] In one embodiment, step 271 (importance evaluation of self-encoding feature parameters, frequency domain feature parameters and time domain feature parameters) includes: inputting the self-encoding feature parameters, frequency domain feature parameters and time domain feature parameters into the welding defect classification model respectively to obtain multiple quality prediction accuracy rates; inputting all the self-encoding feature parameters, frequency domain feature parameters and time domain feature parameters into the welding defect classification model to obtain the overall quality prediction accuracy rate; taking any one of the multiple quality prediction accuracy rates as the target quality prediction accuracy rate, when the ratio of the target quality prediction accuracy rate to the overall quality prediction accuracy rate is greater than or equal to the preset ratio, it is determined that the feature parameters corresponding to the target quality prediction accuracy rate pass the importance evaluation.
[0087] Importance evaluation can be reflected by the percentage of the overall accuracy when only one feature is used as the input of the welding quality monitoring model to the accuracy when all features are used as input. For example, the extracted autoencoder feature parameters, frequency domain feature parameters and time domain feature parameters are respectively input into the welding defect classification model (in order to obtain a wide spectrum of accuracy and generalization performance, and to avoid the influence of the model's prior assumptions, a support vector machine can be used as the model of the evaluation result), and the obtained quality prediction accuracy is compared with that when all features are input, that is, the importance of the feature is obtained. Features that pass the importance evaluation indicate that they have a greater impact on the accuracy of the model output among the features and cannot be removed. Features that fail the importance evaluation can be optimized and deleted.
[0088] In one embodiment, step 271 (performing uniqueness evaluation on self-encoding feature parameters, frequency domain feature parameters and time domain feature parameters) includes: forming a feature set with self-encoding feature parameters, frequency domain feature parameters and time domain feature parameters; inputting the entire feature set into a welding defect classification model to obtain an overall quality prediction accuracy; taking any one of the self-encoding feature parameters, frequency domain feature parameters and time domain feature parameters as a target feature parameter, and deleting the target feature parameter from the feature set to obtain a non-target feature parameter set; inputting the non-target feature parameter set into the welding defect classification model to obtain the prediction accuracy of the non-target feature parameters; calculating the decrease value of the prediction accuracy of the non-target feature parameters relative to the overall quality prediction accuracy; when the percentage of the decrease value in the overall quality prediction accuracy is greater than or equal to the preset percentage, it is determined that the target feature parameter passes the uniqueness evaluation.
[0089] The uniqueness evaluation can be reflected by the decrease in model accuracy when the feature is lost. For example, after deleting the autoencoder feature parameter, frequency domain feature parameter, and time domain feature parameter from the entire feature set, the decrease in model accuracy is reflected as a percentage of the original overall accuracy. This ratio can approximate the uniqueness of the feature. Features that pass the uniqueness evaluation are indispensable among the features, have a greater impact on the model output, and cannot be removed. Features that fail the importance evaluation can be optimized and deleted.
[0090] In addition, importance assessment and uniqueness assessment can be combined for assessment. Features that pass both importance assessment and uniqueness assessment are required to be input, features that pass only one of importance assessment and uniqueness assessment can be selectively input, and features that fail both assessments can be optimized or deleted.
[0091] In one embodiment, the welding status data includes visual signals, acoustic signals and electrical signals. Step 230 (preprocessing the labeled signal segments to obtain filtered valid data) includes: filtering and denoising the labeled visual signal segments to obtain filtered visual valid data; performing spectral analysis on the labeled acoustic signal segments to remove high-frequency noise that is not related to the welding quality to obtain denoised acoustic valid data; performing zero drift compensation on the labeled acoustic signal segments and electrical signal segments to obtain compensated acoustic valid data, current valid data and voltage valid data.
[0092] For image information, various filtering methods represented by mean filtering are used for filtering and noise reduction. The filtering methods include but are not limited to median filtering and Gaussian filtering. For acoustic signals, high-frequency noise may be contained, so Fourier transform is used to obtain the spectrum of the sound, and then high-frequency truncation processing is performed, and then median filtering with a window size of 3 is performed, and then the spectrum is inverse Fourier transform is performed to obtain the sound signal after removing high-frequency noise and noise reduction. For acoustic signal fragments and electrical signal (current signal and voltage signal) fragments, zero point compensation may be performed on them because of zero point drift in the circuit. The purpose of zero point compensation is to eliminate or correct the measurement error caused by internal factors of the equipment (such as sensor offset, circuit imbalance, etc.), so that the reading of the measuring instrument can accurately display zero when there is no external signal input.
[0093] Among them, step 240 (calibrating the valid data on the time axis to obtain the valid data calibrated within a preset time period) includes: calibrating the visual valid data, acoustic valid data, current valid data and voltage valid data on the time axis according to the collection time.
[0094] For example, image acquisition devices (such as cameras and camcorders) usually have the function of recording timestamps, that is, when each image or video frame is taken, an accurate time information is attached. By reading these timestamps, the visual effective data can be accurately matched with its shooting time. Acoustic effective data is digitized at a certain sampling rate, that is, the number of samples collected per second. By knowing the sampling rate and starting timestamp of the audio file, the specific time corresponding to each acoustic effective data can be calculated. In addition, the sensor may contain information such as recording time and encoder settings when collecting. These metadata can be used to assist time calibration. When collecting current effective data and voltage effective data, the timestamp of each sampling point can be recorded. By analyzing these data, the time corresponding to each voltage and current value can be accurately known. In the data processing stage, different types of data are aligned in time according to information such as the timestamp or sampling rate of each device.
[0095] In addition, after the defect detection model and the heat source model are established, the parameters of the model need to be selected. In an exemplary embodiment, a parameter optimization method including but not limited to a hybrid composite optimization algorithm can be used to select and optimize the hyperparameters. The hybrid composite optimization algorithm proposed here refers to an optimization algorithm that combines the advantages of multiple parameter optimization methods. The process is as follows: first, a larger scale grid search is performed, and the uniformity of the grid search is used to avoid the influence of singular points on the overall model parameter search range and focus at the beginning, and try to avoid falling into the local optimal situation; secondly, particle swarm optimization and simulated annealing algorithms are performed on the basis of grid search. Particle swarm optimization can perform effect-based evolution of model parameters near the grid based on grid search, and simulated annealing algorithms can accelerate this process and avoid the problem of increased complexity caused by multiple optimizations. The combination of the two ultimately greatly increases the probability that the result stays at the global optimal; finally, Bayesian optimization based on the above optimization results is used to further select and tune the overall parameters, and further search near the results and where the global optimal position may appear, and finally obtain the parameter optimization result.
[0096] The specific implementation is to first limit the data type of each parameter, then limit the range of its variation, determine the number of steps for random search and optimization search of hyperparameters, and then run the program to select and optimize hyperparameters. Among them, hyperparameters are an important type of parameters in machine learning and deep learning models. They are set before model training, rather than learned through the training process.
[0097] Figure 6 is an overall flow chart of a welding quality monitoring method provided by an exemplary embodiment of the present application. Figure 6As shown, S1: Information collection, using CCD camera, sound sensor (40kHz), voltage sensor, current sensor (currently commonly used Hall sensor can be measured at the same time), can collect visual signals, voltage signals, current signals and sound signals. S2: Data set construction, data classification and label assignment. For example, according to the results of magnetic particle inspection and other flaw detection methods, the data is classified, the welds and welds of the same category are classified into one category, and the signal segments with characteristics are selected for label setting. S3: Information preprocessing, image denoising, high-frequency noise removal and zero point compensation, that is, different filtering methods are implemented for different signals. S4: Alignment and packaging, time calibration, data alignment and data packaging, the packaged data is the training data set of the selected time node, and the category to which the data belongs is the label used in the training process. S5: Feature extraction, convolution feature, statistical feature and frequency domain feature extraction, that is, different feature extraction methods are implemented for different signals. S6: Feature screening and fusion, importance evaluation and uniqueness evaluation, and then feature merging and feature vector construction. S7: Model establishment optimization, integrating different mechanism models as defect detection models, optimizing parameters based on the parameter optimization of hybrid composite optimization algorithm, improving the generalization model of welding heat source neural network adaptive intelligent matching, and forming the heat source model output molten pool state. S8: Evaluation result coupling, coupling evaluation of the results output by the defect detection model and the results output by the heat source model, and realizing the optimization of the a priori-based a posteriori compensation decision. Among them, the a priori-based a posteriori compensation decision optimization is a decision-making method that combines a priori knowledge and a posteriori information, which aims to optimize the decision-making process by utilizing a priori knowledge (knowledge or belief before data observation) and a posteriori information (knowledge or belief updated after data observation).
[0098] Exemplary Devices
[0099] Figure 7 is a structural schematic diagram of a welding quality monitoring device provided by an exemplary embodiment of the present application, such as Figure 7 As shown, the welding quality monitoring device 7 includes: an acquisition module 71, which is used to acquire welding state data and welding site data during the welding process; a processing module 72, which is used to process the welding state data to obtain a feature vector; a first evaluation module 73, which is used to use the feature vector as an input of a defect detection model to obtain a quality evaluation of the defect detection model output; a second evaluation module 74, which is used to use the welding site data and historical welding data as inputs of a heat source model to obtain a molten pool state output by the heat source model; and a coupled evaluation module 75, which is used to couple the molten pool state and quality evaluation to obtain a welding quality evaluation result.
[0100] The welding quality monitoring device and excavator provided in the present application can effectively predict and monitor the welding status based on the welding mechanism. On the basis of quality evaluation based on the welding status data, a heat source model is added to highly simulate the temperature field of the welding process. The welding quality is finally evaluated in combination with the molten pool state and the corresponding quality evaluation. Faults in the welding process are effectively monitored, which greatly improves the stability and generalization performance of the welding process and makes remarkable improvements to the robustness of real-time monitoring of welding.
[0101] In one embodiment, the welding site data may include: initial plate temperature, welding time, density, specific heat capacity, thermal conductivity, thermal diffusion coefficient, heat source efficiency, plate parameters, moving speed, calculation time interval, heat input and heat source parameters; wherein, the second evaluation module 74 may be configured as: inputting the initial plate temperature, welding time, density, specific heat capacity, thermal conductivity, thermal diffusion coefficient, heat source efficiency, plate parameters, moving speed, calculation time interval, heat input and heat source parameters into the temperature field analytical model to obtain temperature field data calculated by the temperature field analytical model; calculating the weld width and weld depth according to the temperature field data to obtain theoretical calculation results; using the theoretical calculation results, historical welding data, workpiece parameters, voltage, current, sound data, weld bead shape, process parameters, image data and heat source parameters as inputs to the heat source model to analyze the welding penetration and obtain the molten pool state.
[0102] In one embodiment, the welding quality monitoring device 7 can be configured as: integrating multiple welding defect classification models as defect detection models, the welding defect classification models include deep neural networks, support vector machines, K-nearest neighbor algorithms, decision trees, and extreme gradient boosting trees; correspondingly, the first evaluation module 73 can also be configured as: taking feature vectors as inputs of multiple welding defect classification models to obtain multiple quality evaluations; weighting multiple quality evaluations based on the evaluation accuracy of the welding defect classification models; summing the predicted types of the weighted multiple welding defect classification models; and taking the predicted type with the greatest predicted possibility as the output of the defect detection model to obtain the final quality evaluation.
[0103] In one embodiment, the processing module 72 can be configured as follows: classify the welding status data according to a preset flaw detection method to obtain a plurality of welding types; select signal segments with corresponding characteristics from the welding types and add labels; preprocess the labeled signal segments to obtain filtered valid data; calibrate the valid data on the time axis to obtain valid data calibrated within a preset time period; align and package the valid data calibrated within the preset time period to form a data packet; extract features from the data packet to obtain extracted feature parameters; perform importance evaluation and uniqueness evaluation on the feature parameters; perform feature merging on the feature parameters that pass the importance evaluation and uniqueness evaluation to obtain a feature vector.
[0104] In one embodiment, the processing module 72 can also be configured as follows: using an autoencoder network as a feature extraction method to extract image feature information from a visual signal to obtain autoencoder feature parameters; extracting the spectrum of an acoustic signal based on Fourier transform, and summing, averaging, and calculating the variance of the acoustic signal according to the interval of the spectrum to obtain frequency domain feature parameters of the sound; extracting time domain statistical features of acoustic signals and electrical signals to obtain time domain feature parameters; and performing importance and uniqueness evaluation on autoencoder feature parameters, frequency domain feature parameters, and time domain feature parameters.
[0105] In one embodiment, the processing module 72 can also be configured as: inputting the self-encoding characteristic parameters, frequency domain characteristic parameters and time domain characteristic parameters into the welding defect classification model respectively to obtain multiple quality prediction accuracy rates; inputting all the self-encoding characteristic parameters, frequency domain characteristic parameters and time domain characteristic parameters into the welding defect classification model to obtain the overall quality prediction accuracy rate; taking any one of the multiple quality prediction accuracy rates as the target quality prediction accuracy rate, when the ratio of the target quality prediction accuracy rate to the overall quality prediction accuracy rate is greater than or equal to the preset ratio, determining that the characteristic parameters corresponding to the target quality prediction accuracy rate pass the importance evaluation.
[0106] In one embodiment, the processing module 72 can also be configured as: forming a feature set with self-encoding feature parameters, frequency domain feature parameters and time domain feature parameters; inputting the entire feature set into a welding defect classification model to obtain an overall quality prediction accuracy; taking any one of the self-encoding feature parameters, frequency domain feature parameters and time domain feature parameters as a target feature parameter, and deleting the target feature parameter from the feature set to obtain a non-target feature parameter set; inputting the non-target feature parameter set into the welding defect classification model to obtain the prediction accuracy of the non-target feature parameters; calculating the decrease value of the prediction accuracy of the non-target feature parameters relative to the overall quality prediction accuracy; when the percentage of the decrease value in the overall quality prediction accuracy is greater than or equal to the preset percentage, it is determined that the target feature parameter passes the uniqueness evaluation.
[0107] In one embodiment, the welding status data includes visual signals, acoustic signals and electrical signals, and the processing module 72 can also be configured to: filter and reduce noise on the visual signal segments after adding labels to obtain filtered visual valid data; perform spectrum analysis on the acoustic signal segments after adding labels to remove high-frequency noise that is not related to the welding quality, and obtain denoised acoustic valid data; perform zero drift compensation on the acoustic signal segments and electrical signal segments after adding labels to obtain compensated acoustic valid data, current valid data and voltage valid data; calibrate the visual valid data, acoustic valid data, current valid data and voltage valid data on the time axis according to the time of collection.
[0108] Exemplary Electronic Devices
[0109] An electronic device includes: a processor; a memory for storing instructions executable by the processor; and the processor is used to execute the welding quality monitoring method described in the embodiment provided in the present application.
[0110] Below, reference Figure 8 The electronic device according to the embodiment of the present application is described. The electronic device may be any one or both of the first device and the second device, or a stand-alone device independent of them, and the stand-alone device may communicate with the first device and the second device to receive the collected input signal from them.
[0111] Figure 8 A block diagram of an electronic device according to an embodiment of the present application is illustrated.
[0112] like Figure 8 As shown, the electronic device 80 includes one or more processors 81 and a memory 82 .
[0113] The processor 81 may be a central processing unit (CPU) or other forms of processing units having data processing capabilities and / or instruction execution capabilities, and may control other components in the electronic device 80 to perform desired functions.
[0114] The memory 82 may include one or more computer program products, which may include various forms of computer-readable storage media, such as volatile memory and / or non-volatile memory. The volatile memory may include, for example, random access memory (RAM) and / or cache memory (cache), etc. The non-volatile memory may include, for example, read-only memory (ROM), hard disk, flash memory, etc. One or more computer program instructions may be stored on the computer-readable storage medium, and the processor 81 may run the program instructions to implement the welding quality monitoring method of each embodiment of the present application described above and / or other desired functions. Various contents such as input signals, signal components, noise components, etc. may also be stored in the computer-readable storage medium.
[0115] In one example, the electronic device 80 may further include: an input device 83 and an output device 84 , and these components are interconnected via a bus system and / or other forms of connection mechanisms (not shown).
[0116] When the electronic device is a stand-alone device, the input device 83 may be a communication network connector for receiving collected input signals from the first device and the second device.
[0117] In addition, the input device 83 may also include, for example, a keyboard, a mouse, etc.
[0118] The output device 84 can output various information to the outside, including the determined distance information, direction information, etc. The output device 84 can include, for example, a display, a speaker, a printer, a communication network and a remote output device connected thereto, and the like.
[0119] Of course, to simplify, Figure 8 Only some of the components related to the present application in the electronic device 80 are shown, and components such as a bus, an input / output interface, etc. are omitted. In addition, according to specific application situations, the electronic device 80 may also include any other appropriate components.
[0120] The computer program product may be written in any combination of one or more programming languages to write program codes for performing the operations of the embodiments of the present application, including object-oriented programming languages, such as Java, C++, etc., and conventional procedural programming languages, such as "C" language or similar programming languages. The program code may be executed entirely on the user computing device, partially on the user device, as an independent software package, partially on the user computing device and partially on a remote computing device, or entirely on a remote computing device or server.
[0121] A computer-readable storage medium stores a computer program, and the computer program is used to execute the welding quality monitoring method described in the embodiment provided in this application.
[0122] The computer readable storage medium can adopt any combination of one or more readable media. The readable medium can be a readable signal medium or a readable storage medium. The readable storage medium can include, for example, but is not limited to, a system, device or device of electricity, magnetism, light, electromagnetic, infrared, or semiconductor, or any combination of the above. More specific examples (non-exhaustive list) of readable storage media include: an electrical connection with one or more wires, a portable disk, a hard disk, a random access memory (RAM), a read-only memory (ROM), an erasable programmable read-only memory (EPROM or flash memory), an optical fiber, a portable compact disk read-only memory (CD-ROM), an optical storage device, a magnetic storage device, or any suitable combination of the above.
[0123] The above description has been given for the purpose of illustration and description. In addition, this description is not intended to limit the embodiments of the present application to the forms disclosed herein. Although multiple example aspects and embodiments have been discussed above, those skilled in the art will recognize certain variations, modifications, changes, additions and sub-combinations thereof.
Claims
1. A welding quality monitoring method, characterized in that: include: Obtain welding status data and welding site data during the welding process; Processing the welding state data to obtain a feature vector; Using the feature vector as an input of a defect detection model to obtain a quality evaluation of an output of the defect detection model; Using the welding site data and historical welding data as inputs of a heat source model, obtaining a molten pool state output by the heat source model; The molten pool state and the quality evaluation are coupled to obtain a welding quality evaluation result.
2. The welding quality monitoring method according to claim 1, characterized in that: Welding site data include: initial plate temperature, welding time, density, specific heat capacity, thermal conductivity, thermal diffusion coefficient, heat source efficiency, plate parameters, moving speed, calculation time interval, heat input and heat source parameters; The welding site data and historical welding data are used as inputs of the heat source model to obtain the molten pool state output by the heat source model, including: Inputting the initial plate temperature, welding time, density, specific heat capacity, thermal conductivity, thermal diffusion coefficient, heat source efficiency, plate parameters, moving speed, calculation time interval, heat input and heat source parameters into the temperature field analytical model to obtain temperature field data calculated by the temperature field analytical model; Calculating the melt width and melt depth according to the temperature field data to obtain theoretical calculation results; The theoretical calculation results, historical welding data, workpiece parameters, voltage, current and sound data, weld bead shape, process parameters, image data and heat source parameters are used as inputs of the heat source model to analyze the welding penetration and obtain the molten pool state.
3. The welding quality monitoring method according to claim 1, characterized in that: Welding quality monitoring methods also include: Integrate multiple welding defect classification models as a defect detection model; wherein the welding defect classification model includes a deep neural network, a support vector machine, a K nearest neighbor algorithm, a decision tree, and an extreme gradient boosting tree; Wherein, taking the feature vector as the input of the defect detection model and obtaining the quality evaluation of the output of the defect detection model includes: Using the feature vector as input to multiple welding defect classification models to obtain multiple quality evaluations; Based on the evaluation accuracy of the welding defect classification model, weighting the plurality of quality evaluations; The predicted types of the multiple welding defect classification models after weighted processing are summed up; The prediction type with the greatest prediction possibility is used as the output of the defect detection model to obtain the final quality evaluation.
4. The welding quality monitoring method according to claim 1, characterized in that: Processing the welding state data to obtain a feature vector includes: Classify welding status data according to preset flaw detection methods to obtain various welding types; Selecting a signal segment with corresponding characteristics in the welding type and adding a label; Preprocess the labeled signal segments to obtain filtered valid data; Calibrate the valid data on the time axis to obtain the valid data calibrated within a preset time period; Align and package the valid data calibrated within a preset time period to form a data packet; Extract features from the data packet to obtain extracted feature parameters; Performing importance evaluation and uniqueness evaluation on the characteristic parameters; Feature merging is performed on the feature parameters that pass the importance evaluation and the uniqueness evaluation to obtain a feature vector.
5. The welding quality monitoring method according to claim 4, characterized in that: The welding state data includes visual signals, acoustic signals and electrical signals; wherein, feature extraction is performed from the data packet to obtain the extracted feature parameters, including: The image feature information in the visual signal is extracted by using the autoencoder network as a feature extraction method to obtain the autoencoder feature parameters; Extracting the spectrum of the acoustic signal based on Fourier transform, and summing, averaging and variance-calculating the acoustic signal according to the interval of the spectrum to obtain the frequency domain characteristic parameters of the sound; Extract the time domain statistical characteristics of acoustic signals and electrical signals to obtain time domain characteristic parameters; The importance and uniqueness of the characteristic parameters are evaluated, including: Importance evaluation and uniqueness evaluation are performed on the self-encoding feature parameter, the frequency domain feature parameter and the time domain feature parameter.
6. The welding quality monitoring method according to claim 5, characterized in that: The importance of the autoencoding characteristic parameter, the frequency domain characteristic parameter and the time domain characteristic parameter is evaluated, including: Inputting the self-encoding characteristic parameter, the frequency domain characteristic parameter and the time domain characteristic parameter into a welding defect classification model respectively to obtain multiple quality prediction accuracy rates; Inputting the self-encoding characteristic parameters, the frequency domain characteristic parameters and the time domain characteristic parameters into a welding defect classification model to obtain an overall quality prediction accuracy rate; Taking any one of the multiple quality prediction accuracy rates as the target quality prediction accuracy rate, when the ratio of the target quality prediction accuracy rate to the overall quality prediction accuracy rate is greater than or equal to a preset ratio, it is determined that the characteristic parameter corresponding to the target quality prediction accuracy rate passes the importance evaluation.
7. The welding quality monitoring method according to claim 5, characterized in that: The uniqueness evaluation of the autoencoding characteristic parameter, the frequency domain characteristic parameter and the time domain characteristic parameter includes: Combining the self-encoding feature parameter, the frequency domain feature parameter and the time domain feature parameter into a feature set; Inputting all the feature sets into the welding defect classification model to obtain the overall quality prediction accuracy; Taking any one of the self-encoding feature parameter, the frequency domain feature parameter and the time domain feature parameter as a target feature parameter, and deleting the target feature parameter from the feature set to obtain a non-target feature parameter set; Inputting the non-target characteristic parameter set into a welding defect classification model to obtain a prediction accuracy rate of the non-target characteristic parameters; Calculating a decrease in the prediction accuracy of the targetless feature parameter relative to the overall quality prediction accuracy; When the percentage of the drop value to the overall quality prediction accuracy is greater than or equal to a preset percentage, it is determined that the target characteristic parameter passes the uniqueness evaluation.
8. The welding quality monitoring method according to claim 4, characterized in that: The welding state data includes visual signals, acoustic signals and electrical signals. The signal segments after adding labels are preprocessed to obtain filtered valid data, including: Filter and reduce noise on the labeled visual signal segments to obtain filtered visual valid data; Perform spectrum analysis on the acoustic signal segments after adding labels, remove high-frequency noise irrelevant to welding quality, and obtain effective acoustic data after denoising; Perform zero drift compensation on the acoustic signal segments and electrical signal segments after adding the labels, and obtain compensated acoustic valid data, current valid data and voltage valid data; The valid data is calibrated on the time axis to obtain the valid data calibrated within a preset time period, including: The visual effective data, acoustic effective data, current effective data and voltage effective data are calibrated on the time axis according to the time of collection.
9. A welding quality monitoring device, characterized in that: include: An acquisition module is used to acquire welding status data and welding site data during the welding process; A processing module, used for processing the welding state data to obtain a feature vector; A first evaluation module, used to use the feature vector as an input of a defect detection model to obtain a quality evaluation of an output of the defect detection model; A second evaluation module is used to use the welding site data and historical welding data as inputs of a heat source model to obtain a molten pool state output by the heat source model; A coupling evaluation module is used to couple the molten pool state and the quality evaluation to obtain a welding quality evaluation result.
10. An excavator, characterized in that: include: A sensor, wherein the sensor is installed on the boom of the excavator, and the sensor is used to collect welding state data and welding site data during the boom welding process; A quality monitoring device, wherein the quality monitoring device is communicatively connected to the sensor, and the quality monitoring device is used to execute the welding quality monitoring method described in any one of claims 1 to 8 to monitor the welding quality of the boom of the excavator.
Citation Information
Cited By
Multi-sensor data fusion processing system in welding process
CN120947752A
Electron beam welding process parameter fusion depth prediction method based on supervised learning
CN121670099A
