A neural network training method, storage medium and electronic device for penetration detection control parameter prediction

Abstract

The present invention discloses a neural network training method, storage medium, and electronic device for predicting penetrant testing control parameters. The method uses a neural network model to obtain predicted values ​​of penetrant testing control parameters based on real-time weld morphology, which are used to guide the real-time adjustment of various control parameters in the automated penetrant testing process and obtain a penetrant testing development image containing complete weld defects. The method includes the following steps: obtaining test data, selecting test data with a defect detection rate of 99% as a training data set, training a neural network model, obtaining optimal penetrant testing control parameters to guide testing implementation, and the like. The present invention uses a neural network model training method to accurately predict optimal penetrant testing control parameters based on the correlation between penetrant testing control parameters and weld morphology, and then finely adjusts the values ​​of various control parameters to achieve uniform and comprehensive coverage of welds, significantly improving the weld defect detection rate, providing reliable guarantees for the integrity and accuracy of test results, and simultaneously achieving economical use of reagents.

Description

Technical Field

[0001] The present invention belongs to the field of artificial intelligence detection technology, and in particular relates to a neural network training method, storage medium and electronic equipment for predicting penetration detection control parameters. Background Art

[0002] Welding is an important processing technology for large workpieces. For example, in the processing of key structures such as special vehicle tanks, frames, and fuel tanks, the quality and reliability of the welds directly affect the performance of large workpieces. Therefore, accurately and completely eliminating weak connection welds in the welding process and ensuring that defective welded structures do not enter the next production link are important prerequisites for ensuring the safe use of large workpieces.

[0003] Penetrant testing is one of the most commonly used nondestructive testing methods for weld surface defects. It can effectively detect internal weld defects such as cracks and inclusions, with a high defect detection rate. However, for large welded components such as special vehicle tanks, frames, and fuel tanks, the workpieces are large and the welds are numerous and long. Existing penetrant testing methods mostly rely on manually pressing portable bottled reagents for spraying. Some companies also use automated equipment for automatic spraying. However, the existing penetration detection methods all have the following problems: First, the bottled reagent is a spray can pressure bottle. Different pressing forces and slight changes in parameters such as spraying position, spraying height, and spraying speed will have a great impact on the amount of reagent sprayed and the coverage of the reagent on the weld. It is easy for the reagent to be sprayed unevenly and incompletely on the weld surface, resulting in insufficient reagent in local areas, and the internal defects of the weld cannot be effectively detected, affecting the stability and reliability of the detection results. Excessive reagent in local areas will cause reagent waste and increase the detection cost; second, differences in the shape and size of welding workpieces, welding positions and specific process parameters will lead to differences in weld morphology such as weld width and residual height, and this difference will directly lead to the actual spraying position and height being unable to accurately adapt to the actual weld morphology, resulting in a decrease in spraying accuracy, and there are local areas that are not covered and omitted, and there are still missed detections, which will also affect the accuracy and stability of the final detection results. Summary of the Invention

[0004] The purpose of the present invention is to propose a neural network training method for predicting control parameters of penetration testing, and a readable storage medium and electronic device containing the same, so as to solve the problem that the existing penetration testing method cannot adaptively adjust and finely control spraying parameters such as spraying position, spraying height, and spraying amount according to the shape and morphology of the weld, and thus cannot effectively ensure the uniform and comprehensive coverage of the weld by the detection reagent, and cannot ensure that the reagent acts effectively and uniformly on the weld, resulting in missed detection and reagent waste.

[0005] The present invention is achieved through the following technical solutions:

[0006] The present invention proposes a neural network training method for predicting penetrant testing control parameters. By using a neural network model to obtain predicted values ​​of penetrant testing control parameters based on real-time weld morphology, the method is used to guide the real-time adjustment of various control parameters in the automated penetrant testing process and obtain a penetrant testing development image containing complete weld defects. The method includes the following steps:

[0007] S1, obtain test data:

[0008] Obtain the corresponding type of weld based on the welding workpiece material, plate thickness, and weld joint form as basic parameters, and collect weld morphology data including weld width and reinforcement;

[0009] The penetrant test is designed and executed using the reagent bottle nozzle pressing height, spraying distance, moving speed, and the rest time after reagent spraying as the penetrant test control parameters, and weld defect information from the penetrant test development image is collected;

[0010] Perform metallographic analysis on the welds after penetrant testing to obtain near-surface weld defect information from the metallographic testing image, compare it with the penetrant testing developed image, and calculate the defect detection rate;

[0011] S2, constructing a training data set: selecting the basic parameters, penetrant inspection control parameters, weld morphology data, penetrant inspection development images, and metallographic inspection images corresponding to a defect detection rate of 99% obtained in S1 as the training data set;

[0012] S3, establishing and training a prediction neural network model: constructing an initial neural network model, based on the initial training data set in step S2, importing the neural network model, using weld width, reinforcement, welding workpiece material, plate thickness, welding joint form, penetrant detection development image, and metallographic detection image as input parameters, and using reagent bottle nozzle pressing height, spraying distance, moving speed, and reagent standing time after spraying as output parameters, training and optimizing the initial neural network model until convergence, completing the construction of a prediction neural network model for penetrant detection control parameters;

[0013] S4, obtain the predicted values ​​of the control parameters of the penetrant inspection: Based on the predictive neural network model constructed in S3, manually input the welding workpiece material, plate thickness and welding joint form information, and use the real-time collected weld width and residual height data as input to obtain the predicted values ​​of the current pressing height, spraying distance, moving speed and standing time after spraying. These are used as the optimal penetrant inspection control parameters to guide the implementation of the penetrant inspection, and finally obtain the penetrant inspection development image containing complete weld defects.

[0014] Based on the above technical scheme, the influence of welding process control parameters such as workpiece material, plate thickness, and welding joint form on weld morphology such as weld width and reinforcement is comprehensively considered, and the penetration detection control parameters such as pressing height, spraying distance, moving speed and static time after spraying, as well as the weld defect detection results are combined as the training data basis of the neural network model to carry out targeted model training and optimization, obtain the intrinsic mapping model of penetration detection control parameters and weld morphology information including weld width and reinforcement, and finally obtain the accurate prediction value of penetration detection control parameters, realize fine control of the whole process of penetration detection, ensure uniform and comprehensive coverage of the weld by reagents to ensure the integrity and accuracy of the detection results, and avoid the waste of reagents to the greatest extent, saving detection costs.

[0015] Preferably, the initial training data set of step S2 is divided into different training data subsets according to the welding workpiece material, and each training data subset executes step S3 to train and optimize the initial neural network model, obtains the input-output mapping model and data subset corresponding to the different welding workpiece materials, and re-assembles the input-output mapping data subsets into a training data set, re-imports the neural network model optimized by the training data subset of any material, executes step S3 and finally constructs a comprehensive prediction neural network model for penetrant detection control parameter prediction. Based on this training process, according to the different characteristics of the weld morphology and defect types formed by different welding workpiece materials, the neural network model is optimized and trained in a targeted manner, so that the neural network model can distinguish the materials and accurately obtain the mapping relationship between the weld morphology and the penetrant detection control parameters. At the same time, the neural network model also has the ability to predict the penetrant detection control parameters based on any material, thereby expanding the scope of application of the model.

[0016] More preferably, when obtaining the predicted value of the penetration detection control parameter according to step S4, the following process is also included: according to the manually input welding workpiece material and plate thickness data, the corresponding input-output mapping model and data subset are automatically retrieved as the basic data source of the prediction neural network model, and the real-time collected weld width and residual height data are intelligently matched with the basic data source. When the match is successful, the predicted value of the optimal penetration detection control parameter is directly obtained; when the match fails, the real-time acquired weld width and residual height data are used as new input parameters, and the prediction neural network model is trained again until convergence, and the re-acquired optimal penetration detection control parameter predicted value is used for penetration detection. Based on this training process, on the basis of the constructed neural network model, the model operation frequency is reduced, thereby reducing the software and hardware requirements, while being able to adapt to the model optimization of the input data deviation, ensuring that the real-time weld morphology is always matched to accurately obtain the optimal penetration detection control parameters.

[0017] Furthermore, it also includes S5, the optimization training of the prediction neural network model, which specifically includes:

[0018] S501, establishing an optimization model for a prediction neural network, importing the initial data training set of S2 into the optimization model for model training and optimization, and obtaining standard mapping parameters between weld width, reinforcement height and metallographic inspection images, and standard mapping parameters between penetrant inspection development images and metallographic inspection images;

[0019] S502, using the weld width, reinforcement height, and penetrant detection and development image acquired in real time in S4 as input parameters, importing the optimization model trained in step S501, and obtaining actual mapping parameters of the weld width, reinforcement height and the metallographic detection image, as well as actual mapping parameters of the penetrant detection and development image and the metallographic detection image;

[0020] S503, using the weld width, reinforcement height, penetration test development image, and the standard mapping parameters and the actual mapping parameters obtained in real time by S4 as input parameters, compares the standard mapping parameters obtained by S501 with the actual mapping parameters obtained by S502, and takes the deviation between the two being no more than 5% as the optimization control standard, and takes the pressing height, spraying distance, moving speed and standing time after spraying as output conditions, imports the neural network model constructed in S3 for real-time online optimization training to obtain the optimal penetration test control parameters that meet the current weld width and reinforcement height conditions.

[0021] Based on the above optimization training process, the accuracy of the predicted values ​​of the control parameters of the penetrant inspection is evaluated by changing the mapping relationship between the weld width, reinforcement height and the metallographic inspection image, and between the penetrant inspection development image and the metallographic inspection image. The neural network model is corrected to ensure the accuracy of the predicted values ​​of the control parameters of the penetrant inspection.

[0022] Furthermore, the automated penetration testing process is completed by an automated penetration testing device, which includes at least an automatic image acquisition module, a posture adjustment module, an automatic pressing and spraying module and a time control module, wherein the automatic image acquisition module is used to collect the weld image before spraying to obtain the weld width and residual height data, and to collect the penetration testing development image, the posture adjustment module adjusts the device spatial coordinates, spraying distance and moving speed in real time according to the predicted value of the predictive neural network model, and the automatic pressing and spraying module adjusts the pressing height and pressing time in real time according to the predicted value of the predictive neural network model.

[0023] Furthermore, the process of acquiring and adjusting the spraying distance is as follows: acquiring the spatial model of the welding workpiece through the automatic image acquisition module, and acquiring the weld foot position, weld width and excess height information on the spatial model before detection, simulating and generating the plane where the weld foot position is located and determining the center vertical line of the plane; adjusting the current nozzle of the reagent to be sprayed to be located at the center vertical line through the posture adjustment module, and then adjusting the distance between the current nozzle of the reagent to be sprayed and the plane where the weld foot position is located along the center vertical line according to the design value or the predicted value of step S4.

[0024] The present invention also provides a computer-readable storage medium storing computer-executable instructions, wherein the computer-executable instructions are used to execute any of the aforementioned neural network training methods for predicting penetration detection control parameters.

[0025] The present invention also proposes an electronic device, including a processor and a memory, wherein the memory is used to store computer programs; the processor is used to implement any of the aforementioned neural network training methods for predicting penetration detection control parameters when executing the program stored in the memory.

[0026] The technical solution of the present invention is proposed based on the fact that the welding workpiece material, plate thickness, and welding joint form will lead to differences in weld width and reinforcement, where reinforcement refers to the height difference between the highest point of the central raised part and the lowest point of the weld after the weld is formed. The size of the reinforcement directly affects the amount of reagent used and the area of ​​action of the reagent and the weld. The difference in weld width and reinforcement means that there are differences in the welding melting area, which means that there are also differences in the locations where weld defects may occur. When this difference is manifested in penetrant testing, slight fluctuations in the spraying position, spraying height, and reagent dosage may cause the penetrant testing reagent to be unable to achieve comprehensive and uniform coverage of the weld surface. In particular, when performing penetrant testing on large welded workpieces, due to the large welds and diverse shapes of the welding areas, the weld widths and reinforcements in different areas of the same weld are not the same, which makes it difficult to accurately control the parameters of the penetrant testing process and cannot effectively guarantee the detection rate of weld defects.

[0027] Based on the above problems, the beneficial effects of the present invention are:

[0028] (1) By adopting the neural network training method, with the metallographic detection results of weld defects and the defect detection of the penetrant detection image reaching 99% as the premise, the mapping database between the basic parameters (material, plate thickness, welding joint form), weld width, reinforcement height, and penetrant detection control parameters (pressing height, spraying distance, moving speed and standing time after spraying, etc.) is used as the training sample to predict the penetrant detection control parameters that fit the real-time weld width and reinforcement height. Based on this, the values ​​of each control parameter are adjusted and balanced in real time and finely, thereby achieving uniform and comprehensive coverage of the weld by each detection reagent, obtaining a penetrant detection development image containing complete weld defects, significantly improving the detection rate of weld defects, and providing reliable guarantee for the integrity and accuracy of the detection results.

[0029] (2) Based on the same training data set as the predictive neural network model, another neural network model is used to obtain the standard mapping parameters and actual mapping parameters of the metallographic inspection image and the weld width and residual height, and the metallographic inspection image and the penetrant inspection development image, and the two and their comparison results are introduced into the predictive neural network model for model correction, thereby ensuring the accuracy of the predicted values ​​of the penetrant inspection control parameters. BRIEF DESCRIPTION OF THE DRAWINGS

[0030] Figure 1 A flowchart of a neural network training for predicting control parameters of penetration detection according to the present invention; DETAILED DESCRIPTION

[0031] The present invention will be further described in detail below with reference to the examples, but the embodiments of the present invention are not limited thereto. Example

[0032] The present invention discloses a neural network training method for predicting control parameters of penetrant testing. The method uses a neural network model to obtain predicted values ​​of penetrant testing control parameters based on real-time weld morphology, which are used to guide the real-time adjustment of various control parameters in the automated penetrant testing process and obtain a penetrant testing development image containing complete weld defects. The method comprises the following steps:

[0033] S1, obtain test data:

[0034] Based on the welding workpiece material, plate thickness, and welding joint form (including conventional welding joint forms such as butt joints, T-joints, fillet joints, and lap joints) as basic parameters, the workpiece containing the corresponding weld type is obtained, and the weld shape data including weld width and reinforcement is collected;

[0035] The test was designed and carried out using the reagent bottle nozzle pressing height, spraying distance, moving speed and the rest time after reagent spraying as the control parameters for penetrant testing. The weld defect information of the penetrant testing development image was collected.

[0036] Perform metallographic analysis on the weld mark position after penetrant testing to obtain near-surface weld defect information from the metallographic test image, compare it with the penetrant test development image at the corresponding position, and calculate the defect detection rate;

[0037] S2, constructing a training data set: selecting the basic parameters, penetrant inspection control parameters, weld morphology data, penetrant inspection development images, and metallographic inspection images corresponding to a defect detection rate of 99% obtained in S1 as the training data set;

[0038] S3, establishing and training a prediction neural network model: constructing an initial neural network model, based on the initial training data set in step S2, importing the neural network model, using weld width, reinforcement, welding workpiece material, plate thickness, welding joint form, penetrant detection development image, and metallographic detection image as input parameters, and using reagent bottle nozzle pressing height, spraying distance, moving speed, and reagent standing time after spraying as output parameters, training and optimizing the initial neural network model until convergence, completing the construction of a prediction neural network model for penetrant detection control parameters;

[0039] Preferably, when constructing and training the predictive neural network model, the initial training data set of step S2 is first divided into different training data subsets according to the welding workpiece material, and each training data subset executes step S3 to train and optimize the initial neural network model, obtains the input-output mapping model and data subset corresponding to different welding workpiece materials, and reorganizes the input-output mapping data subsets into a training data set, re-imports the neural network model optimized by the training data subset of any material, executes step S3 and finally constructs a comprehensive predictive neural network model for predicting penetration detection control parameters. The predictive neural network model constructed based on this training process can not only accurately provide the best predicted value of penetration detection control parameters for welds formed on the basis of a certain material, but also provide the best predicted value of penetration detection control parameters for welds formed on the basis of any material, thereby expanding the scope of application of the predictive neural network model.

[0040] S4, obtain the predicted values ​​of the control parameters of the penetrant inspection: Based on the predictive neural network model constructed in S3, manually input the welding workpiece material, plate thickness and welding joint form information, and use the real-time collected weld width and residual height data as input to obtain the predicted values ​​of the current pressing height, spraying distance, moving speed and standing time after spraying. These are used as the optimal penetrant inspection control parameters to guide the implementation of the penetrant inspection, and finally obtain the penetrant inspection development image containing complete weld defects.

[0041] Preferably, when obtaining the predicted value of the penetration detection control parameter, the following process is also included: according to the manually input welding workpiece material and plate thickness data, the corresponding input-output mapping model and data subset are automatically retrieved as the basic data source of the prediction neural network model, and the real-time collected weld width and residual height data are intelligently matched with the basic data source. When the match is successful, the predicted value of the optimal penetration detection control parameter is directly obtained; when the match fails, the real-time acquired weld width and residual height data are used as new input parameters, and the prediction neural network model is trained again until convergence, and the re-acquired optimal penetration detection control parameter predicted value is used for penetration detection. Based on this training process, on the basis of the constructed neural network model, the model operation frequency is reduced, thereby reducing the software and hardware requirements, while being able to adapt to the model optimization of the input data deviation, ensuring that the real-time weld morphology is always matched to accurately obtain the optimal penetration detection control parameter.

[0042] Furthermore, in the aforementioned steps, the weld morphology data and the penetration detection control parameters in the automated penetration detection process are collected by an automated penetration detection device, and the automated penetration detection device includes at least an automatic image acquisition module, a posture adjustment module, an automatic pressing and spraying module and a time control module, wherein the automatic image acquisition module is used to collect the weld image before spraying to obtain weld morphology data such as weld width and excess height, and to collect all penetration detection development images after the actual spraying is completed and the standing still is completed. The posture adjustment module adjusts the spatial position coordinates of the entire device (including the reagent nozzle), the spraying distance (including the spraying height and the spraying spatial position coordinates, used to accurately control the spraying surface coverage position and range) and the moving speed (i.e., the moving speed along the length direction of the weld during reagent spraying) in real time according to the predicted value of the predictive neural network model. The automatic pressing and spraying module adjusts the pressing height (a characterization value of the pressing pressure) and the pressing time in real time according to the predicted value of the predictive neural network model, thereby accurately controlling the reagent spraying amount. The time control module is used to automatically control the length of the standing time after each reagent is sprayed.

[0043] Furthermore, in the aforementioned steps, the process of acquiring and adjusting the spraying distance is as follows: acquiring the spatial model of the welding workpiece through the automatic image acquisition module, and acquiring the weld foot position, weld width and excess height information on the spatial model before detection, simulating and generating the plane where the weld foot position is located and determining the center vertical line of the plane; adjusting the current nozzle of the reagent to be sprayed to be located at the center vertical line through the posture adjustment module, and then adjusting the distance between the current nozzle of the reagent to be sprayed and the plane where the weld foot position is located along the center vertical line according to the design value or the predicted value of step S4, so that the spraying surface can accurately cover the entire surface of the weld, ensuring uniform spraying without omissions.

[0044] The advantages of this embodiment are:

[0045] By adopting the neural network training method, the penetration testing control parameters that fit the actual weld state are accurately obtained, and based on this, the values ​​of each control parameter are adjusted and balanced in real time and finely, thereby achieving uniform and comprehensive coverage of the weld by each detection reagent. Finally, a penetration testing development image containing complete weld defects is obtained, which significantly improves the detection rate of weld defects and provides reliable guarantees for the integrity and accuracy of the detection results in the next link. Example

[0046] This embodiment discloses a neural network training method for predicting penetration detection control parameters, which differs from the first embodiment in that:

[0047] It also includes S5, the optimization training of the prediction neural network model, which specifically includes:

[0048] S501, establishing an optimization model for a prediction neural network, importing the initial data training set of S2 into the optimization model for model training and optimization, and obtaining standard mapping parameters between weld width, reinforcement height and metallographic inspection images, and standard mapping parameters between penetrant inspection development images and metallographic inspection images;

[0049] S502, using the weld width, reinforcement height, and penetrant detection and development image acquired in real time in S4 as input parameters, importing the optimization model trained in step S501, and obtaining actual mapping parameters of the weld width, reinforcement height and the metallographic detection image, as well as actual mapping parameters of the penetrant detection and development image and the metallographic detection image;

[0050] S503, using the weld width, reinforcement height, penetration test development image, and the standard mapping parameters and the actual mapping parameters obtained in real time by S4 as input parameters, compares the standard mapping parameters obtained by S501 with the actual mapping parameters obtained by S502, and takes the deviation between the two being no more than 5% as the optimization control standard, and takes the pressing height, spraying distance, moving speed and standing time after spraying as output conditions, imports the neural network model constructed in S3 for real-time online optimization training to obtain the optimal penetration test control parameters that meet the current weld width and reinforcement height conditions.

[0051] The advantages of this embodiment are: by changing the mapping relationship between weld width, reinforcement height and metallographic detection image, and between penetration detection development image and metallographic detection image, the accuracy of the prediction value of penetration detection control parameter is evaluated, and the neural network model is corrected to eliminate the cumulative error based only on the initial training data set, thereby ensuring the accuracy of the prediction value of penetration detection control parameter and guaranteeing the reliability of the detection results. Example

[0052] This embodiment discloses a computer-readable storage medium storing computer-executable instructions, wherein the computer-executable instructions are used to execute the neural network training method for predicting penetration detection control parameters as described in the first and second embodiments. Example

[0053] This embodiment discloses an electronic device, including a processor and a memory, wherein the memory is used to store computer programs; the processor is used to implement the neural network training method for predicting penetration detection control parameters as described in Examples 1 and 2 when executing the program stored in the memory.

[0054] The above description is merely a preferred embodiment of the present invention and does not limit the technical scope of the present invention. Therefore, any minor modifications, equivalent changes and modifications made to the above embodiments based on the technical essence of the present invention still fall within the scope of protection of the present invention.

Claims

1. A neural network training method for predicting control parameters of penetration detection, characterized by: A neural network model is used to obtain the predicted values ​​of the penetrant inspection control parameters based on the real-time weld morphology. This is used to guide the real-time adjustment of the control parameters in the automated penetrant inspection process and obtain a penetrant inspection development image containing complete weld defects. The process includes the following steps: S1, obtain test data: Obtain the corresponding type of weld based on the welding workpiece material, plate thickness, and weld joint form as basic parameters, and collect weld morphology data including weld width and reinforcement; The penetrant test is designed and executed using the reagent bottle nozzle pressing height, spraying distance, moving speed, and the rest time after reagent spraying as the penetrant test control parameters, and weld defect information from the penetrant test development image is collected; Perform metallographic analysis on the welds after penetrant testing to obtain near-surface weld defect information from the metallographic testing image, compare it with the penetrant testing developed image, and calculate the defect detection rate; S2, constructing a training data set: selecting the basic parameters, penetrant inspection control parameters, weld morphology data, penetrant inspection development images, and metallographic inspection images corresponding to a defect detection rate of 99% obtained in S1 as the training data set; S3, establishing and training a prediction neural network model: constructing an initial neural network model, based on the initial training data set in step S2, importing the neural network model, using weld width, reinforcement, welding workpiece material, plate thickness, welding joint form, penetrant detection development image, and metallographic detection image as input parameters, and using reagent bottle nozzle pressing height, spraying distance, moving speed, and reagent standing time after spraying as output parameters, training and optimizing the initial neural network model until convergence, completing the construction of a prediction neural network model for penetrant detection control parameters; S4, obtain the predicted values ​​of the control parameters of the penetrant inspection: Based on the predictive neural network model constructed in S3, manually input the welding workpiece material, plate thickness and welding joint form information, and use the real-time collected weld width and residual height data as input to obtain the predicted values ​​of the current pressing height, spraying distance, moving speed and standing time after spraying. These are used as the optimal penetrant inspection control parameters to guide the implementation of the penetrant inspection, and finally obtain the penetrant inspection development image containing complete weld defects.

2. A neural network training method for predicting penetration detection control parameters according to claim 1, characterized in that: The initial training data set of step S2 is divided into different training data subsets according to the welding workpiece material. Step S3 is executed for each training data subset to train and optimize the initial neural network model, and the input-output mapping models and data subsets corresponding to different welding workpiece materials are obtained. The input-output mapping data subsets are reorganized into a training data set, and the neural network model optimized by the training data subset of any material is re-imported. Step S3 is executed and finally a comprehensive prediction neural network model for penetration detection control parameter prediction is constructed.

3. A neural network training method for predicting penetration detection control parameters according to claim 2, characterized in that: When obtaining the predicted value of the penetration detection control parameter according to step S4, the following process is also included: according to the manually input welding workpiece material and plate thickness data, the corresponding input-output mapping model and data subset are automatically retrieved as the basic data source of the prediction neural network model, and the real-time collected weld width and residual height data are intelligently matched with the basic data source. When the match is successful, the predicted value of the optimal penetration detection control parameter is directly obtained; when the match fails, the real-time acquired weld width and residual height data are used as new input parameters, and the prediction neural network model is trained again until convergence, and the re-acquired optimal penetration detection control parameter predicted value is used for penetration detection.

4. A neural network training method for penetration detection control parameter prediction according to claim 1, characterized in that: It also includes S5, the optimization training of the prediction neural network model, which specifically includes: S501, establishing an optimization model for a prediction neural network, importing the initial data training set of S2 into the optimization model for model training and optimization, and obtaining standard mapping parameters between weld width, reinforcement height and metallographic inspection images, and standard mapping parameters between penetrant inspection development images and metallographic inspection images; S502, using the weld width, reinforcement height, and penetrant detection and development image acquired in real time in S4 as input parameters, importing the optimization model trained in step S501, and obtaining actual mapping parameters of the weld width, reinforcement height and the metallographic detection image, as well as actual mapping parameters of the penetrant detection and development image and the metallographic detection image; S503, using the weld width, reinforcement height, penetration test development image, and the standard mapping parameters and the actual mapping parameters obtained in real time by S4 as input parameters, compares the standard mapping parameters obtained by S501 with the actual mapping parameters obtained by S502, and takes the deviation between the two being no more than 5% as the optimization control standard, and takes the pressing height, spraying distance, moving speed and standing time after spraying as output conditions, imports the neural network model constructed in S3 for real-time online optimization training to obtain the optimal penetration test control parameters that meet the current weld width and reinforcement height conditions.

5. The neural network training method for predicting penetration detection control parameters according to claim 1, characterized in that: The automated penetration testing process is completed by an automated penetration testing device, which includes at least an automatic image acquisition module, a posture adjustment module, an automatic pressing and spraying module and a time control module, wherein the automatic image acquisition module is used to collect the weld image before spraying to obtain the weld width and residual height data, and to collect the penetration testing development image, the posture adjustment module adjusts the device spatial coordinates, spraying distance and moving speed in real time according to the predicted value of the predictive neural network model, and the automatic pressing and spraying module adjusts the pressing height and pressing time in real time according to the predicted value of the predictive neural network model.

6. A neural network training method for penetration detection control parameter prediction according to claim 5, characterized in that: The acquisition and adjustment process of the spray distance is as follows: the spatial model of the welded workpiece is acquired by the automatic image acquisition module, and the weld leg position, weld width and reinforcement information are acquired on the spatial model before detection, the plane where the weld leg position is located is simulated and the center perpendicular line of the plane is determined; After adjusting the current reagent nozzle to be sprayed to be located at the central vertical line through the posture adjustment module, the distance between the current reagent nozzle to be sprayed and the plane where the weld foot position is located is adjusted along the central vertical line according to the design value or the predicted value of step S4.

7. A computer-readable storage medium, characterized in that Computer-executable instructions are stored, and the computer-executable instructions are used to execute the neural network training method for penetration detection control parameter prediction described in any one of claims 1-6.

8. An electronic device, characterized in that: It comprises a processor and a memory, wherein the memory is used to store computer programs; the processor is used to implement the neural network training method for penetration detection control parameter prediction described in any one of claims 1-6 when executing the program stored in the memory.

Images