Magnetic defect detector with intelligent defect identification function

Abstract

The invention relates to the technical field of magnetic defect detectors, in particular to a magnetic defect detector with an intelligent defect recognition function, which comprises the following steps: acquiring different-scale output characteristics of a to-be-detected piece; constructing a feature model based on the global overall features and the local crack detail features; designing a convolutional neural network based on the feature model, performing feature extraction on the feature model, sampling the extracted features, and pooling sampling information according to different scales of the detection model through a spatial pooling pyramid to perform feature extraction and reaggregation; based on the aggregation information, calling a database to carry out state judgment on the to-be-detected part, and carrying out frame selection and alarm on the identified crack position on the to-be-detected part; and constructing a learning model based on the state of the to-be-detected part and the feature model, and importing the learning model into a database to construct a rapid judgment model. The invention aims to quickly and stably process data and improve the detection precision and robustness of cracks on the to-be-detected piece.

Description

Technical Field

[0001] The invention relates to the technical field of magnetic particle flaw detectors, and in particular to a magnetic particle flaw detector with a defect intelligent recognition function. Background Art

[0002] Magnetic particle flaw detector is a device used to detect surface and near-surface cracks and defects of metal workpieces. It is based on the principle of magnetic particle method and uses magnetic particle detection technology to generate a magnetic field on the metal surface and apply magnetic particles to find and display defects such as cracks, inclusions, pores, etc. on the object being tested, thereby evaluating and testing the quality of metal workpieces.

[0003] In related technologies, data processing is achieved through image processing technology, and simple linear operations such as summation and concatenation are used in data processing and feature fusion problems at different scales. This feature fusion method between different scales is too simple and rough, and cannot effectively bridge the features of feature layers at different scales. The inability to integrate information between feature layers at different scales will lead to poor accuracy and robustness of subsequent target detection. Summary of the invention

[0004] The main purpose of the present invention is to provide a magnetic particle flaw detector with a defect intelligent recognition function, aiming to process data quickly and stably and improve the detection accuracy and robustness of cracks on a workpiece to be inspected.

[0005] To achieve the above object, the present invention proposes a magnetic particle flaw detector with intelligent defect recognition function, comprising the following steps:

[0006] S100, detecting the test piece with magnetic powder and detection magnetic field through an imager, and obtaining output features of the test piece at different scales, including global overall features and local crack detail features;

[0007] S200, constructing a feature model based on the global overall feature and the local crack detail feature, and backing up the feature model by the database;

[0008] S300, designs a convolutional neural network based on the feature model, extracts features from the feature model, samples the extracted features, and pools the sampled information at different scales of the detection model through a spatial pooling pyramid to extract and aggregate features;

[0009] S400, based on the aggregated information and calling the database, the status of the part to be inspected is determined, and the status of the part to be inspected is displayed by setting an imaging module, and the crack position on the identified part to be inspected is framed and an alarm is issued;

[0010] S500, building a learning model based on the state of the part to be detected and the feature model, and importing the learning model into a database to build a fast judgment model.

[0011] In an embodiment of the present application, after step S300, the method further includes:

[0012] S310, designing a function module based on the feature model, and performing preliminary feature extraction on different scales through the function module;

[0013] S320, providing an advanced analysis module based on the feature model, and analyzing the extracted preliminary features through the advanced analysis module to obtain a key information set;

[0014] S330, sampling is performed through maximum pooling based on the key information set, and the sampled information is further analyzed and aggregated using a spatial pooling pyramid to obtain aggregated information.

[0015] In an embodiment of the present application, after step S310, the following steps are further included:

[0016] S311, setting a convolution layer based on the feature model and configuring a filter to extract and detect different feature information at different positions from an input image in the feature model;

[0017] S312, setting a batch normalization layer based on the feature information to normalize the data input of different layers by accelerating the convergence process of the neural network;

[0018] S313, setting the Silu activation function based on the data input, and learning complex feature representations by calling the network for activation feature analysis.

[0019] In one embodiment of the present application, after step S320, the following steps are further included:

[0020] S321, a depth expansion module is provided based on the extracted preliminary features, and the gradient vanishing or exploding problem in the deep neural network is solved by cross-layer connection of the depth expansion module;

[0021] S322, setting a convolution layer based on the depth expansion module, learning deep analysis through the convolution layer, and performing deep analysis on the preliminary features.

[0022] In an embodiment of the present application, after step S400, the method further includes:

[0023] S410, designing a neck structure based on the aggregated information, and collecting more comprehensive status information of the to-be-detected component through the neck structure;

[0024] S420, an attention mechanism is designed based on the neck structure to guide the global information and local information of features at different levels to complement each other.

[0025] In one embodiment of the present application, after step S420, the following steps are further included:

[0026] S421, based on the attention mechanism, performs data processing on the global information from the channel level. The processing method is

[0027]

[0028] Among them, X represents the feature map of the input attention, Mc() represents the channel-wise operation on all feature maps, and H and W represent the height and width of each feature map.

[0029] In one embodiment of the present application, after step S421, the following steps are further included:

[0030] S422, the feature map is focused and weighted from the spatial level, so as to generate a mask through convolution to enhance the local information of the feature map from the spatial level, which can be expressed as

[0031]

[0032] Among them, Ms() represents the spatial operation of all feature maps, C represents the number of channels of the feature map, and δ represents the feature enhancement operation of each channel feature map.

[0033] In one embodiment of the present application, after step S422, the following steps are further included:

[0034] S423, the feature maps of different scales are fused, the deep feature map and the shallow feature map obtained by sampling are jointly processed, and finally the features are fused at different scales to obtain the state features of the part to be detected, which can be expressed as

[0035]

[0036]

[0037] Where M is the final enhanced feature of the input feature map, represents element-by-element multiplication, N is the newly generated fusion feature, F up( X l+1 ) represents the underlying features from upsampling, X l represents shallow features, and [] represents the merging operation.

[0038] In one embodiment of the present application, after step S423, the following steps are further included:

[0039] S424, based on the newly generated fusion features, the number of feature channels is adjusted through RepConv structure re-parameterization, and structure prediction is performed on feature maps of different scales to output the corresponding target position and category information.

[0040] In one embodiment of the present application, a magnetic particle flaw detection system is applied to a magnetic particle flaw detector with a defect intelligent recognition function, and the magnetic particle flaw detection system includes:

[0041] A data collection module, wherein the data collection module is used to connect to the magnetic particle flaw detector and collect the global overall characteristics and local crack detail characteristics of the to-be-detected part at different scales;

[0042] A data sorting module, which is used to receive global overall features and local crack detail features, build a feature model, and temporarily store them in a database;

[0043] A feature extraction module, which extracts state information and feature keywords of different scales according to a feature model, and obtains a state diagram of the part to be detected through a function;

[0044] The display adjustment module displays the status of the part to be inspected based on the status chart of the part to be inspected, issues an alarm for existing cracks, and updates the database in combination with the feature model.

[0045] The technical solution of the present invention adopts the method of pre-cleaning the surface of the to-be-detected piece before detecting whether there is a wound on the to-be-detected piece to ensure that there is no oil or impurities, placing a magnetic particle flaw detector, adjusting a remote camera, and setting the magnetic field strength and the dilution ratio of the magnetic powder to ensure that it is suitable for specific detection needs. The magnetic powder is applied to the potential defect in a powder or liquid state, and a magnetic particle flaw detector is used to apply a magnetic field, and the state data of the to-be-detected piece is detected and collected in combination with a remote camera to judge the injury. Specifically, it is first necessary to obtain different scale output features of the potential defects of the to-be-detected piece through instruments such as an imager, a detector, and a remote camera, including global overall features and local crack detail features. These features will be received by the flaw detection management software prefabricated in the magnetic particle flaw detector, and a feature model is constructed based on these data. The feature model is conducive to the subsequent calculation of the state, and the feature model is backed up by a database. After the backup is completed in combination with the final calculation result, it can be conducive to the system to perform deep learning on the flaw detection process and method, and it is conducive to improving the accuracy and efficiency of the subsequent judgment of the injury of the to-be-detected piece. A convolutional neural network can be set up correspondingly through the feature model, and the network can efficiently and accurately collect and judge the to-be-detected piece. The injury condition can first be extracted by using the activation function through the information points of the feature model. The system is also learning advanced semantics through networking, which is convenient for screening the initially extracted features, obtaining key information, and sampling. The sampling information is combined with the feature model to be pooled from different scales to extract features and then aggregated through the sampling space pooling pyramid method to improve the robustness and accuracy of the algorithm. The above aggregated information is calculated and judged, and the instances that have been judged in the historical database can be called for auxiliary judgment, and then the crack category information and position information of the crack of the to-be-detected piece in the image collected by the magnetic particle flaw detector can be judged to realize the automatic detection of crack identification and flaw detection. The imaging module configured by the magnetic particle flaw detector, such as the display screen, marks the wound position of the to-be-detected piece in the form of an image and alarms, so that the user can know the detection result. When the detection of the state of the to-be-detected piece is completed, the database will construct a learning model with the corresponding result together with the feature model, and optimize the fast judgment model together with the previous data in the database, so as to facilitate the subsequent call when judging the state of the to-be-detected piece, so as to achieve the purpose of quickly and accurately judging the state of the current to-be-detected piece, and improve the detection accuracy and robustness. BRIEF DESCRIPTION OF THE DRAWINGS

[0046] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on the structures shown in these drawings without paying creative work.

[0047] Figure 1This is a flow chart of the magnetic particle flaw detector with intelligent defect recognition function of the present invention;

[0048] Figure 2 It is a data analysis flow chart of the magnetic particle flaw detector with intelligent defect recognition function of the present invention;

[0049] Figure 3 It is a preliminary feature extraction flow chart of the magnetic particle flaw detector with defect intelligent recognition function of the present invention;

[0050] Figure 4 It is a flow chart of the depth analysis of the magnetic particle flaw detector with intelligent defect recognition function of the present invention;

[0051] Figure 5 It is a crack identification flow chart of the magnetic particle flaw detector with intelligent defect identification function of the present invention;

[0052] Figure 6 It is a structural schematic diagram of the magnetic particle flaw detection system of the present invention.

[0053] The realization of the purpose, functional features and advantages of the present invention will be further explained in conjunction with embodiments and with reference to the accompanying drawings. DETAILED DESCRIPTION

[0054] In order to make the purpose, technical solution and advantages of the present application more clearly understood, the present application is further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present application and are not used to limit the present application.

[0055] To solve the problems in the background technology, such as Figures 1 to 5 As shown, the present invention proposes a magnetic particle flaw detector with a defect intelligent identification function, comprising the following steps:

[0056] S100, detecting the test piece with magnetic powder and detection magnetic field through an imager, and obtaining output features of the test piece at different scales, including global overall features and local crack detail features;

[0057] S200, constructing a feature model based on the global overall feature and the local crack detail feature, and backing up the feature model by the database;

[0058] S300, designs a convolutional neural network based on the feature model, extracts features from the feature model, samples the extracted features, and pools the sampled information at different scales of the detection model through a spatial pooling pyramid to extract and aggregate features;

[0059] S400, based on the aggregated information and calling the database, the status of the part to be inspected is determined, and the status of the part to be inspected is displayed by setting an imaging module, and the crack position on the identified part to be inspected is framed and an alarm is issued;

[0060] S500, building a learning model based on the state of the part to be detected and the feature model, and importing the learning model into a database to build a fast judgment model.

[0061] Understandably, before detecting whether there is a wound on the part to be inspected, the surface of the part to be inspected is cleaned in advance to ensure that there is no oil or impurities, the magnetic particle flaw detector is placed, and the remote camera is adjusted. At the same time, the magnetic field strength and the dilution ratio of the magnetic powder are set to ensure that it is suitable for specific inspection needs, and the magnetic powder is applied to the potential defects in a powder or liquid state, and the magnetic particle flaw detector is used to apply a magnetic field, and the remote camera is combined to detect and collect the state data of the part to be inspected for injury judgment. Specifically, it is first necessary to obtain different scale output features of the potential defects of the part to be inspected through instruments such as imagers, detectors, and remote cameras, including global overall features and local crack detail features. These features will be received by the flaw detection management software prefabricated in the magnetic particle flaw detector, and a feature model is constructed based on these data. The feature model is conducive to the subsequent calculation of the state, and the feature model is backed up by the database. After the backup is completed in combination with the final calculation result, it can be conducive to the system to conduct deep learning of the flaw detection process and method, and to improve the accuracy and efficiency of the subsequent judgment of the injury of the part to be inspected. Through the feature model, a convolutional neural network can be set up accordingly, which can efficiently and accurately collect and judge the injury of the part to be inspected. First, the activation function can be used to perform preliminary feature extraction through the information points of the feature model. The system is also learning advanced semantics through networking, which is convenient for screening the initially extracted features, obtaining key information, and sampling. Through the sampling space pooling pyramid method, the sampling information is combined with the feature model to pool from different scales to extract features and then aggregate them to improve the robustness and accuracy of the algorithm. The above aggregated information is calculated and judged, and the instances that have been judged in the historical database can be called for auxiliary judgment, and then the crack category information and position information of the crack of the to-be-detected piece in the image collected by the magnetic particle flaw detector are judged, so as to realize the automatic detection of crack identification and flaw detection, and the imaging module configured by the magnetic particle flaw detector, such as the display screen, marks the wound position of the to-be-detected piece in the form of an image and alarms, so that the user can know the detection results. When the state detection of the to-be-detected piece is completed, the database will construct a learning model with the corresponding results together with the feature model, and optimize the fast judgment model together with the previous data in the database, so as to facilitate the subsequent call when judging the state of the to-be-detected piece, so as to achieve the purpose of quickly and accurately judging the state of the current to-be-detected piece, and improve the detection accuracy and robustness.

[0062] S310, designing a function module based on the feature model, and performing preliminary feature extraction on different scales through the function module.

[0063] It can be understood that by designing the function module, it can output the corresponding position information after receiving the target information. Through the setting of the function module, the possible hidden danger information can be preliminarily extracted and processed from the features of different scales, which is convenient for the subsequent higher-precision calculation of the relevant information and the status information of the test piece can be obtained.

[0064] S320, an advanced analysis module is provided based on the feature model, and the extracted preliminary features are analyzed by the advanced analysis module to obtain a key information set.

[0065] It can be understood that in order to better process the features collected initially, an advanced analysis module is also set relative to the feature model. This module can use a convolution block with residuals to continuously learn abstract semantic information to improve the system's ability to analyze information, obtain key information sets, and improve flaw detection accuracy.

[0066] S330, sampling is performed through maximum pooling based on the key information set, and the sampled information is further analyzed and aggregated using a spatial pooling pyramid to obtain aggregated information.

[0067] It can be understood that after obtaining the key information set, the maximum pooling is used for sampling operations, which can reduce the spatial dimension of the feature model, retain the most significant features by taking the maximum value of each area, and improve the computational efficiency by reducing the amount of calculation. The sampling information is further processed by the spatial pooling pyramid. The pyramid-shaped pooling operation allows multi-scale features to be extracted from different areas of the input feature model to deal with defects or features of different sizes and proportions, and integrate the multi-scale feature models extracted from the backbone network through feature model aggregation, so as to improve the robustness of the algorithm and the detection accuracy.

[0068] S311, setting a convolution layer based on the feature model and configuring a filter to extract and detect different feature information at different positions from the input image in the feature model.

[0069] It can be understood that in order to improve the ability to collect and process information on risky parts of the inspected parts, a convolution-batch normalization layer-Silu activation function module is relatively set up to perform preliminary feature sampling, wherein the convolution layer is used to extract features from the image input into the system by the remote camera, and a filter is set so that the collected feature information can be decomposed into different positions and different features to avoid mutual interference of information and improve the accuracy of the data.

[0070] S312, setting a batch normalization layer based on feature information to standardize data inputs of different layers by accelerating the convergence process of the neural network.

[0071] It can be understood that in order to standardize the data and avoid data confusion and distortion, a batch normalization layer is set relative to the feature information, so that the feature collection will input the data layer by layer relative to the different layer states of the inspection parts to avoid quantitative confusion and incomputation.

[0072] S313, setting the Silu activation function based on the data input, and learning complex feature representations by calling the network for activation feature analysis.

[0073] It can be understood that in order to activate the entire module and facilitate the collection of data required for calculating cracks, a Silu activation function is set, which can learn auxiliary feature representations. The activation function is a smooth and nonlinear activation function, which can adaptively scale the input value and improve the accuracy of the input data, thereby improving the robustness and accuracy of the algorithm.

[0074] S321, a depth expansion module is set based on the extracted preliminary features, and the gradient disappearance or explosion problem in the deep neural network is solved through cross-layer connection of the depth expansion module.

[0075] Understandably, in order to make the collected preliminary feature data more stable and to extract more critical information, convolutional blocks with residuals are used to learn abstract semantic information. Residual connections can solve the gradient vanishing or exploding problem in deep neural networks through cross-layer connections, which helps to expand the depth of the network.

[0076] S322, setting a convolution layer based on the depth expansion module, learning deep analysis through the convolution layer, and performing deep analysis on the preliminary features.

[0077] Understandably, the convolutional blocks with residuals can be used to perform in-depth analysis on the preliminary feature data, optimize the data therein, and improve subsequent computational efficiency.

[0078] S410, a neck structure is designed based on the aggregated information, and more comprehensive status information of the to-be-detected component is collected through the neck structure.

[0079] It can be understood that in order to make the calculated data more accurate to determine the wound location of the inspected piece, the neck structure is redesigned so that the system can obtain and utilize more comprehensive semantic and spatial information from the features, which is beneficial to the characteristic expression of cracks and fissures in the images of real industrial flaw detection scenes, so that multi-scale and small target defects can be detected more accurately and comprehensively, achieving the goal of improving the accuracy of toilets.

[0080] S420, an attention mechanism is designed based on the neck structure to guide the global information and local information of features at different levels to complement each other.

[0081] Understandably, in order to comprehensively consider the semantic information and spatial information in global features and local features, an attention mechanism is designed to guide the global information and local information of features at different levels to complement each other, which helps to fuse features at different scales and merge to generate features that are more discriminative than the input features, thereby improving the accuracy of the target detection algorithm.

[0082] S421, based on the attention mechanism, performs data processing on the global information from the channel level. The processing method is Among them, X represents the feature map of the input attention, Mc() represents the channel-wise operation on all feature maps, and H and W represent the height and width of each feature map.

[0083] It can be understood that in order to express the feature map of the input attention more clearly, the feature module is operated in the channel direction so that the different levels of the to-be-detected parts can be expressed more clearly.

[0084] S422, the feature map is focused and weighted from the spatial level, so as to generate a mask through convolution to enhance the local information of the feature map from the spatial level, which can be expressed as Among them, Ms() represents the spatial operation of all feature maps, C represents the number of channels of the feature map, and δ represents the feature enhancement operation of each channel feature map.

[0085] It can be understood that the features are further clustered and weighted at the spatial level to strengthen the local information of the feature model to improve the detection accuracy.

[0086] S423, the feature maps of different scales are fused, the deep feature map and the shallow feature map obtained by sampling are jointly processed, and finally the features are fused at different scales to obtain the state features of the part to be detected, which can be expressed as Where M is the final enhanced feature of the input feature map, represents element-by-element multiplication, N is the newly generated fusion feature, F up (X l+1 ) represents the underlying features from upsampling, X l represents shallow features, and [] represents the merging operation.

[0087] Understandably, the features are finally fused and finally generate features that are more discriminative than the input features, so as to achieve the purpose of improving robustness and making the detection effect better.

[0088] S424, based on the newly generated fusion features, the number of feature channels is adjusted through RepConv structure re-parameterization, and structure prediction is performed on feature maps of different scales to output the corresponding target position and category information.

[0089] It can be understood that after obtaining the final fusion feature expression, the number of feature channels is adjusted through RepConv structure reparameterization for final output, and the crack category information and position information of the crack in the image are determined to realize the automated detection of crack identification and flaw detection.

[0090] Reference Figure 6 As shown, the present invention also proposes a magnetic particle flaw detection system, which includes the above-mentioned magnetic particle flaw detector with defect intelligent recognition function. The specific structure of the magnetic particle flaw detector with defect intelligent recognition function refers to the above-mentioned embodiment. Since this magnetic particle flaw detection system adopts all the technical solutions of all the above-mentioned embodiments, it at least has all the beneficial effects brought by the technical solutions of the above-mentioned embodiments, which will not be repeated here one by one.

[0091] The magnetic particle flaw detection system proposed by the present invention comprises a data collection module, a data sorting module, a feature extraction module and a display adjustment module. The data collection module is used to be connected to a magnetic particle flaw detector and collect global overall features and local crack detail features of a to-be-detected piece at different scales; the data sorting module is used to receive global overall features and local crack detail features, construct a feature model, and temporarily store them through a database; the feature extraction module extracts state information and feature keywords of different scales according to the feature model, and obtains a state chart of the to-be-detected piece through a function; the display adjustment module displays the state of the to-be-detected piece based on the state chart of the to-be-detected piece, gives an alarm for existing crack conditions, and updates the database in combination with the feature model.

[0092] It can be understood that in this flaw detection system or flaw detection management software, the perception module first detects and collects information, which is also the data collection module. After collecting the information, the data sorting module collects the information and classifies and organizes it according to different layers, different areas, and different types to facilitate subsequent calculations. The feature extraction module will classify and summarize the data according to the feature model, and calculate the status of the parts to be inspected. The display adjustment module displays the test results to the user, and the collected results are sorted together with the feature model, and stored and learned by the database to continuously train the database and improve subsequent detection efficiency.

[0093] The same or similar numbers in the drawings of this embodiment correspond to the same or similar parts; in the description of this application, it should be understood that if the terms "upper", "lower", "left", "right", etc. indicate an orientation or position relationship based on the orientation or position relationship shown in the drawings, it is only for the convenience of describing this application and simplifying the description, rather than indicating or implying that the device or element referred to must have a specific orientation, be constructed and operated in a specific orientation. Therefore, the terms describing the position relationship in the drawings are only used for illustrative purposes and cannot be understood as a limitation on this patent. For ordinary technicians in this field, the specific meanings of the above terms can be understood according to specific circumstances.

[0094] The above are only preferred embodiments of the present application and are not intended to limit the present application. Any modifications, equivalent substitutions and improvements made within the spirit and principles of the present application should be included in the protection scope of the present application.

Claims

1. A magnetic particle flaw detector with intelligent defect recognition function, characterized in that: The following steps are involved: S100, detecting the test piece with magnetic powder and detection magnetic field through an imager, and obtaining output features of the test piece at different scales, including global overall features and local crack detail features; S200, constructing a feature model based on the global overall feature and the local crack detail feature, and backing up the feature model by the database; S300, designs a convolutional neural network based on the feature model, extracts features from the feature model, samples the extracted features, and pools the sampled information at different scales of the detection model through a spatial pooling pyramid to extract and aggregate features; S400, based on the aggregated information and calling the database, the status of the part to be inspected is determined, and the status of the part to be inspected is displayed by setting an imaging module, and the crack position on the identified part to be inspected is framed and an alarm is issued; S500, building a learning model based on the state of the part to be detected and the feature model, and importing the learning model into a database to build a fast judgment model.

2. A magnetic particle flaw detector with intelligent defect recognition function according to claim 1, characterized in that: After step S300, the method further includes: S310, designing a function module based on the feature model, and performing preliminary feature extraction on different scales through the function module; S320, providing an advanced analysis module based on the feature model, and analyzing the extracted preliminary features through the advanced analysis module to obtain a key information set; S330, sampling is performed through maximum pooling based on the key information set, and the sampled information is further analyzed and aggregated using a spatial pooling pyramid to obtain aggregated information.

3. The magnetic particle flaw detector with intelligent defect recognition function according to claim 2 is characterized in that: After step S310, the method further includes: S311, setting a convolution layer based on the feature model and configuring a filter to extract and detect different feature information at different positions from an input image in the feature model; S312, setting a batch normalization layer based on the feature information to normalize the data input of different layers by accelerating the convergence process of the neural network; S313, setting the Silu activation function based on the data input, and learning complex feature representations by calling the network for activation feature analysis.

4. The magnetic particle flaw detector with intelligent defect recognition function according to claim 3 is characterized in that: After step S320, the method further includes: S321, a depth expansion module is provided based on the extracted preliminary features, and the gradient vanishing or exploding problem in the deep neural network is solved by cross-layer connection of the depth expansion module; S322, setting a convolution layer based on the depth expansion module, learning deep analysis through the convolution layer, and performing deep analysis on the preliminary features.

5. The magnetic particle flaw detector with intelligent defect recognition function according to claim 4 is characterized in that: After step S400, the method further includes: S410, designing a neck structure based on the aggregated information, and collecting more comprehensive status information of the to-be-detected component through the neck structure; S420, based on the neck structure, is designed with an attention mechanism to guide the global and local information of features at different levels to complement each other.

6. The magnetic particle flaw detector with intelligent defect recognition function according to claim 5, characterized in that: After step S420, the method further includes: S421, based on the attention mechanism, performs data processing on the global information from the channel level. The processing method is Among them, X represents the feature map of the input attention, Mc() represents the channel-wise operation on all feature maps, and H and W represent the height and width of each feature map.

7. The magnetic particle flaw detector with intelligent defect recognition function according to claim 6 is characterized in that: After step S421, the method further includes: S422, the feature map is focused and weighted from the spatial level, so as to generate a mask through convolution to enhance the local information of the feature map from the spatial level, which can be expressed as Among them, Ms() represents the spatial operation of all feature maps, C represents the number of channels of the feature map, and δ represents the feature enhancement operation of each channel feature map.

8. The magnetic particle flaw detector with intelligent defect recognition function according to claim 7, characterized in that: After step S422, the method further includes: S423, the feature maps of different scales are fused, the deep feature map and the shallow feature map obtained by sampling are jointly processed, and finally the features are fused at different scales to obtain the state features of the part to be detected, which can be expressed as Where M is the final enhanced feature of the input feature map, represents element-by-element multiplication, N is the newly generated fusion feature, F up (X l+1 ) represents the underlying features from upsampling, X l represents shallow features, and [] represents the merging operation.

9. The magnetic particle flaw detector with intelligent defect recognition function according to claim 8, characterized in that: After step S423, the method further includes: S424, based on the newly generated fusion features, the number of feature channels is adjusted through RepConv structure re-parameterization, and structure prediction is performed on feature maps of different scales to output the corresponding target position and category information.

10. A magnetic particle flaw detection system, comprising a magnetic particle flaw detector with a defect intelligent recognition function as claimed in any one of claims 1 to 9, characterized in that: The magnetic particle flaw detection system comprises: A data collection module, wherein the data collection module is used to connect to the magnetic particle flaw detector and collect the global overall characteristics and local crack detail characteristics of the to-be-detected part at different scales; A data sorting module, which is used to receive global overall features and local crack detail features, build a feature model, and temporarily store them in a database; A feature extraction module, which extracts state information and feature keywords of different scales according to a feature model, and obtains a state diagram of the part to be detected through a function; The display adjustment module displays the status of the part to be inspected based on the status chart of the part to be inspected, issues an alarm for existing cracks, and updates the database in combination with the feature model.

Images