A method and system for precision metal part defect detection based on multimodal large model

Abstract

The present invention provides a precision metal part defect detection method and system based on a multimodal large model. The method obtains laser scattering image data of the precision metal part in an electromagnetic shielding environment; performs multi-scale spatial analysis on the laser scattering image data to extract a multimodal image feature set including the grain boundary topology structure of the precision metal part surface and the scattering spot distribution gradient characteristics of the sub-surface; converts the process parameters corresponding to the precision metal part in an existing metal heat treatment process database into a thermal history feature vector, wherein the thermal history feature vector includes a temperature gradient distribution feature and a time series phase change feature; and fuses the thermal history feature vector and the multimodal image feature set through a multimodal large model to obtain a fusion processing result to generate defect detection data. The present invention improves the detection rate and detection efficiency of micron-level defects in complex electromagnetic environments.

Description

Technical Field

[0001] The present invention relates to the field of defect detection technology, and in particular to a method and system for detecting defects in precision metal parts based on a multimodal large model. Background Art

[0002] As the reliability requirements for precision metal parts continue to increase in high-end manufacturing fields such as aerospace and nuclear power equipment, online detection of micron-scale grain boundary cracks and subsurface defects has become a key challenge. When these components are subjected to complex operating conditions (such as high-cycle fatigue), defects often exhibit multi-physics coupling characteristics.

[0003] The current mainstream technical solution uses a defect detection system based on laser speckle imaging and convolutional neural networks. Its technical implementation process includes laser speckle image acquisition, single-modal feature extraction, and defect detection using a binocular vision system. Existing technical solutions have some drawbacks, such as relying solely on visual modality data, resulting in a high rate of missed defect detection; and image acquisition in an open environment, resulting in the loss of data on microcracks (e.g., 5μm). Summary of the Invention

[0004] The present invention provides a precision metal parts defect detection method and system based on a multimodal large model, which is used to solve the problems in the prior art of relying solely on visual modal data, resulting in a high defect missed detection rate and a lack of microcrack data collection.

[0005] In a first aspect, the present invention provides a method for detecting defects in precision metal parts based on a multimodal large model, comprising:

[0006] Acquire laser scattering image data of precision metal parts in an electromagnetic shielding environment;

[0007] Performing multi-scale spatial analysis on the laser scattering image data to extract a multimodal image feature set including the grain boundary topology of the surface of the precision metal part and the scattering spot distribution gradient characteristics of the subsurface;

[0008] Converting the process parameters corresponding to the precision metal part in the existing metal heat treatment process database into a thermal history feature vector, wherein the thermal history feature vector includes a temperature gradient distribution feature and a time series phase change feature;

[0009] fusing the thermal history feature vector and the multimodal image feature set through a multimodal large model to obtain a fusion processing result;

[0010] Based on the fusion processing result, defect detection data is generated, wherein the defect detection data includes three-dimensional position information of grain boundary cracks and prediction data of crack propagation direction.

[0011] Optionally, the thermal history feature vector and the multimodal image feature set are fused using a multimodal large model to obtain a fusion result, including:

[0012] generating a weight control parameter set based on material property parameters of the precision metal part, wherein the weight control parameter set includes a thermal history weight vector and an image weight vector;

[0013] In a thermal history feature encoding channel of a multimodal large model, mapping processing is performed on the thermal history feature vector according to the thermal history weight vector to generate a process feature encoding sequence with temperature gradient correlation characteristics;

[0014] In an image feature encoding channel of a multimodal large model, a sliding window process is performed on the multimodal image feature set according to the image weight vector to generate an image feature encoding sequence having grain boundary topological correlation characteristics;

[0015] Performing weighted processing on the process feature coding sequence and the image feature coding sequence according to the weight control parameter set to generate a fused coding tensor;

[0016] The fused coding tensor is screened to generate an optimized fusion processing result.

[0017] Optionally, weighting the process feature coding sequence and the image feature coding sequence according to the weight control parameter set to generate a fused coding tensor includes:

[0018] Performing element-by-element multiplication operation on each element value of the thermal history weight vector in the weight control parameter set and the corresponding element value of the process feature coding sequence to generate a weighted process feature sequence;

[0019] Performing element-by-element multiplication on each element value of the image weight vector in the weight control parameter set and the corresponding element value of the image feature coding sequence to generate a weighted image feature sequence;

[0020] Determining a splicing ratio coefficient according to material property parameters of the precision metal part, and performing dimension adjustment processing on the weighted process feature sequence and the weighted image feature sequence according to the splicing ratio coefficient to obtain an adjusted weighted process feature sequence and an adjusted weighted image feature sequence;

[0021] The adjusted weighted process feature sequence and the adjusted weighted image feature sequence are interleaved and spliced ​​in element position order to generate initial fusion coding data;

[0022] Performing target redundant dimension labeling processing on the initial fused coded data to obtain labeled fused coded data;

[0023] After removing the labels, all marked target redundant dimensions in the fused coded data are fused to generate a fused coded tensor.

[0024] In a second aspect, the present invention provides a precision metal part defect detection system based on a multimodal large model, comprising:

[0025] An acquisition module is used to acquire laser scattering image data of precision metal parts in an electromagnetic shielding environment;

[0026] An analysis module, configured to perform multi-scale spatial analysis on the laser scattering image data to extract a multimodal image feature set including the grain boundary topology of the surface of the precision metal part and the scattering spot distribution gradient characteristics of the subsurface;

[0027] a conversion module, configured to convert the process parameters corresponding to the precision metal part in an existing metal heat treatment process database into a thermal history feature vector, wherein the thermal history feature vector includes a temperature gradient distribution feature and a time series phase change feature;

[0028] A fusion module, configured to fuse the thermal history feature vector and the multimodal image feature set using a multimodal large model to obtain a fusion result;

[0029] A generation module is used to generate defect detection data based on the fusion processing result, wherein the defect detection data includes three-dimensional position information of grain boundary cracks and prediction data of crack propagation direction.

[0030] In a third aspect, the present invention provides a computing device comprising a processor and a memory, wherein the memory stores a computer program, and the processor is configured to run the computer program to execute a precision metal part defect detection method based on a multimodal large model as described in any one of the first aspects.

[0031] In a fourth aspect, the present invention provides a computer storage medium having computer program instructions stored thereon, wherein the computer program instructions, when executed by a processor, implement a precision metal part defect detection method based on a multimodal large model as described in any one of the first aspects.

[0032] In the present invention, laser scattering image data of a precision metal part is acquired in an electromagnetic shielding environment; multi-scale spatial analysis is performed on the laser scattering image data to extract a multimodal image feature set including the grain boundary topological structure of the surface of the precision metal part and the gradient characteristics of the scattering spot distribution of the subsurface; the process parameters corresponding to the precision metal part in an existing metal heat treatment process database are converted into a thermal history feature vector, and the thermal history feature vector includes temperature gradient distribution characteristics and time series phase change characteristics; the thermal history feature vector and the multimodal image feature set are fused through a multimodal large model to obtain a fusion processing result; based on the fusion processing result, defect detection data is generated, and the defect detection data includes three-dimensional position information of grain boundary cracks and predicted data of crack propagation direction. The technical solution provided by the present invention suppresses high-frequency interference signals in industrial sites through an electromagnetic shielding environment, ensures the stable capture of micron-level defect features, and solves the problem that detection signals in open environments are susceptible to electromagnetic pollution and lead to missed detection; through multi-scale analysis, the cross-dimensional features of surface grain boundary topology (macrostructure) and sub-surface scattering spot gradient (microdistribution) are simultaneously extracted, breaking through the limitations of traditional single-scale feature extraction in characterizing complex defects, and enhancing the joint detection capability of grain boundary cracks and inclusions; through time and space feature vectors, a physical correlation between heat treatment history and defect formation is established, providing process constraint prior knowledge for multimodal fusion; by fusing process history data with real-time detection data, the problem of misjudgment of multi-physical field coupling defects (such as thermal stress cracks) by traditional single-modal models is solved, and the robustness of defect identification under complex working conditions is improved; the composite detection results break through the geometric distortion defects of traditional two-dimensional positioning, and provide high-precision input for life prediction and maintenance decisions.

[0033] These and other aspects of the present invention will become more readily apparent from the following description of the embodiments. BRIEF DESCRIPTION OF THE DRAWINGS

[0034] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following is a brief introduction to the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0035] Figure 1 A flowchart of a method for detecting defects in precision metal parts based on a multimodal large model provided by an embodiment of the present invention;

[0036] Figure 2 A schematic structural diagram of a precision metal parts defect detection system based on a multimodal large model provided by an embodiment of the present invention;

[0037] Figure 3A schematic diagram of the structure of a computing device provided in an embodiment of the present invention. DETAILED DESCRIPTION

[0038] In order to enable those skilled in the art to better understand the solutions of the present invention, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the accompanying drawings in the embodiments of the present invention.

[0039] In some of the processes described in the specification and claims of the present invention and the above-mentioned figures, multiple operations that appear in a specific order are included, but it should be clearly understood that these operations may not be executed in the order in which they appear in this article or may be executed in parallel. The serial numbers of the operations, such as 101, 102, etc., are only used to distinguish between different operations, and the serial numbers themselves do not represent any execution order. In addition, these processes may include more or fewer operations, and these operations may be executed in sequence or in parallel. It should be noted that the descriptions of "first", "second", etc. in this article are used to distinguish different messages, devices, modules, etc., and do not represent the order of precedence, nor do they limit "first" and "second" to be different types.

[0040] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without making any creative efforts shall fall within the scope of protection of the present invention.

[0041] Figure 1 The present invention provides a flowchart of a method for detecting defects in precision metal parts based on a multimodal large model. Figure 1 As shown, the method includes:

[0042] Aiming at the core problems of insufficient micro-defect detection accuracy of precision metal parts in complex electromagnetic interference environments and difficulty in identifying defects coupled with multiple physical fields, this technology breaks through the limitations of traditional single-modal detection solutions. It suppresses high-frequency noise interference by constructing an electromagnetic shielding environment, innovatively integrates multi-scale characteristics of laser scattering with historical data of heat treatment processes, and designs a multi-modal large model to dynamically associate process parameters with real-time detection features. It solves the problems of missed detection of sub-surface defects and distortion in crack propagation direction prediction caused by the lack of thermal-optical cross-modal fusion in existing technologies. At the same time, it realizes precise defect positioning through three-dimensional spatial mapping and physical constraint modeling, overcomes the technical bottleneck of full-dimensional detection of micron-level defects in industrial field environments, and forms a high-precision defect detection closed-loop system driven by process data and multi-modal collaboration. Based on this, the present invention provides a precision metal parts defect detection method based on a multi-modal large model, such as Figure 1 ,include:

[0043] Step 101: Acquire laser scattering image data of a precision metal part in an electromagnetic shielding environment.

[0044] In this step, the electromagnetic shielding environment refers to a closed detection space constructed with conductive shielding structures and absorbing materials. It is used to suppress external electromagnetic interference signals (frequency range 10kHz-6GHz) and ensure the purity of laser scattering signal acquisition. Laser scattering image data refers to the distribution of reflected light spots captured by an optical sensor when a polarized laser beam is illuminated on the surface of a precision metal part. It contains surface microstructure morphology and diffraction characteristics of subsurface defects.

[0045] In this step, a closed detection space is constructed in an electromagnetic shielding environment through a Faraday detection cavity. The cavity adopts a composite structure of multiple layers of conductive shielding layers and absorbing materials to attenuate and suppress electromagnetic interference signals in the frequency range of 10kHz-6GHz in the detection area; in the closed detection space, a laser emitting device is used to project a polarized laser beam with a wavelength of 532nm onto the surface of the precision metal part, so that the polarized laser beam is scattered by the surface of the alloy part and a reflected light spot distribution is generated; the reflected light distribution data in the closed detection space is received in real time by an optical sensing device to generate laser scattering image data to be optimized containing surface microstructure features; based on the electromagnetic interference suppression environment provided by the closed detection space, dark current correction and background noise filtering are performed on the laser scattering image data to be optimized to eliminate signal distortion introduced by external electromagnetic disturbances, and obtain pre-processed laser scattering image data with a high signal-to-noise ratio as the input source for subsequent multi-scale spatial analysis.

[0046] Step 102: Perform multi-scale spatial analysis on the laser scattering image data to extract a multi-modal image feature set including the grain boundary topology structure of the surface of the precision metal part and the scattering spot distribution gradient characteristics of the sub-surface.

[0047] In this step, the grain boundary topology refers to the spatial network morphology formed by the grain boundaries on the surface of the metal part, including the grain boundary strike angle, node density, and connectivity parameters, reflecting the microscopic deformation under mechanical stress. The scattering spot distribution gradient characteristic refers to the spatial intensity change rate of the scattered light spot caused by sub-surface defects, which characterizes the defect depth, size, and degree of stress concentration within the material. The multimodal image feature set refers to a cross-dimensional feature set that integrates the surface grain boundary topology (macroscale) and the sub-surface scattering spot distribution gradient (microscale) to form a multi-physics field representation for defect detection.

[0048] This step performs multi-scale spatial analysis on the pre-processed laser scattering image data, specifically including: extracting grain boundary topological structure features using pixel gradient analysis in the spatial domain to obtain grain boundary strike angle distribution maps and node density distribution maps; performing two-dimensional wavelet decomposition on the laser scattering image data in the frequency domain to extract the energy distribution characteristics of scattering spots at different scales and generate sub-surface scattering spot spectrum feature maps; aligning the grain boundary topological structure features with the sub-surface scattering spot spectrum feature maps in feature space to construct a multimodal image feature set containing surface-subsurface joint features as the feature input for multimodal fusion processing.

[0049] Step 103: converting the process parameters corresponding to the precision metal part in the existing metal heat treatment process database into a thermal history feature vector, wherein the thermal history feature vector includes a temperature gradient distribution feature and a time series phase change feature.

[0050] In this step, the thermal history feature vector encodes the heat treatment process parameters (quenching temperature gradient, tempering time series) into continuous vector data, containing the quantitative characteristics of the spatial distribution of the temperature field and the temporal phase transformation process. The temperature gradient distribution characteristics refer to the rate and direction of temperature change in three dimensions during the heat treatment process, reflecting the distribution of thermal stress during the material cooling process. The temporal phase transformation characteristics refer to the dynamic characteristics of the metal phase transformation process in the time dimension, including parameters such as the duration of the phase transformation critical point and the frequency of temperature fluctuations.

[0051] This step extracts a set of process parameters corresponding to the complex alloy part from the metal heat treatment process database, and the process parameter set includes temperature sampling point data, time series data and phase change critical point records; the temperature sampling points in the temperature sampling point data are expanded to generate temperature gradient distribution characteristics, and the time series data and the phase change critical point records are time-series modeled to generate time series phase change characteristics; the temperature gradient distribution characteristics and the time series phase change characteristics are tensor-spliced ​​to form a thermal history feature vector as the process data input of the multimodal large model.

[0052] Step 104: fusing the thermal history feature vector and the multimodal image feature set using a large multimodal model to obtain a fusion result.

[0053] This step constructs a multimodal large model based on the attention mechanism. The multimodal large model includes a dual-channel architecture of a thermal history feature encoder and an image feature encoder; the cross-modal association weights are dynamically configured according to the material property parameters of the precision metal parts, and the material property parameters include the proportion of alloy element composition and the crystal structure type; the thermal history feature vector is nonlinearly transformed by the thermal history feature encoder to generate a process feature embedding vector; the multimodal image feature set is convolutionally extracted by the image feature encoder to generate a visual feature embedding vector; the process feature embedding vector is cross-modally fused with the visual feature embedding vector using cross-modal attention to generate a fusion processing result that includes material process constraints.

[0054] Step 105: Based on the fusion processing result, defect detection data is generated, where the defect detection data includes three-dimensional position information of grain boundary cracks and prediction data of crack propagation direction.

[0055] In this step, the fusion processing results are input into the defect detection decoder, which includes a three-dimensional space reconstruction module and a crack extension prediction module. The three-dimensional space reconstruction module is used to perform voxel-level positioning of grain boundary cracks to generate three-dimensional crack position information containing XYZ coordinates. The crack extension prediction module analyzes the stress field distribution characteristics at the crack tip and calculates the crack extension direction angle and potential extension path based on the temperature gradient distribution characteristics in the thermal history feature vector. The results of the above steps are integrated to generate defect detection data with spatial topological relationships.

[0056] The embodiments of the present invention suppress industrial site interference by constructing an electromagnetic shielding environment, and combine multi-scale feature analysis of laser scattering with heat treatment process data fusion to break through the bottleneck of traditional single-modal detection in identifying sub-surface defects and multi-physical field coupling defects, achieve three-dimensional precise positioning and expansion direction prediction of grain boundary cracks, improve the detection rate and efficiency of micron-level defects in complex electromagnetic environments, and systematically solve the problems of missed detection and misjudgment caused by data fragmentation and environmental sensitivity in existing technologies.

[0057] The present invention provides a specific embodiment, step 104, fusing the thermal history feature vector and the multimodal image feature set using a multimodal large model to obtain a fusion result, specifically comprising the following steps:

[0058] Step 401: Based on the material property parameters of the precision metal part, a weight control parameter set is generated, where the weight control parameter set includes a thermal history weight vector and an image weight vector.

[0059] In this step, material property parameters refer to a set of parameters that characterize the physical properties of precision metal parts. These parameters include the mass percentage composition of the main alloying elements (e.g., 85% titanium, 10% aluminum, and 5% vanadium) and the digital encoding of the crystal structure type (e.g., body-centered cubic is coded as 001, face-centered cubic is coded as 010). The weight control parameter set refers to a set of parameters consisting of a thermal history weight vector and an image weight vector. These parameters are generated from the material property parameters and are used to control dual-channel feature fusion. The thermal history weight vector is used to adjust the mapping intensity of the temperature gradient feature in the thermal history feature encoding channel. The vector length matches the dimension of the thermal history feature vector. The image weight vector is used to control the extraction ratio of grain boundary topology features in the image feature encoding channel. The vector length matches the number of multimodal image feature types.

[0060] This step extracts the composition ratio parameters of the main alloying elements and the encoding parameters of the crystal structure type from the material property database of precision metal parts to generate a composition ratio vector and a crystal structure encoding vector, wherein the composition ratio parameter is a vector composed of the mass percentages of each main alloying element; the encoding parameter is a one-hot encoding vector classified according to the symmetry of the crystal structure; the composition ratio vector and the crystal structure encoding vector are connected in head-to-tail order to form a combined attribute vector; the combined attribute vector is input into a preset weight generation function to generate a thermal history weight vector and an image weight vector, and the weight generation function contains an independent fully connected operation branch; wherein the length of the thermal history weight vector is equal to the number of input nodes of the thermal history feature encoding channel, which is used to adjust the mapping strength of the process features; the length of the image weight vector is equal to the number of input channels of the image feature encoding channel, which is used to control the extraction weight of the image features; the thermal history weight vector and the image weight vector together constitute a cross-modal weight control parameter set.

[0061] Step 402: In the thermal history feature encoding channel of the multimodal large model, the thermal history feature vector is mapped according to the thermal history weight vector to generate a process feature encoding sequence with temperature gradient correlation characteristics.

[0062] In this step, the thermal history feature encoding channel refers to the data processing channel in the multimodal large model that specifically processes the thermal history feature vector. Its input interface receives the thermal history feature vector, and its output interface generates a process feature encoding sequence. The number of nodes inside the channel is equal to the dimensional value of the thermal history feature vector. The temperature gradient correlation characteristic refers to the degree of correlation between each element value in the process feature encoding sequence and the temperature gradient distribution characteristics in the thermal history feature vector, which is specifically manifested as the characteristic that the element value increases as the temperature difference between adjacent temperature sampling points increases. The process feature coding sequence refers to the characteristic data sequence output by the thermal history feature encoding channel. Each element value in the sequence reflects the comprehensive influence of the temperature gradient at a specific spatial position on the phase change process of the material. The order of element arrangement is consistent with the dimensional order of the thermal history feature vector.

[0063] The thermal history feature encoding channel in this step is one of the dual-channel feature encoding architecture structures of the multimodal large model. The number of its input nodes is equal to the total dimension value of the thermal history feature vector, and the total dimension value is the sum of the dimensions of the temperature gradient distribution characteristics and time series phase change characteristics contained in the thermal history feature vector. Its input interface is directly connected to the data storage location of the thermal history feature vector; when performing nonlinear mapping processing in the thermal history feature encoding channel, each element value of the thermal history weight vector is multiplied by the corresponding dimension value of the thermal history feature vector to generate a weighted initial eigenvalue; the initial eigenvalue is subjected to nonlinear transformation processing, and the output value of the nonlinear transformation processing is added to the temperature gradient change output of the previous layer to generate the current element value of the process feature encoding sequence, and finally a process feature encoding sequence with temperature gradient correlation characteristics is formed based on the current element value, in which each element value increases linearly with the increase of the temperature gradient change.

[0064] Step 403: In the image feature coding channel of the multimodal large model, sliding window processing is performed on the multimodal image feature set according to the image weight vector to generate an image feature coding sequence with grain boundary topological correlation characteristics.

[0065] In this step, the image feature encoding channel refers to a data processing channel in the multimodal large model that specifically processes the multimodal image feature set. Its input interface receives grain boundary topological structure features and subsurface scattering spot distribution features, and its output interface generates an image feature encoding sequence. The channel's sliding window size is inversely proportional to the average grain boundary spacing. The grain boundary topological correlation characteristic refers to the degree of correlation between each element value in the image feature encoding sequence and the spatial distribution characteristics of the grain boundary structure, specifically manifested as the characteristic that the element value increases with increasing grain boundary density. The image feature encoding sequence refers to the feature data sequence output by the image feature encoding channel. Each element value in the sequence reflects the joint feature intensity of the grain boundary topological structure and the scattering spot distribution in a specific window area. The element arrangement order is consistent with the image acquisition space coordinate order.

[0066] The image feature encoding channel in this step is one of the dual-channel feature encoding architectures of the multimodal large model, and its number of input channels is equal to the total number of types of grain boundary topological structure features in the multimodal image features plus the number of types of sub-surface scattering spot distribution features; its input interface is directly connected to the data storage location of the multimodal image features; when performing sliding window processing in the image feature encoding channel, the window step size is calculated based on the average grain boundary spacing, and the window step size is the inverse of the average grain boundary spacing multiplied by a preset proportional coefficient; the window step size is used to slide on the grain boundary topological structure feature map, and the area covered by each sliding is the square of the window step size; the local statistics of the sub-surface scattering spot distribution characteristics are extracted at each window position, and the local statistics include the mean number of scattering spots and the variance of the distribution density; the local statistics are multiplied by the corresponding channel value of the image weight vector to generate the current element value of the image feature encoding sequence, and finally an image feature encoding sequence with grain boundary topological correlation characteristics is formed based on the current element value.

[0067] Step 404: performing weighted processing on the process feature coding sequence and the image feature coding sequence according to the weight control parameter set to generate a fusion coding tensor.

[0068] In this step, the fused coding tensor refers to a multidimensional data set formed by aligning the process feature coding sequence and the image feature coding sequence according to the element position and then weighted splicing them. Its number of dimensions is the sum of the number of dimensions of the two input sequences. The value of each position in the tensor is generated by superimposing the two input sequence values ​​at the corresponding position according to the weight.

[0069] In this step, each element value of the thermal history weight vector in the weight control parameter set is multiplied point by point with the corresponding element value of the process feature coding sequence to obtain the weighted process feature component; at the same time, each element value of the image weight vector in the weight control parameter set is multiplied point by point with the corresponding element value of the image feature coding sequence to obtain the weighted image feature component; the weighted process feature component and the weighted image feature component are aligned according to the element position and then spliced ​​to form a fused coding tensor, whose data dimension is the sum of the number of dimensions of the process feature coding sequence and the number of dimensions of the image feature coding sequence.

[0070] Step 405: Filter the fused coding tensor to generate an optimized fusion processing result.

[0071] This step extracts the data distribution of each feature dimension from the fusion coding tensor. The feature dimension is an independent data channel representing different feature types in the fusion coding tensor; the spatial distribution correlation of each feature dimension is calculated. The spatial distribution correlation is obtained by correlating the data sequence of the feature dimension at all detection positions with the pre-established grain boundary crack spatial distribution map, specifically the average value of the product of the numerical values ​​of the corresponding positions of the two data sequences minus the product of their respective average values; at the same time, the directional sensitivity of each feature dimension is calculated. The directional sensitivity is obtained by dividing the data variance value of the feature dimension in the main direction of crack propagation by the data variance value in the direction perpendicular to the main direction. The main direction of crack propagation is determined according to the stress field distribution model of precision metal parts; a spatial distribution correlation threshold and a directional sensitivity threshold are set, and feature dimensions that satisfy both a correlation greater than the threshold and a directional sensitivity greater than 1 are retained; the retained feature dimensions are reorganized according to the original arrangement order in the fusion coding tensor to generate an optimized fusion processing result.

[0072] The embodiments of the present invention realize the physical consistency fusion of thermal history process data and real-time image detection data, solve the problem of feature correlation inaccuracy caused by fixed weights in traditional multimodal fusion; enhance the coupling characterization capability of sub-surface defects and thermal stress cracks; and improve the recognition accuracy and model generalization performance of multi-physical field coupling defects in complex electromagnetic environments.

[0073] The present invention provides a specific embodiment, step 404, performing weighted processing on the process feature coding sequence and the image feature coding sequence according to the weight control parameter set to generate a fused coding tensor, specifically comprising the following steps:

[0074] Step 411: performing element-by-element multiplication operation on each element value of the thermal history weight vector in the weight control parameter set and the corresponding element value of the process feature coding sequence to generate a weighted process feature sequence.

[0075] In this step, the weighted process feature sequence refers to the process feature data sequence adjusted by the thermal history weight vector, and each element value reflects the weight enhancement result of the temperature gradient feature.

[0076] This step extracts each element value from the thermal history weight vector in sequence, and at the same time extracts the corresponding element value, that is, the temperature gradient association value, from the same position in the process feature coding sequence; multiplies the thermal history weight element value of each position with the process feature element value to generate the current element value of the weighted process feature sequence, and combines the current element values ​​to form a weighted process feature sequence. The number of dimensions of the weighted process feature sequence is consistent with the original process feature coding sequence.

[0077] Step 412: performing element-by-element multiplication operation on each element value of the image weight vector in the weight control parameter set and the corresponding element value of the image feature coding sequence to generate a weighted image feature sequence.

[0078] In this step, the weighted image feature sequence refers to the image feature data sequence adjusted by the image weight vector, and each element value reflects the weight enhancement result of the grain boundary topological feature.

[0079] This step extracts each element value from the image weight vector in channel order, and extracts the grain boundary topology association value from the corresponding channel position of the image feature coding sequence; multiplies the image weight element value of each channel with the image feature element value to generate the current channel value of the weighted image feature sequence; combines the current channel values ​​to form a weighted process feature sequence, and the number of channels of the weighted image feature sequence remains consistent with the original image feature coding sequence

[0080] Step 413: Determine a splicing ratio coefficient based on the material property parameters of the precision metal part, and perform dimension adjustment processing on the weighted process feature sequence and the weighted image feature sequence based on the splicing ratio coefficient to obtain an adjusted weighted process feature sequence and an adjusted weighted image feature sequence.

[0081] In this step, the stitching ratio coefficient refers to the fusion ratio control parameter generated by the material composition and crystal structure, which determines the dimensional adjustment ratio of the process and image features.

[0082] This step extracts the main alloying element composition ratio parameters from the material property parameters, assigns preset weights to each element based on its sensitivity to crack propagation, and calculates the weighted sum of all elements. This weighted sum is multiplied by the crystal structure encoding parameter (a numerical encoding, such as 1.2 for body-centered cubic and 1.5 for face-centered cubic) to generate a splicing scale factor, which is used to control the fusion ratio of process and image features. The original dimension of the weighted process feature sequence is multiplied by the splicing scale factor to expand its dimension to the new value. Simultaneously, the original dimension of the weighted image feature sequence is divided by the splicing scale factor to compress its dimension to the new value. After adjustment, the dimensions of the two sequences are equal to ensure the feasibility of subsequent splicing.

[0083] Step 414: interlacing and splicing the adjusted weighted process feature sequence and the adjusted weighted image feature sequence according to the order of element positions to generate initial fused coded data.

[0084] In this step, element position order refers to the order in which the elements in the data sequence are arranged according to the spatial coordinates of the detection area, ensuring a one-to-one correspondence between features and physical locations. Initial fused encoded data refers to the raw fused data formed by the interleaving of weighted processes and image features, including both valid and redundant features.

[0085] When performing interleaved stitching in this step, the adjusted weighted process feature sequence elements and the adjusted weighted image feature sequence elements are alternately arranged in the order of element positions to generate initial fused coding data. Specifically, odd-numbered positions are filled with process feature elements, and even-numbered positions are filled with image feature elements; the alternating arrangement order is consistent with the physical spatial distribution of the detection area.

[0086] Step 415: performing target redundant dimension labeling processing on the initial fused coded data to obtain labeled fused coded data.

[0087] This step calculates the positional correlation between each dimension of the initial fused encoded data and the spatial distribution map of grain boundary cracks: if the numerical fluctuation amplitude of a certain dimension in the high-incidence crack area is less than a preset threshold, it is marked as a redundant dimension; the spatial distribution map of grain boundary cracks is the output of a pre-established defect distribution statistical model.

[0088] Step 416: Remove all marked target redundant dimensions in the fused coded data after the marking to generate a fused coded tensor.

[0089] This step traverses all dimensions of the initial fused encoded data, deletes all dimension data marked as redundant, reorganizes the retained dimension data into the original spatial order, and generates a fused encoded tensor whose number of dimensions is the total number of retained valid dimensions.

[0090] The embodiments of the present invention solve the problems of unreasonable weight distribution of process features and detection features in traditional methods, poor adaptability caused by fixed multimodal feature fusion ratio, destruction of spatial correlation by traditional serial splicing, and increased false detection rate caused by feature redundancy; and break through the feature extraction bias caused by static weights.

[0091] The present invention provides a specific embodiment, step 101, obtaining laser scattering image data of a precision metal part in an electromagnetic shielding environment, specifically comprising the following steps:

[0092] Step 111: constructing a closed detection space having a multi-layer conductive shielding structure, wherein the multi-layer conductive shielding structure includes alternately arranged metal mesh layers and absorbing material layers.

[0093] In this step, the multi-layer conductive shielding structure refers to a composite barrier formed by alternating stacking of two or more layers of materials with different conductive properties. Each layer of material has specific electromagnetic wave reflection and absorption characteristics, which is used to block the interference of external electromagnetic fields on the detection process. The closed detection space refers to a physically isolated area surrounded by a multi-layer conductive shielding structure. Its internal environment and the external electromagnetic field achieve attenuation isolation of at least 40dB, providing a controlled electromagnetic environment for laser detection. The metal mesh layer refers to a mesh structure layer woven from conductive metal wires. The mesh geometric parameters match the frequency band of the electromagnetic wave to be shielded, and block the penetration of high-frequency electromagnetic waves through reflection. The absorbing material layer refers to a polymer composite material layer filled between the metal mesh layers, which converts the energy of the transmitted electromagnetic wave into heat energy through dielectric loss, thereby suppressing multiple reflections of the electromagnetic wave in the shielding structure.

[0094] In this step, the metal mesh layer and the absorbing material layer are first stacked in an alternating arrangement to perform a multi-layer structure assembly to form a closed detection space. A laser emitting device and an optical sensing device are set in the closed detection space. The laser emitting device projects a polarized laser beam onto the surface of the precision metal part at a preset incident angle, and the preset incident angle is dynamically adjusted according to the surface curvature of the precision metal part; wherein, the surface roughness of the inner wall of the closed detection space is controlled within 0.4μm to reduce internal light reflection interference; the aperture size of each grid in the metal mesh layer is determined according to the maximum wavelength of the electromagnetic wave to be suppressed, specifically, the maximum wavelength value is divided by 5 as the upper limit of the aperture size; the thickness of the absorbing material layer is adjusted according to the conductivity gradient of the adjacent metal mesh layer to ensure that the dielectric constant of each layer of absorbing material matches the impedance characteristics of the adjacent metal mesh layer.

[0095] Step 112: adjusting the laser incident angle of the laser emitting device in the closed detection space according to the surface curvature distribution data of the precision metal part to obtain an adjusted laser incident angle.

[0096] In this step, surface curvature distribution data refers to a set of quantitative parameters that characterize the degree of curvature at each point on the surface of a precision metal part. This data is acquired using a non-contact 3D profilometer and is presented as a distribution graph of curvature radius values ​​as a function of surface coordinates. The laser incident angle, defined as the angle between the laser beam projected by the emitting device and the normal to the inspection point on the surface of the precision metal part, is dynamically adjusted based on the surface curvature distribution data to optimize signal acquisition efficiency.

[0097] This step first obtains the surface curvature distribution data of the precision metal part through the surface profile measuring device, and calculates the corresponding laser incident angle compensation value based on the curvature radius of each measuring point in the surface curvature distribution data. The compensation value is the product of the inverse of the curvature radius and a preset coefficient; the laser emitting device dynamically adjusts the projection direction of the polarized laser beam according to the compensation value, so that the deviation angle between the incident direction of the laser beam and the surface normal direction of the detection point is always less than 2 degrees, and finally the adjusted laser incident angle is obtained; at the same time, the field of view angle range of the optical sensor device is dynamically and synchronously adjusted according to the incident angle of the laser emitting device to ensure that the overlap between the reflected light signal receiving coverage area and the laser projection area is greater than 95%.

[0098] Step 113: Scan the surface of the precision metal part based on the adjusted laser incident angle to obtain reflected light signal data.

[0099] In this step, the reflected light signal data refers to the set of original electrical signals generated by photoelectric conversion of the laser signal reflected from the surface of the precision metal part, which includes the signal modulation characteristics caused by the surface morphology characteristics and internal defects.

[0100] In this step, in a closed detection space, the surface of the precision metal part is scanned by a polarized laser beam with an adjusted laser incident angle, and reflected light signal data scattered by the surface of the precision metal part is received by an optical sensing device, wherein the reflected light signal data includes light intensity distribution data modulated by surface grain boundary characteristics and light spot morphology data modulated by sub-surface defects; specifically, the photoelectric conversion unit in the optical sensing device is used to convert the light signal into a current signal, wherein the portion of the current signal modulated by the surface grain boundary characteristics manifests as periodic intensity fluctuations, and its fluctuation frequency is proportional to the grain boundary density; the portion modulated by the sub-surface defects manifests as topological distortion of the light spot morphology, and the degree of topological distortion is exponentially related to the defect depth; the signal transmission channel in the optical sensing device adopts a coaxial shielding structure, and the grounding impedance of the shielding layer is less than 0.1Ω, so as to isolate electromagnetic interference outside the closed detection space.

[0101] Step 114: Eliminate the signal distortion component of the reflected light signal data to generate optical response data, perform correction processing on the optical response data, and generate laser scattering image data to be optimized.

[0102] In this step, the signal distortion component refers to the non-target-related signal component in the reflected light signal introduced by external electromagnetic field coupling or equipment noise, manifesting as random fluctuations or periodic interference unrelated to the surface physical characteristics. The optical response data refers to the subset of reflected light signals retained after eliminating the signal distortion component. Its signal amplitude and frequency characteristics have a definite correspondence with the physical properties of the surface of precision metal parts. The laser scattering image data to be optimized refers to the initial image data set generated by mapping the optical response data according to spatial coordinates. It contains a mixture of effective detection signals and residual interference signals.

[0103] This step first establishes an electromagnetic field intensity distribution model in a closed detection space. The model input includes the aperture distribution data of the metal grid layer and the absorption efficiency parameters of the absorbing material layer. Based on the spatial electromagnetic field intensity gradient map output by the model, the signal distortion component related to the gradient map distribution is separated from the reflected light signal data. The signal distortion component is eliminated by a signal subtractor, and only the light intensity change data directly caused by the physical characteristics of the surface of the precision metal part is retained in the remaining signal components to generate optical response data. Based on the proportional relationship between the physical size of the pixel unit in the optical sensing device and the actual size of the surface of the precision metal part, a mapping rule between pixel coordinates and actual spatial coordinates is established. A material refractive index parameter correction factor is introduced into the mapping rule for correction. The correction factor is the ratio of the material refractive index to the standard reference refractive index. The corrected mapping rule is applied to the optical response data to generate the grayscale value of each pixel point in the initial laser scattering image data. The grayscale value is linearly positively correlated with the light intensity value at the corresponding position, ultimately forming the laser scattering image data to be optimized.

[0104] Step 115: performing interference component separation processing on the laser scattering image data to be optimized to extract target scattering data and generate laser scattering image data.

[0105] In this step, the interference component refers to invalid image information in the laser scattering image data to be optimized, originating from non-target features such as background reflections from the detection environment and thermal noise from the equipment. The target scattering data refers to the subset of scattered signals directly related to the surface of the precision metal part, extracted from the data to be optimized through physical spatial filtering. Its data distribution characteristics reflect the true physical state of surface and subsurface defects.

[0106] In this step, a three-dimensional coordinate system is established in a closed detection space with the surface of the precision metal part as the reference plane, and the vertical distance from each scattered signal source to the reference plane is calculated. A fixed distance threshold is set to 10 times the laser wavelength, and scattered signal components with vertical distances exceeding the fixed distance threshold are filtered out through a distance filter. Spatial clustering analysis is performed on the remaining signal components, and cluster area data with signal density greater than the preset threshold is retained. Finally, laser scattering image data is generated, in which each pixel corresponds to a scattering signal source that has passed the screening.

[0107] The embodiments of the present invention effectively suppress the influence of electromagnetic interference in industrial environments on scattered signals; improve the imaging quality of surface and sub-surface defects of complex curved test pieces; and systematically solve the problem of missed detection of micron-level defects caused by electromagnetic pollution, surface geometric distortion and noise coupling in existing open environment detection solutions.

[0108] The present invention provides a specific embodiment, step 114, eliminating the signal distortion component of the reflected light signal data to generate optical response data, and performing correction processing on the optical response data to generate laser scattering image data to be optimized, specifically includes the following steps:

[0109] Step 121: Detect the spectrum distribution pattern of electromagnetic interference in the closed detection space and extract the target interference frequency component.

[0110] In this step, electromagnetic interference refers to the remaining electromagnetic noise signal within the shielded closed test space, with a frequency range of 10kHz-6GHz. The spectrum distribution pattern refers to the energy distribution characteristics of the electromagnetic interference signal at different frequencies, including the fundamental wave, harmonics, and random noise components. The target interference frequency components are electromagnetic signal components unrelated to surface defects in precision metal parts and need to be suppressed or eliminated.

[0111] When detecting the spectrum distribution pattern of residual electromagnetic interference in the closed detection space, a wide-band spectrum analyzer is used to scan the electromagnetic signal in the 10kHz-6GHz frequency band, and record the distribution curve of the signal amplitude changing with frequency; the periodic peak component synchronized with the laser scanning timing and the broadband noise component related to the ambient electromagnetic field in the distribution curve are extracted as the target interference frequency component.

[0112] Step 122: Based on the aperture size of the metal mesh layer and the attenuation coefficient of the absorbing material layer, generate a compensation signal waveform with a phase opposite to that of the target interference frequency component.

[0113] In this step, the attenuation coefficient of the absorbing material layer refers to the material's ability to attenuate electromagnetic waves per unit thickness (dB / m), which is related to the material's dielectric constant and magnetic permeability. The compensation signal waveform is a cancellation signal with an opposite phase and amplitude to the interference signal waveform, and its duration is synchronized with the laser scanning period.

[0114] This step calculates the cutoff frequency based on the aperture size of the metal mesh layer. The calculation formula is: speed of light / (5×aperture size); based on the attenuation coefficient of the absorbing material layer at the cutoff frequency (dB / m), based on: original interference amplitude × = compensation signal amplitude, calculate the compensation signal amplitude, and generate a compensation signal waveform that is identical to the interference frequency component waveform and opposite in phase, and the duration of the waveform is strictly synchronized with the laser scanning period.

[0115] Step 123: superimpose the compensation signal waveform and the reflected light signal data to obtain compensated reflected light signal data.

[0116] In this step, the compensated reflected light signal data refers to the reflected light signal after the compensation signal is superimposed, the periodic interference is suppressed, and the effective defect signal is retained.

[0117] In this step, the compensation signal waveform and reflected light signal data are input into a differential amplifier to adjust the compensation signal delay so that it is time-aligned with the interference component. The amplitude of the superimposed compensation signal is adjusted by a gain controller to obtain a superimposed signal, in which the amplitude of its periodic interference component is attenuated to less than 5% of the original signal, thereby generating the compensated reflected light signal data.

[0118] Step 124: Eliminate the periodic interference component in the compensated reflected light signal data, retain the target light signal, process the target light signal, and generate optical response data including light intensity distribution and light spot morphology information.

[0119] In this step, the periodic interference component refers to a sinusoidal or pulsed interference signal that is homologous to the laser scanning frequency and appears as discrete peaks in the spectrum. The target optical signal is the effective optical signal modulated by surface grain boundaries and subsurface defects, carrying defect characteristic information. The light intensity distribution refers to the spatial distribution of the reflected light intensity at each point in the detection area, reflecting surface microstructural differences. The spot morphology information refers to the geometric characteristics of the reflected light spot, including parameters such as diameter, ellipticity, and edge sharpness.

[0120] To eliminate periodic interference components, this step performs a sliding window energy analysis on the compensated reflected light signal data, with a window width three times the laser pulse width. If the standard deviation of the signal energy fluctuation within the window is less than a set threshold, it is determined to be residual random noise and filtered out, retaining the target light signal modulated by surface grain boundary scattering and subsurface defect diffraction. A photodiode array is used to convert the target light signal into a current signal, where the current value I and the light intensity P satisfy I = kP (k is the photoelectric conversion coefficient). The current signal is quantized using a 16-bit analog-to-digital converter at a 1MHz sampling rate to generate optical response data containing light intensity distribution (P value at each point) and spot morphology information (spot diameter and ellipticity).

[0121] Step 125: Correct the optical response data according to the adjusted laser incident angle and the surface refractive index parameter of the precision metal part to generate laser scattering image data to be optimized.

[0122] In this step, the surface refractive index parameter refers to the refractive characteristic value of the surface of the precision metal part to the laser wavelength.

[0123] In this step, when the image data is generated by correction processing, the surface normal direction correction factor C=cosθ / n² (n is the surface refractive index parameter) is calculated according to the adjusted laser incident angle θ; the light intensity value P in the optical response data is multiplied by C and arranged into a two-dimensional pixel array according to the XY coordinate order of the scanning path; and the nonlinear mapping function is used to calculate the surface normal direction correction factor C=cosθ / n² (n is the surface refractive index parameter). Generate pixel grayscale values ​​to form laser scattering image data to be optimized.

[0124] The embodiments of the present invention solve the problem of incomplete residual electromagnetic interference detection in traditional methods, improve recognition accuracy, and cover a large interference frequency band; effectively eliminate periodic interference and micro-crack signal amplitude distortion; can eliminate most geometric distortions of curved workpieces and reduce spatial positioning errors.

[0125] The present invention provides a specific embodiment, step 103, converting the process parameters corresponding to the precision metal part in the existing metal heat treatment process database into a thermal history feature vector, wherein the thermal history feature vector includes temperature gradient distribution characteristics and time series phase change characteristics, specifically comprising the following steps:

[0126] Step 301: Extract a set of process parameters corresponding to the precision metal part from an existing metal heat treatment process database, wherein the set of process parameters includes temperature sampling point data of a high-temperature treatment stage higher than a preset temperature value and time series data of a low-temperature holding stage lower than a preset temperature value.

[0127] In this step, the preset temperature value refers to the critical temperature threshold that separates the high-temperature treatment phase from the low-temperature hold phase. It is dynamically set based on the material's melting point characteristics and is calculated as: material melting point × 0.6. Temperature sampling point data refers to the set of temperature measurements recorded at specific spatial locations and time points during the heat treatment process, consisting of three-dimensional coordinates (X, Y, Z) and temperature values ​​(T). Time series data refers to the set of temperature values ​​recorded during the low-temperature hold phase in timestamp order, structured as time-temperature pairs (t, T).

[0128] In this step, when extracting the process parameter set corresponding to the precision metal parts from the existing metal heat treatment process database, the process record table in the database is first matched according to the unique identification code of the precision metal parts; the extraction conditions of the process parameter set are: all temperature sampling point data whose temperature values ​​exceed the preset temperature value in the high-temperature treatment stage, and all time series data whose temperature values ​​are lower than the preset temperature value in the low-temperature holding stage, where the preset temperature value is set according to the material melting point characteristics of the precision metal parts.

[0129] Step 302: Based on the spatial position information of each temperature sampling point and the temperature variation of adjacent points in the temperature sampling point data, the temperature sampling points are expanded to generate temperature gradient distribution features.

[0130] In this step, spatial location information refers to the coordinate data of the temperature sampling points in three-dimensional space, which is used to construct the spatial correlation relationship of the temperature distribution. The temperature change of adjacent points refers to the temperature difference between two adjacent temperature sampling points in the same grid divided by the distance between their nodes. This is used to calculate the direction of the temperature gradient. The temperature gradient distribution feature refers to the continuous temperature distribution data generated by interpolation and expansion, which characterizes the temperature gradient variation characteristics in three-dimensional space.

[0131] In this step, the temperature sampling point data includes the three-dimensional spatial position information and corresponding temperature values ​​recorded at each sampling moment; a three-dimensional grid coordinate system is constructed based on the three-dimensional spatial position information, and the temperature change between each grid node and its adjacent nodes is calculated. The temperature change calculation formula is: the difference between the temperature values ​​of adjacent nodes divided by the node spacing; the temperature value of the area not directly sampled is extrapolated based on the temperature change, and the discrete temperature sampling points are expanded into a continuously distributed temperature gradient surface through linear interpolation method; the grid node temperature values ​​of the temperature gradient surface are arranged in spatial coordinate order to construct the temperature gradient distribution feature.

[0132] Step 303: extracting the critical duration of the phase change critical point and the temperature fluctuation frequency in the time series data to generate a time series phase change feature.

[0133] In this step, the phase transition critical point refers to the starting time point when the temperature fluctuations during the material phase transition process tend to stabilize, and is determined by detecting the change in the slope of the temperature curve. The critical duration refers to the duration of the stable state corresponding to the phase transition critical point, calculated by subtracting the start time from the end time. The temperature fluctuation frequency characteristic refers to the number of temperature fluctuations per unit time, calculated by dividing the number of fluctuations by the duration. The time series phase transition feature refers to the feature vector formed by encoding the duration feature and the temperature fluctuation frequency feature in chronological order, which characterizes the dynamic characteristics of the phase transition process.

[0134] This step extracts the phase transition critical point from the time series data. The phase transition critical point is the starting time point when the absolute value of the slope in the temperature fluctuation curve is first less than the preset threshold. The duration of the stable interval corresponding to each phase transition critical point is counted, and its characteristic is the difference between the end time and the start time of the critical point. The number of temperature fluctuations of the temperature wave is divided by the duration to calculate the temperature fluctuation frequency characteristics within the stable interval. The duration and temperature fluctuation frequency characteristics are encoded in chronological order to form the time series phase transition characteristics.

[0135] Step 304: The temperature gradient distribution feature and the time series phase change feature are combined to generate a thermal history feature vector. The dimension of the thermal history feature vector is proportional to the number of material components of the precision metal part.

[0136] This step connects the one-dimensional array of temperature gradient distribution characteristics and the one-dimensional array of time series phase change characteristics in head-to-tail order, and inserts a dimension separation identifier at the connection point to generate a thermal history feature vector, wherein the dimension separation identifier contains the number of material component types of the precision metal part, specifically a binary encoding of the number of component types; the total number of dimensions of the thermal history feature vector is equal to the sum of the number of dimensions of the temperature gradient distribution characteristics and the number of dimensions of the time series phase change characteristics, and is proportional to the number of material component types of the precision metal part.

[0137] The embodiments of the present invention solve the problem of insufficient defect-process correlation modeling caused by discrete input of traditional process data; achieve accurate mapping of heat treatment history and defect formation mechanism of different alloy systems, and improve the detection sensitivity of multimodal large models for process-related defects such as thermal stress cracks and phase transformation-induced defects.

[0138] The present invention provides a specific embodiment, step 105, generating defect detection data based on the fusion processing result, wherein the defect detection data includes three-dimensional position information of grain boundary cracks and prediction data of crack propagation direction, specifically comprising the following steps:

[0139] Step 501: converting the detection position corresponding to the characteristic dimension data related to the spatial distribution of grain boundary cracks in the fusion processing result into a three-dimensional coordinate value in a preset physical space coordinate system to generate spatial coordinate data.

[0140] In this step, the spatial distribution of grain boundary cracks refers to the location and orientation characteristics of grain boundary cracks in the three-dimensional space of precision metal parts, including parameters such as length, depth, and number of bifurcations. Feature dimension data refers to the data channels in the fusion processing results that are directly related to grain boundary crack detection, with each dimension corresponding to a defect-related feature. The preset physical space coordinate system refers to the coordinate system established based on the calibration parameters of the laser scattering system, with the origin at the geometric center of the detection area and the coordinate axes aligned with the scanning direction. Spatial coordinate data refers to a structured data set that includes the three-dimensional positions (X, Y, Z) of all detection points and the corresponding feature dimension values.

[0141] This step first extracts the detection position information corresponding to each feature dimension in the feature dimension data. Combined with the calibration parameters of the preset physical space coordinate system, which include lateral resolution and longitudinal resolution, the detection position information is converted into three-dimensional coordinate values. The conversion formula is: X coordinate = lateral pixel index × lateral resolution, Y coordinate = longitudinal pixel index × longitudinal resolution, Z coordinate = feature dimension value × inter-layer resolution. Spatial coordinate data containing the three-dimensional coordinates of all detection points is generated. The lateral pixel index is the column number of the pixel point in the laser scattering image data, and the longitudinal pixel index is the row number, which are automatically generated by the image acquisition system according to the scanning sequence.

[0142] Step 502: Identify the coordinates of the numerical mutation region and the target gradual change region of the spatial coordinate data to generate an initial crack position set, wherein the initial crack position set includes crack tip positions and branch node positions.

[0143] In this step, the region of numerical mutation refers to the spatial location where the characteristic dimension value changes dramatically in a local area (standard deviation > 3 times the mean). The target gradual change region refers to the spatial location where the characteristic dimension value changes steadily along a specific direction (rate of change > 10% / point). The crack tip location refers to the stress concentration point at the crack propagation front, where the characteristic dimension value changes most significantly. The branch node location refers to the key point where the crack path forks or turns, where the characteristic dimension value exhibits a continuous gradual change characteristic.

[0144] In this step, the standard deviation of the characteristic dimension values ​​within the region with a radius of three times the grain size around each coordinate point in the spatial coordinate data is calculated; if the standard deviation exceeds three times the average value of the adjacent region, it is marked as a numerical mutation region; the change rate of the characteristic dimension values ​​of adjacent coordinate points is detected, and if the change rate of five consecutive points is in the same direction and exceeds 10% / point, it is marked as a target gradual change region; the coordinates of all marked points are merged to form an initial crack position set containing the crack tip position and branch node position.

[0145] Step 503: Calculate the stress field distribution vector at the crack tip position based on the temperature gradient distribution feature in the thermal history feature vector and the sub-surface scattering spot distribution gradient feature in the multimodal image feature set.

[0146] In this step, the stress field distribution vector refers to a three-dimensional vector that characterizes the force direction and intensity at the crack tip, and its direction is synthesized by the temperature gradient and the sub-surface scattering spot density gradient.

[0147] In this step, the temperature gradient direction vector of the area corresponding to the crack tip position is extracted from the temperature gradient distribution characteristics; the scattering spot density gradient direction vector of the same area is extracted from the subsurface scattering spot distribution characteristics; the two direction vectors are synthesized according to the weights, according to: stress field distribution vector = temperature gradient direction vector × ln (peak temperature / 800) + density gradient direction vector × , generating stress field distribution vectors.

[0148] Step 504: Bind the coordinates of the crack tip position with the corresponding stress field distribution vector to generate bound data, and combine the change rate of the characteristic dimension value of the connection path between the branch node position and the adjacent crack tip position to generate primary defect data.

[0149] In this step, primary defect data refers to a set of unoptimized original defect information, including location, orientation, and preliminary topological relationship.

[0150] In this step, the three-dimensional coordinate data of the crack tip position is bound to the corresponding stress field distribution vector in the following format: [X coordinate, Y coordinate, Z coordinate, vector X component, vector Y component, vector Z component]. For each branch node position, the change rate of the characteristic dimension value of the connection path between it and the adjacent tip position is recorded to generate primary defect data containing spatial topological relationships.

[0151] Step 505: Perform topological connection verification on the primary defect data to generate defect detection data.

[0152] This step verifies the topological connectivity of the primary defect data. Specifically, it detects isolated coordinate points in the primary defect data (no other crack points within 5 times the grain size) and deletes them if they exist. For the fracture paths in the primary defect data, it inserts completion points based on the consistency of the vector directions of adjacent paths and the continuity of the characteristic dimension changes (difference in change rate <5%). Finally, it outputs the optimized defect detection data.

[0153] The embodiments of the present invention solve the problem of deviation between virtual coordinates and actual object positions in traditional detection methods, achieve submillimeter precision positioning through physical calibration parameters; overcome the problem of missed detection of microcracks and realize automatic identification of microcracks; solve the problem of inaccurate prediction of crack propagation direction; eliminate pseudo cracks caused by detection noise, and ensure that the crack propagation direction conforms to the laws of material fracture mechanics.

[0154] Figure 2 The present invention provides a structural diagram of a precision metal parts defect detection system based on a multi-modal large model, as shown in FIG. Figure 2 As shown, the system includes:

[0155] An acquisition module 21 is used to acquire laser scattering image data of a precision metal part in an electromagnetic shielding environment;

[0156] An analysis module 22 is configured to perform multi-scale spatial analysis on the laser scattering image data to extract a multimodal image feature set including the grain boundary topology of the surface of the precision metal part and the scattering spot distribution gradient characteristics of the subsurface;

[0157] A conversion module 23 is used to convert the process parameters corresponding to the precision metal part in the existing metal heat treatment process database into a thermal history feature vector, wherein the thermal history feature vector includes a temperature gradient distribution feature and a time series phase change feature;

[0158] A fusion module 24 is configured to fuse the thermal history feature vector and the multimodal image feature set using a multimodal large model to obtain a fusion result;

[0159] The generation module 25 is configured to generate defect detection data based on the fusion processing result, wherein the defect detection data includes three-dimensional position information of grain boundary cracks and prediction data of crack propagation direction.

[0160] Figure 2 The precision metal parts defect detection system based on multi-modal large model can be performed Figure 1 The implementation principle and technical effects of the multimodal large model-based precision metal part defect detection method described in the illustrated embodiment will not be elaborated on here. The specific manner in which each module and unit performs operations in the multimodal large model-based precision metal part defect detection system described in the above embodiment has been described in detail in the relevant embodiments of the method and will not be elaborated on here.

[0161] In one possible design, Figure 2 The embodiment shown is a precision metal parts defect detection system based on a multimodal large model that can be implemented as a computing device, such as Figure 3 As shown, the computing device may include a storage component 31 and a processing component 32;

[0162] The storage component 31 stores one or more computer instructions, wherein the one or more computer instructions are called and executed by the processing component 32 .

[0163] The processing component 32 is used to: obtain laser scattering image data of precision metal parts in an electromagnetic shielding environment; perform multi-scale spatial analysis on the laser scattering image data to extract a multimodal image feature set including the grain boundary topology structure of the surface of the precision metal parts and the gradient characteristics of the scattering spot distribution of the sub-surface; convert the process parameters corresponding to the precision metal parts in the existing metal heat treatment process database into thermal history feature vectors, and the thermal history feature vectors include temperature gradient distribution characteristics and time series phase change characteristics; fuse the thermal history feature vectors and the multimodal image feature set through a multimodal large model to obtain a fusion processing result; based on the fusion processing result, generate defect detection data, and the defect detection data includes three-dimensional position information of grain boundary cracks and predicted data of crack extension direction.

[0164] The processing component 32 may include one or more processors to execute computer instructions to complete all or part of the steps in the above method. Of course, the processing component may also be implemented as one or more application-specific integrated circuits (ASICs), digital signal processors (DSPs), digital signal processing devices (DSPDs), programmable logic devices (PLDs), field programmable gate arrays (FPGAs), controllers, microcontrollers, microprocessors, or other electronic components to perform the above method.

[0165] The storage component 31 is configured to store various types of data to support operations at the terminal. The storage component can be implemented by any type of volatile or non-volatile memory device, or a combination thereof, such as static random access memory (SRAM), electrically erasable programmable read-only memory (EEPROM), erasable programmable read-only memory (EPROM), programmable read-only memory (PROM), read-only memory (ROM), magnetic memory, flash memory, magnetic disk, or optical disk.

[0166] Of course, a computing device may also include other components, such as input / output interfaces, display components, communication components, etc.

[0167] The input / output interface provides an interface between the processing component and the peripheral interface module, which can be an output device, an input device, etc.

[0168] The communication component is configured to facilitate, among other things, wired or wireless communications between the computing device and other devices.

[0169] Among them, the computing device can be a physical device or an elastic computing host provided by a cloud computing platform, etc. In this case, the computing device can refer to a cloud server, and the above-mentioned processing components, storage components, etc. can be basic server resources rented or purchased from the cloud computing platform.

[0170] The embodiment of the present invention further provides a computer storage medium storing a computer program, which can achieve the above-mentioned Figure 1 The illustrated embodiment is a method for detecting defects in precision metal parts based on a multimodal large model.

[0171] Those skilled in the art will clearly understand that, for the convenience and brevity of description, the specific working processes of the systems, devices and units described above can refer to the corresponding processes in the aforementioned method embodiments and will not be repeated here.

[0172] The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate, and the components shown as units may or may not be physical units, i.e., they may be located in one location or distributed across multiple network units. Some or all of the modules may be selected based on actual needs to achieve the objectives of the present embodiment. Persons of ordinary skill in the art will be able to understand and implement the present invention without inventive effort.

[0173] Through the above description of the embodiments, those skilled in the art will clearly understand that each embodiment can be implemented using software plus a necessary general-purpose hardware platform, or of course, hardware. Based on this understanding, the essence of the above technical solution, or the portion that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, a magnetic disk, or an optical disk, and includes a number of instructions for causing a computer device (such as a personal computer, server, or network device) to execute the methods described in each embodiment or certain portions of the embodiments.

[0174] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or make equivalent replacements for some of the technical features therein. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the various embodiments of the present invention.

Claims

1. A method for detecting defects in precision metal parts based on a multimodal large model, characterized in that: include: Acquire laser scattering image data of precision metal parts in an electromagnetic shielding environment; Performing multi-scale spatial analysis on the laser scattering image data to extract a multimodal image feature set including the grain boundary topology of the surface of the precision metal part and the scattering spot distribution gradient characteristics of the subsurface; Converting the process parameters corresponding to the precision metal part in the existing metal heat treatment process database into a thermal history feature vector, wherein the thermal history feature vector includes a temperature gradient distribution feature and a time series phase change feature; fusing the thermal history feature vector and the multimodal image feature set through a multimodal large model to obtain a fusion processing result; generating defect detection data based on the fusion processing result, wherein the defect detection data includes three-dimensional position information of grain boundary cracks and prediction data of crack propagation direction; The thermal history feature vector and the multimodal image feature set are fused using a multimodal large model to obtain a fusion processing result, including: Based on the material property parameters of the precision metal part, a weight control parameter set is generated, and the weight control parameter set includes a thermal history weight vector and an image weight vector; in the thermal history feature encoding channel of the multimodal large model, the thermal history feature vector is mapped according to the thermal history weight vector to generate a process feature encoding sequence with temperature gradient correlation characteristics; in the image feature encoding channel of the multimodal large model, a sliding window processing is performed on the multimodal image feature set according to the image weight vector to generate an image feature encoding sequence with grain boundary topology correlation characteristics; the process feature encoding sequence and the image feature encoding sequence are weighted according to the weight control parameter set to generate a fused encoding tensor; the fused encoding tensor is screened to generate a fusion processing result; Based on the fusion processing results, defect detection data is generated, including: The fusion processing results are input into a defect detection decoder, which includes a three-dimensional space reconstruction module and a crack propagation prediction module. The grain boundary cracks are located at the voxel level by the three-dimensional space reconstruction module to generate an initial crack position set. The stress field distribution vector at the crack tip position is analyzed by the crack propagation prediction module, and combined with the temperature gradient distribution characteristics in the thermal history feature vector, defect detection data with a spatial topological relationship is generated.

2. The method according to claim 1, characterized in that The process feature coding sequence and the image feature coding sequence are weighted according to the weight control parameter set to generate a fusion coding tensor, including: Performing element-by-element multiplication operation on each element value of the thermal history weight vector in the weight control parameter set and the corresponding element value of the process feature coding sequence to generate a weighted process feature sequence; Performing element-by-element multiplication on each element value of the image weight vector in the weight control parameter set and the corresponding element value of the image feature coding sequence to generate a weighted image feature sequence; Determining a splicing ratio coefficient according to material property parameters of the precision metal part, and performing dimension adjustment processing on the weighted process feature sequence and the weighted image feature sequence according to the splicing ratio coefficient to obtain an adjusted weighted process feature sequence and an adjusted weighted image feature sequence; The adjusted weighted process feature sequence and the adjusted weighted image feature sequence are interleaved and spliced ​​in element position order to generate initial fusion coding data; Performing target redundant dimension labeling processing on the initial fused coded data to obtain labeled fused coded data; After removing the labels, all marked target redundant dimensions in the fused coded data are fused to generate a fused coded tensor.

3. The method according to claim 1, characterized in that Acquire laser scattering image data of precision metal parts in an electromagnetic shielding environment, including: Constructing a closed detection space having a multi-layer conductive shielding structure, wherein the multi-layer conductive shielding structure includes alternating metal mesh layers and absorbing material layers; adjusting the laser incident angle of the laser emitting device in the closed detection space according to the surface curvature distribution data of the precision metal part to obtain an adjusted laser incident angle; Scanning the surface of the precision metal part based on the adjusted laser incident angle to obtain reflected light signal data; Eliminating a signal distortion component of the reflected light signal data to generate optical response data, and performing correction processing on the optical response data to generate laser scattering image data to be optimized; Interference component separation processing is performed on the laser scattering image data to be optimized to extract target scattering data and generate laser scattering image data.

4. The method according to claim 3, characterized in that Eliminating the signal distortion component of the reflected light signal data to generate optical response data, and performing correction processing on the optical response data to generate laser scattering image data to be optimized, including: detecting a spectrum distribution pattern of electromagnetic interference within the enclosed detection space and extracting a target interference frequency component; generating a compensation signal waveform having a phase opposite to that of the target interference frequency component based on the aperture size of the metal mesh layer and the attenuation coefficient of the absorbing material layer; Superimposing the compensation signal waveform with the reflected light signal data to obtain compensated reflected light signal data; Eliminating the periodic interference component in the compensated reflected light signal data, retaining the target light signal, and processing the target light signal to generate optical response data including light intensity distribution and light spot morphology information; The optical response data is corrected according to the adjusted laser incident angle and the surface refractive index parameter of the precision metal part to generate laser scattering image data to be optimized.

5. The method according to claim 1, wherein The process parameters corresponding to the precision metal part in the existing metal heat treatment process database are converted into a thermal history feature vector. The thermal history feature vector includes temperature gradient distribution characteristics and time series phase change characteristics, including: Extracting a set of process parameters corresponding to the precision metal part from an existing metal heat treatment process database, the set of process parameters including temperature sampling point data of a high-temperature treatment stage above a preset temperature value and time series data of a low-temperature holding stage below a preset temperature value; Based on the spatial position information of each temperature sampling point and the temperature variation of adjacent points in the temperature sampling point data, the temperature sampling points are expanded to generate temperature gradient distribution features; Extracting the critical duration and temperature fluctuation frequency of the phase transition critical point in the time series data to generate a time series phase transition feature; The temperature gradient distribution feature and the time series phase change feature are spliced ​​together to generate a thermal history feature vector, wherein the number of dimensions of the thermal history feature vector is proportional to the number of types of material components of the precision metal part.

6. The method according to claim 1, characterized in that Based on the fusion processing result, defect detection data is generated, wherein the defect detection data includes three-dimensional position information of grain boundary cracks and prediction data of crack propagation direction, including: Converting the detection position corresponding to the characteristic dimension data related to the spatial distribution of the grain boundary crack in the fusion processing result into a three-dimensional coordinate value in a preset physical space coordinate system to generate spatial coordinate data; Identifying coordinates of a numerical mutation region and a target gradual change region of the spatial coordinate data to generate an initial crack position set, wherein the initial crack position set includes crack tip positions and branch node positions; Calculating the stress field distribution vector at the crack tip position based on the temperature gradient distribution feature in the thermal history feature vector and the sub-surface scattering spot distribution gradient feature in the multimodal image feature set; Binding the coordinates of the crack tip position with the corresponding stress field distribution vector to generate bound data, and combining the change rate of the characteristic dimension value of the connection path between the branch node position and the adjacent crack tip position to generate primary defect data; Topological connection verification is performed on the primary defect data to generate defect detection data.

7. A precision metal parts defect detection system based on a multimodal large model, characterized in that: include: An acquisition module is used to acquire laser scattering image data of precision metal parts in an electromagnetic shielding environment; An analysis module, configured to perform multi-scale spatial analysis on the laser scattering image data to extract a multimodal image feature set including the grain boundary topology of the surface of the precision metal part and the scattering spot distribution gradient characteristics of the subsurface; a conversion module, configured to convert the process parameters corresponding to the precision metal part in an existing metal heat treatment process database into a thermal history feature vector, wherein the thermal history feature vector includes a temperature gradient distribution feature and a time series phase change feature; A fusion module, configured to fuse the thermal history feature vector and the multimodal image feature set using a multimodal large model to obtain a fusion result; A generation module, configured to generate defect detection data based on the fusion processing result, wherein the defect detection data includes three-dimensional position information of grain boundary cracks and prediction data of crack propagation direction; The thermal history feature vector and the multimodal image feature set are fused using a multimodal large model to obtain a fusion processing result, including: Based on the material property parameters of the precision metal part, a weight control parameter set is generated, and the weight control parameter set includes a thermal history weight vector and an image weight vector; in the thermal history feature encoding channel of the multimodal large model, the thermal history feature vector is mapped according to the thermal history weight vector to generate a process feature encoding sequence with temperature gradient correlation characteristics; in the image feature encoding channel of the multimodal large model, a sliding window processing is performed on the multimodal image feature set according to the image weight vector to generate an image feature encoding sequence with grain boundary topology correlation characteristics; the process feature encoding sequence and the image feature encoding sequence are weighted according to the weight control parameter set to generate a fused encoding tensor; the fused encoding tensor is screened to generate a fusion processing result; Based on the fusion processing results, defect detection data is generated, including: The fusion processing results are input into a defect detection decoder, which includes a three-dimensional space reconstruction module and a crack propagation prediction module. The grain boundary cracks are located at the voxel level by the three-dimensional space reconstruction module to generate an initial crack position set. The stress field distribution vector at the crack tip position is analyzed by the crack propagation prediction module, and combined with the temperature gradient distribution characteristics in the thermal history feature vector, defect detection data with a spatial topological relationship is generated.

8. A computing device, characterized in that It includes a processing component and a storage component; the storage component stores one or more computer instructions; the one or more computer instructions are used to be called and executed by the processing component to implement a precision metal part defect detection method based on a multimodal large model as described in any one of claims 1 to 6.

9. A computer storage medium, characterized in that A computer program is stored, and when the computer program is executed by a computer, a precision metal part defect detection method based on a multimodal large model as described in any one of claims 1 to 6 is implemented.

Images