Quality evaluation method for industrial detection of tiny defects in metal part

Abstract

The invention discloses an industrial detection metal part internal tiny defect quality evaluation method, and relates to the technical field of metal part quality detection, and the method comprises the steps: employing a Peephole-LSTM model to carry out the dynamic weight distribution of the credibility coefficient of multi-modal data, carrying out the fusion, generating a defect feature pattern of a metal part, and carrying out the calculation of the defect feature pattern of the metal part; carrying out local CT scanning based on the defect feature map of the metal part to obtain metal part defect reconstruction data, calculating the equivalent diameter, the maximum sectional area and the volume of the metal part defect based on the metal part defect reconstruction data, obtaining defect quantization parameters, and calculating the defect quality score of the metal part according to the defect quantization parameters. And dividing quality grade intervals for comparison based on the defect quality score of the metal part, and obtaining a quality evaluation result of the metal part. According to the metal part quality evaluation method, the safety and reliability of the high-precision advanced equipment metal part in actual service are effectively guaranteed.

Description

Technical Field

[0001] This invention relates to the field of metal parts quality inspection technology, and in particular to an industrial method for evaluating the quality of minute internal defects in metal parts. Background Technology

[0002] With the rapid development of modern industrial manufacturing towards high precision and high reliability, metal parts are increasingly widely used in aerospace, rail transportation, energy and power and other fields. Non-destructive testing and quality evaluation of their internal minute defects have become a core technical link to ensure structural safety and product reliability. In recent years, with the advancement of multi-sensor fusion technology and digital signal processing algorithms, multimodal collaborative detection has gradually become a research hotspot. By simultaneously acquiring multi-source data such as ultrasonic A-scan waveforms, eddy current impedance signals and X-ray images, information complementarity can be achieved in the spatial, temporal and physical mechanism dimensions, improving defect detection rate and positioning accuracy. At the same time, the introduction of deep learning and artificial intelligence technologies has further promoted the development of automated and intelligent detection.

[0003] Nevertheless, existing methods for evaluating the quality of metal parts often set thresholds based solely on geometric dimensions, neglecting the fundamental differences in mechanical behavior and failure risk among defect types (such as cracks, porosity, and inclusions). This leads to a disconnect between the quality evaluation results of metal parts and their actual service performance, thereby weakening their scientific validity and interpretability. Summary of the Invention

[0004] In view of the aforementioned existing problems, the present invention is proposed.

[0005] Therefore, this invention provides a method for evaluating the quality of internal minute defects in metal parts in industrial applications, which solves the problem of disjointed quality evaluation results caused by neglecting defect types.

[0006] To solve the above-mentioned technical problems, the present invention provides the following technical solution: This invention provides a method for evaluating the quality of internal minute defects in metal parts in industrial applications, comprising: Collect multimodal data of the target metal parts, identify defects, and determine the suspected defect areas of the metal parts; Calculate the local signal-to-noise ratio of multimodal data in the suspected defect area of ​​metal parts, and generate the credibility coefficient of multimodal data; The Peephole-LSTM model is used to dynamically assign weights to the credibility coefficients of multimodal data and then fuse them to generate defect feature maps of metal parts. Local CT scanning is performed based on the defect feature map of the metal part to obtain defect reconstruction data of the metal part; Based on the reconstructed data of metal part defects, the equivalent diameter, maximum cross-sectional area and volume of the defects in the metal parts are calculated to obtain the defect quantification parameters; The defect quality score of the metal parts is calculated based on the defect quantification parameters. Based on the defect quality score of the metal parts, the quality level range is divided for comparison to obtain the quality evaluation result of the metal parts.

[0007] As a preferred embodiment of the industrial inspection method for evaluating the quality of minute internal defects in metal parts according to the present invention, the multimodal data includes ultrasonic A-scan waveforms, eddy current impedance signals, and X-ray images.

[0008] As a preferred embodiment of the industrial inspection method for evaluating the quality of minute internal defects in metal parts according to the present invention, the step of defect judgment and obtaining suspected defect areas in the metal parts specifically involves: When the signal amplitude of the ultrasonic A-scan waveform in a metal part exceeds N times the standard deviation of the mean background noise, it is identified as a suspected defect area of ​​the ultrasonic A-scan waveform. The point where the rate of change of the real part of the impedance in the eddy current impedance signal exceeds the rate of change threshold is taken as the suspected defect area in eddy current detection. Regions in X-ray images with gray values ​​below a gray threshold are considered as suspected defect areas in the X-ray images. By integrating the coordinates of suspected defect areas from multimodal data, the suspected defect areas of metal parts are obtained through spatial alignment.

[0009] As a preferred embodiment of the industrial inspection method for evaluating the quality of minute internal defects in metal parts according to the present invention, the step of calculating the local signal-to-noise ratio of multimodal data in the suspected defect area of ​​the metal part to generate the reliability coefficient of the multimodal data specifically involves: Select a background region with the same size as the suspected defect region of the metal part outside the suspected defect region of the metal part, and calculate the noise power in the background region. Based on noise power, the local signal-to-noise ratio of the ultrasonic A-scan waveform, eddy current impedance signal, and X-ray image in the suspected defect area of ​​the metal part is calculated sequentially, and then normalized to obtain the reliability coefficient of the multimodal data.

[0010] As a preferred embodiment of the industrial inspection method for evaluating the quality of minute internal defects in metal parts according to the present invention, the method of dynamically assigning weights to the reliability coefficients of multimodal data using the Peephole-LSTM model specifically involves: The confidence coefficients of ultrasound A-scan waveforms, eddy current impedance signals and X-ray images in multimodal data are constructed into an input sequence in chronological order and input into the Peephole-LSTM model. The Peephole-LSTM model directly accesses the cell state through forget gate, input gate and output gate to calculate temporal features and outputs the dynamic fusion weights of multimodal data. The dynamic fusion weights of the multimodal data include the dynamic fusion weights of ultrasound A-scan waveforms, eddy current impedance signals, and X-ray images.

[0011] As a preferred embodiment of the industrial inspection method for evaluating the quality of minute internal defects in metal parts according to the present invention, the generation of the defect feature map of the metal part refers to using dynamic fusion weighting of ultrasonic A-scan waveform, dynamic fusion weighting of eddy current impedance signal, and dynamic fusion weighting of X-ray image to sequentially weight the signal amplitude of ultrasonic A-scan waveform, the real part of impedance of eddy current impedance signal, and the gray value in X-ray image in the suspected defect area of ​​the metal part, and then performing pixel-level superposition and fusion of the weighted multimodal data in the spatial domain to generate the defect feature map of the metal part.

[0012] As a preferred embodiment of the industrial inspection method for evaluating the quality of minute internal defects in metal parts according to the present invention, the step of performing local CT scanning based on the defect feature map of the metal part to obtain defect reconstruction data of the metal part specifically involves... Based on the defect feature map of metal parts, CT is used to perform local projection scanning of metal parts to obtain multi-angle X-ray projection data; A filtered back-projection algorithm is used to perform three-dimensional reconstruction of multi-angle X-ray projection data. During the reconstruction process, total variational regularization constraints are added to suppress artifacts and generate reconstructed data of defects in metal parts.

[0013] As a preferred embodiment of the industrial inspection method for evaluating the quality of minute internal defects in metal parts according to the present invention, the method involves: calculating the equivalent diameter, maximum cross-sectional area, and volume of the defects based on the reconstructed data of the metal parts, and obtaining the defect quantification parameters. Specifically, Extract the defect voxel distribution matrix from the defect reconstruction data of metal parts; Based on the defect voxel distribution matrix, the maximum projected area of ​​all cross sections in the defect of the metal part is calculated as the maximum cross-sectional area of ​​the defect, and the equivalent diameter of the defect is calculated based on the maximum cross-sectional area of ​​the defect. The defect volume is calculated by spatial integration of the defect voxel distribution matrix. The equivalent diameter of the defect, the maximum cross-sectional area of ​​the defect, and the volume of the defect are used as parameters for defect quantification.

[0014] As a preferred embodiment of the industrial inspection method for evaluating the quality of minute internal defects in metal parts according to the present invention, the step of calculating the defect quality score of the metal part based on defect quantification parameters specifically involves: Convolutional neural networks are used to classify the morphological features of metal part defect reconstruction data and obtain the defect type risk coefficient. After normalizing the defect quantification parameters, a weighted sum is calculated with the defect type risk coefficient to generate a defect quality score for the metal part.

[0015] As a preferred embodiment of the industrial inspection method for evaluating the quality of minute internal defects in metal parts according to the present invention, the step of dividing quality grade intervals based on the defect quality score of the metal parts for comparison and obtaining the quality evaluation result of the metal parts specifically involves: Based on the defect quality score of metal parts, Y quality level intervals are divided. Each quality level interval corresponds to a different quality evaluation result. The defect quality score of metal parts is mapped to the corresponding quality level interval, and the corresponding quality evaluation result of metal parts is output.

[0016] The beneficial effects of this invention are as follows: By collecting multimodal data and combining local signal-to-noise ratio analysis with Peephole-LSTM dynamic weight allocation, a high-confidence defect feature map is generated, enabling precise localization and enhanced characterization of minute internal defects in metal parts. Then, based on the feature map, local CT scans are guided, and total variational regularization is introduced to suppress artifacts, obtaining high-precision three-dimensional defect reconstruction data. Finally, a convolutional neural network is used to identify defect types and a risk coefficient based on failure probability is introduced. Weighted scoring is performed using quantitative parameters such as the equivalent diameter, maximum cross-sectional area, and volume of the defect. This overcomes the limitations of previous methods that relied solely on geometric size criteria, achieving a leap from "size evaluation" to a comprehensive quality assessment of "type-size-risk," improving the accuracy, scientific rigor, and engineering practicality of the detection results, and effectively ensuring the safety and reliability of high-precision equipment metal components in actual service. Attached Figure Description

[0017] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the following description of the embodiments will be briefly introduced. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0018] Figure 1 A flowchart for a method of evaluating the quality of minute internal defects in industrial metal parts.

[0019] Figure 2 A flowchart for generating defect reconstruction data for metal parts.

[0020] Figure 3 A flowchart for generating dynamic weights.

[0021] Figure 4 A flowchart for generating defect quality scores for metal parts. Detailed Implementation

[0022] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings.

[0023] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein, and those skilled in the art can make similar extensions without departing from the spirit of the invention. Therefore, the invention is not limited to the specific embodiments disclosed below.

[0024] Secondly, the term "one embodiment" or "embodiment" as used herein refers to a specific feature, structure, or characteristic that may be included in at least one implementation of the present invention. The phrase "in one embodiment" appearing in different places in this specification does not necessarily refer to the same embodiment, nor is it a single or selective embodiment that is mutually exclusive with other embodiments.

[0025] Reference Figures 1-4 This is one embodiment of the present invention, which provides a method for evaluating the quality of internal minute defects in metal parts in industrial applications, comprising the following steps: S1. Collect multimodal data of the target metal part, perform defect judgment, and obtain the suspected defect area of ​​the metal part; S1.1 When acquiring the ultrasonic A-scan waveform of the target metal part, an ultrasonic probe is moved across the surface of the target metal part, emitting ultrasonic waves and receiving the reflected echoes from inside the target metal part, recording the signal amplitude. Simultaneously, eddy current impedance signals are acquired, and the changes in the electromagnetic properties of the target metal part are measured using an eddy current probe, recording the real part of the impedance. At the same time, an X-ray machine is used to penetrate the target metal part, and a flat panel detector receives the X-ray signals returning from the target metal part, generating an X-ray image. The ultrasonic A-scan waveform, eddy current impedance signal, and X-ray image are simultaneously recorded in a storage device.

[0026] S1.2 When the signal amplitude of the ultrasonic A-scan waveform in a metal part exceeds N times the standard deviation of the mean background noise, it is determined to be a suspected defect area of ​​the ultrasonic A-scan waveform. It should be noted that an ultrasonic A-scan waveform is acquired using a brand-new, defect-free metal part, and the arithmetic mean and standard deviation of the corresponding signal amplitude are calculated as the mean and standard deviation of the background noise. Based on the mean and standard deviation of the background noise, the defect judgment threshold is set to the mean background noise plus N times (e.g., 3 times) the standard deviation of the background noise. Based on the ultrasonic A-scan waveform of the target metal part in step S1.1, when the signal amplitude of the ultrasonic A-scan waveform at any position of the target metal part exceeds the mean background noise plus 3 times the standard deviation of the background noise, the coordinates of this position are marked as the suspected defect area of ​​the ultrasonic A-scan waveform.

[0027] S1.3 The abrupt change point where the rate of change of the real part of the impedance in the eddy current impedance signal exceeds the rate of change threshold is taken as the suspected defect area of ​​eddy current detection. It should be noted that the rate of change of the real part of the impedance is calculated by spatial differentiation of the eddy current impedance signal of the target metal part, and expressed by the formula: ; in, This represents the rate of change of the real part of the impedance. Indicates the first The real part of the eddy current impedance at +1 sampling points Indicates the first The real part of the eddy current impedance at each sampling point This indicates the spatial distance between adjacent sampling points (determined by the eddy current probe's movement step resolution, for example: 0.1 mm / step). Based on the rate of change of the real part of the eddy current impedance signal, a rate of change threshold is set. When the absolute value of the rate of change of the real part of the eddy current impedance signal at any position of the target metal part exceeds the rate of change threshold (e.g., 10%), it is marked as a mutation point. The spatial coordinate region where all mutation points are located is taken as the suspected defect region of eddy current detection.

[0028] S1.4. Regions in X-ray images with gray values ​​below the gray threshold are considered as suspected defect regions in the X-ray images. X-ray images of brand-new, defect-free metal parts are acquired, and the arithmetic mean of the grayscale values ​​in the X-ray images is calculated as the material's average grayscale. Based on this average grayscale, a grayscale threshold is set as the product of the material's average grayscale and 20%. Using the X-ray images of the target metal part, the grayscale values ​​of each pixel in the X-ray images are compared with the grayscale threshold. Pixels with grayscale values ​​below the threshold are marked. Morphological dilation is performed on all pixels below the threshold to form connected regions. The bounding rectangles of all connected regions are considered as suspected defect areas in the X-ray images. The selection of 20% is based on a comprehensive optimization of defect physical characteristics and multimodal collaboration, ensuring a high detection rate while effectively controlling the risk of false positives.

[0029] S1.5. Based on the coordinates of suspected defect areas from the integrated multimodal data, the suspected defect areas of metal parts are obtained through spatial alignment. Furthermore, the coordinates of the suspected defect areas from the ultrasonic A-scan waveform, the eddy current detection, and the X-ray image are input into a spatial registration algorithm. Through feature point matching and coordinate transformation matrix calculation, the coordinates of the suspected defect areas from the multimodal data are unified into the same coordinate system of the metal part. The logical union of the suspected defect areas from the ultrasonic A-scan waveform, the eddy current detection, and the X-ray image after spatial coordinate alignment is taken to generate the suspected defect areas of the metal part.

[0030] S2. Calculate the local signal-to-noise ratio of multimodal data in the suspected defect area of ​​metal parts, and generate the credibility coefficient of multimodal data; S2.1 Select a background area with the same size as the suspected defect area of ​​the metal part outside the suspected defect area, and calculate the noise power in the background area; It should be noted that a region of the same size as the suspected defect area of ​​the metal part is selected as the background region (i.e., the background region of the same size) around the suspected defect area of ​​the metal part. The signal amplitude of the ultrasonic A-scan waveform, the real part of the impedance of the eddy current impedance signal, and the gray value of the X-ray image are extracted from the background region of the same size. The variance of the signal amplitude of the ultrasonic A-scan waveform is calculated as the ultrasonic noise power, the variance of the real part of the impedance of the eddy current impedance signal is calculated as the eddy current noise power, and the variance of the gray value of the X-ray image is calculated as the X-ray noise power.

[0031] S2.2 Based on noise power, calculate the local signal-to-noise ratio of the ultrasonic A-scan waveform, eddy current impedance signal and X-ray image in the suspected defect area of ​​the metal part in the multimodal data in turn, and perform normalization processing to obtain the credibility coefficient of the multimodal data. Furthermore, the signal amplitude of the ultrasonic A-scan waveform, the real part of the eddy current impedance signal, and the gray value of the X-ray image in the suspected defect area of ​​the metal part are successively averaged and squared to obtain the ultrasonic signal power, eddy current signal power, and X-ray signal power.

[0032] Based on the ultrasonic signal power, eddy current signal power, and X-ray signal power, and combined with the three noise powers from step S2.1, the local signal-to-noise ratio of the ultrasonic A-scan waveform, the local signal-to-noise ratio of the eddy current impedance signal, and the local signal-to-noise ratio of the X-ray image are calculated sequentially using the same formula. For example, the local signal-to-noise ratio of the ultrasonic A-scan waveform is calculated using the following formula: ; in, This indicates the local signal-to-noise ratio of the ultrasound A-scan waveform. This indicates the ultrasonic signal power of the A-scan waveform within the suspected defect area of ​​the metal part. This indicates the total number of ultrasonic A-scan waveform sampling points within the suspected defect area of ​​the metal part. This indicates the first position of the ultrasonic A-scan waveform within the suspected defect area of ​​the metal part. The signal amplitude at each sampling point Indicates ultrasonic noise power; The local signal-to-noise ratios of the ultrasound A-scan waveform, eddy current impedance signal, and X-ray image were subjected to minimum-maximum normalization to obtain the reliability coefficients of the ultrasound A-scan waveform, eddy current impedance signal, and X-ray image, which are the reliability coefficients of the multimodal data.

[0033] S3. The Peephole-LSTM model is used to dynamically assign weights to the credibility coefficients of multimodal data and then fuse them to generate defect feature maps of metal parts. S3.1 Construct an input sequence by arranging the confidence coefficients of ultrasound A-scan waveforms, eddy current impedance signals and X-ray images in multimodal data in chronological order, and input it into the Peephole-LSTM model. The Peephole-LSTM model directly accesses the cell state through the forget gate, input gate and output gate to calculate the temporal features and outputs the dynamic fusion weights of the multimodal data. It should be noted that a metal part of the same specifications as the target metal part is prepared. At the same time, the historical ultrasound A-scan waveform confidence coefficient, eddy current impedance signal confidence coefficient, and X-ray image confidence coefficient of the same specification metal part are collected and divided into training set and validation set. The training set is input into the Peephole-LSTM model for training. During the training process, the gradient is calculated by backpropagation through the time algorithm. The weight matrix and bias vector of the Peephole-LSTM model are iteratively updated by the adaptive moment estimation optimizer to minimize the mean square error loss between the prediction fusion weight and the true value weight. Training is stopped when the validation set loss function no longer decreases for 10 consecutive training cycles, and the trained Peephole-LSTM model is obtained.

[0034] The reliability coefficients of ultrasound A-scan waveform, eddy current impedance signal, and X-ray image are arranged in time-stamp order to construct a time series as input data, which is then input into the trained Peephole-LSTM model. The forget gate, input gate, and output gate of the Peephole-LSTM model simultaneously receive the input data, access the temporal dependencies of the reliability coefficients of ultrasound A-scan waveform, eddy current impedance signal, and X-ray image in the input data, and iteratively calculate the hidden state vector through a gating mechanism. The hidden state vector is then processed through the fully connected layer and Softmax function of the Peephole-LSTM model to output the dynamic fusion weights of ultrasound A-scan waveform, eddy current impedance signal, and X-ray image.

[0035] Using dynamic fusion weights for ultrasonic A-scan waveforms, eddy current impedance signals, and X-ray images, the signal amplitude of the ultrasonic A-scan waveform, the real part of the impedance signal, and the gray value in the X-ray image within the suspected defect area of ​​the metal part are weighted sequentially. The weighted three modal data are then superimposed and fused at the pixel level in the spatial domain to generate a defect feature map of the metal part.

[0036] S4. Perform local CT scanning based on the defect feature map of the metal part to obtain defect reconstruction data of the metal part; S4.1 Defect feature map based on metal parts: Use CT to perform local projection scanning on metal parts to obtain multi-angle X-ray projection data; It should be noted that, based on the defect feature map of the metal part, the metal part is fixed on a rotating platform. The X-ray source and the flat panel detector are aligned with the area marked on the defect feature map as the center. The rotating platform is controlled to rotate around the area marked on the defect feature map from 0 degrees to 180 degrees at fixed angular intervals. At each rotation angle position, the X-ray source emits a cone-shaped X-ray beam that penetrates the metal part, and the flat panel detector receives the attenuated X-ray signal and records it as a two-dimensional projection image, ultimately obtaining multi-angle X-ray projection data.

[0037] S4.2. A filtered back projection algorithm is used to perform three-dimensional reconstruction of multi-angle X-ray projection data. During the reconstruction process, total variational regularization constraints are added to suppress artifacts and generate reconstructed data of metal part defects. Furthermore, each two-dimensional projection image in the multi-angle X-ray projection data is convolved with a Ram-Lak filter to obtain filtered projection data. The filtered projection data is then back-projected along the projection path. The three-dimensional voxel matrix is ​​reconstructed by accumulating the back-projection results from all angles. A total variational regularization constraint is added as a penalty term during the back-projection process to suppress noise and artifacts in the three-dimensional voxel matrix. Finally, a high-quality three-dimensional voxel matrix is ​​obtained, which is the defect reconstruction data of the metal part.

[0038] S5. Based on the reconstructed data of metal part defects, calculate the equivalent diameter, maximum cross-sectional area and volume of the metal part defects, and obtain the defect quantification parameters; S5.1 Extract the defect voxel distribution matrix from the defect reconstruction data of metal parts; It should be noted that in the metal part defect reconstruction data, by calculating the arithmetic mean and standard deviation of the gray values ​​of the metal part matrix material region, the defect segmentation threshold is set as the average gray value of the matrix material minus 3 times the standard deviation. All voxels in the metal part defect reconstruction data are traversed, and voxels with gray values ​​lower than the defect segmentation threshold are marked as defect voxels, while voxels with gray values ​​higher than or equal to the defect segmentation threshold are marked as matrix material voxels. All voxel coordinates in the metal part defect reconstruction data are traversed, and the marking status of each voxel is checked. The X, Y, and Z coordinate values ​​of all voxels marked as defects are extracted. The extracted defect voxel coordinates are mapped to a matrix of the same size as the metal part defect reconstruction data according to the three-dimensional spatial grid position. Values ​​of 1 are filled in the matrix positions corresponding to defect voxels, and values ​​of 0 are filled in the matrix positions corresponding to non-defect voxels, generating a defect voxel distribution matrix in binary form representing the spatial distribution of defects.

[0039] S5.2 Based on the defect voxel distribution matrix, calculate the maximum projected area of ​​all cross sections in the defect of the metal part as the maximum cross-sectional area of ​​the defect, and calculate the equivalent diameter of the defect based on the maximum cross-sectional area of ​​the defect. It should be noted that, along the X, Y, and Z axes of the three-dimensional voxel distribution matrix of defects in the metal part, slices are made layer by layer. The number of defect voxels with a value of 1 on each slice plane is calculated and multiplied by the physical area of ​​each individual defect voxel to obtain the defect projected area of ​​each slice plane. The maximum value of the defect projected areas of all slice planes is taken as the maximum cross-sectional area of ​​the defect in the metal part. The maximum cross-sectional area of ​​the defect in the metal part is then substituted into the formula for the area of ​​a circle to calculate the equivalent diameter of the defect, which is expressed by the formula: ; in, Indicates the equivalent diameter of the defect. Represents pi (π). This indicates the maximum cross-sectional area of ​​a defect in a metal part.

[0040] S5.3 Calculate the defect volume by spatial integration of the defect voxel distribution matrix; It should be noted that the number of defect voxels with a value of 1 in the defect voxel distribution matrix of the metal part is counted, and the defect volume of the metal part is obtained by multiplying the number of voxels by the physical volume of the defect voxels. The physical volume of a single defect voxel is equal to the product of the defect voxel dimensions in the X-axis direction, the Y-axis direction, and the Z-axis direction. The equivalent diameter of the defect, the maximum cross-sectional area of ​​the defect, and the defect volume are used as defect quantification parameters.

[0041] S6. Calculate the defect quality score of the metal parts based on the defect quantification parameters, divide the quality level range based on the defect quality score of the metal parts for comparison, and obtain the quality evaluation result of the metal parts. S6.1 Use a convolutional neural network to classify the morphological features of the reconstructed defect data of metal parts and obtain the defect type risk coefficient; It should be noted that historical metal part defect reconstruction data and corresponding defect type labels are collected and divided into training and validation sets. The training set is input into a convolutional neural network for training. During training, the backpropagation algorithm is used to calculate the cross-entropy loss between the predicted output of the convolutional neural network and the defect type label. The convolutional layer weights and fully connected layer parameters of the convolutional neural network are iteratively updated through the gradient descent optimizer to minimize the cross-entropy loss function. Training is stopped when the accuracy of the validation set no longer improves for 10 consecutive training cycles, and the trained convolutional neural network is obtained.

[0042] Historical data on metal part failures were collected, and the probabilities of three defect types—cracks, porosity, and inclusions—leading to failure were statistically analyzed. These probabilities were then linearly normalized and mapped to a risk coefficient range of 0.4 to 0.9. For example, cracks, with the highest failure probability, were assigned a risk coefficient of 0.9; porosity, with the lowest, was assigned a risk coefficient of 0.4; and inclusions, with a moderate failure probability, were assigned a risk coefficient of 0.5. Finally, a correspondence table between defect types and risk coefficients was established—a defect type risk coefficient mapping table. The risk coefficient range of 0.4 to 0.9 accurately reflects the severity of defects and facilitates integration into existing quality evaluation systems.

[0043] The reconstructed data of metal part defects is input into the trained convolutional neural network. The convolutional neural network extracts local features of defect morphology through multiple convolutional layers, reduces feature dimensionality through pooling layers, maps features to classification space through fully connected layers, and outputs defect type probability distribution. Based on the defect type probability distribution, the defect type risk coefficient mapping table is queried to obtain the corresponding defect type risk coefficient.

[0044] S6.2 After normalizing the defect quantification parameters, a weighted sum is calculated with the defect type risk coefficient to generate a defect quality score for the metal part. It should be noted that the equivalent diameter, maximum cross-sectional area, and volume of the defect are subjected to minimum-maximum normalization. The normalized equivalent diameter, maximum cross-sectional area, and volume are then weighted with the defect type risk coefficient distribution (e.g., the normalized equivalent diameter has a weight of 0.4, the normalized maximum cross-sectional area has a weight of 0.3, the normalized volume has a weight of 0.2, and the defect type risk coefficient has a weight of 0.1) and summed to obtain the defect quality score for the metal part, expressed by the formula: ; in, This indicates the quality score for defects in metal parts. This represents the normalized defect equivalent diameter. This represents the maximum cross-sectional area of ​​the normalized defect. Represents the normalized defect volume. Indicates the risk coefficient of the trap type. The weights representing the normalized defect equivalent diameter The weights representing the maximum cross-sectional area of ​​the normalized defect are: The weights representing the normalized defect volume, The weight of the risk coefficient for each trap type; S6.3 Based on the defect quality score of metal parts, divide the quality into Y quality level intervals, each quality level interval corresponding to a different quality evaluation result; It should be noted that, based on the quality score of metal parts defects, three quality level ranges are preset. For example, when the quality score of metal parts defects is greater than or equal to 0.8, it is classified into the first quality level range and the corresponding quality evaluation result is "qualified"; when the quality score of metal parts defects is greater than or equal to 0.6 and less than 0.8, it is classified into the second quality level range and the corresponding quality evaluation result is "critical and requires re-inspection"; when the quality score of metal parts defects is less than 0.6, it is classified into the third quality level range and the corresponding quality evaluation result is "unqualified". The quality score of metal parts defects is mapped to the corresponding quality level range, and the corresponding quality evaluation result of metal parts is output.

[0045] This embodiment also provides a computer device applicable to the industrial inspection of a method for evaluating the quality of minute internal defects in metal parts, comprising: a memory and a processor; the memory is used to store computer-executable instructions, and the processor is used to execute the computer-executable instructions to implement the industrial inspection of a method for evaluating the quality of minute internal defects in metal parts as proposed in the above embodiment.

[0046] The computer device can be a terminal, comprising a processor, memory, communication interface, display screen, and input devices connected via a system bus. The processor provides computing and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system and computer programs. The internal memory provides an environment for the operation of the operating system and computer programs stored in the non-volatile storage media. The communication interface is used for wired or wireless communication with external terminals; wireless communication can be achieved through Wi-Fi, carrier networks, NFC (Near Field Communication), or other technologies. The display screen can be an LCD screen or an e-ink screen. The input devices can be a touch layer covering the display screen, buttons, a trackball, or a touchpad on the computer device's casing, or an external keyboard, touchpad, or mouse.

[0047] This embodiment also provides a storage medium storing a computer program. When executed by a processor, the program implements the quality evaluation method for detecting minute internal defects in metal parts as proposed in the above embodiments. The storage medium can be implemented by any type of volatile or non-volatile storage 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 Red-Only Memory (PROM), Read-Only Memory (ROM), magnetic storage, flash memory, magnetic disk, or optical disk.

[0048] In summary, this invention achieves precise localization and enhanced characterization of minute internal defects in metal parts by: collecting multimodal data, combining local signal-to-noise ratio analysis and Peephole-LSTM dynamic weight allocation to generate high-confidence defect feature maps; guiding local CT scans based on feature maps and introducing total variational regularization to suppress artifacts, thereby obtaining high-precision three-dimensional defect reconstruction data; and finally using convolutional neural networks to identify defect types and introducing risk coefficients based on failure probability, combined with quantitative parameters such as the equivalent diameter, maximum cross-sectional area, and volume of the defect for weighted scoring. This breakthrough overcomes the limitations of previous methods that relied solely on geometric size criteria, achieving a leap from "size evaluation" to a comprehensive quality assessment of "type-size-risk," improving the accuracy, scientific rigor, and engineering practicality of the detection results, and effectively ensuring the safety and reliability of high-precision equipment metal components in actual service.

[0049] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A method for evaluating the quality of minute internal defects in industrial metal parts, characterized in that: include, Collect multimodal data of the target metal parts, identify defects, and determine the suspected defect areas of the metal parts; Calculate the local signal-to-noise ratio of multimodal data in the suspected defect area of ​​metal parts, and generate the credibility coefficient of multimodal data; The Peephole-LSTM model is used to dynamically assign weights to the credibility coefficients of multimodal data and then fuse them to generate defect feature maps of metal parts. Local CT scanning is performed based on the defect feature map of the metal part to obtain defect reconstruction data of the metal part; Based on the reconstructed data of metal part defects, the equivalent diameter, maximum cross-sectional area and volume of the defects in the metal parts are calculated to obtain the defect quantification parameters; The defect quality score of the metal parts is calculated based on the defect quantification parameters. Based on the defect quality score of the metal parts, the quality level range is divided for comparison to obtain the quality evaluation result of the metal parts.

2. The method for evaluating the quality of internal minute defects in industrial metal parts as described in claim 1, characterized in that: The multimodal data includes ultrasound A-scan waveforms, eddy current impedance signals, and X-ray images.

3. The method for evaluating the quality of internal minute defects in metal parts as described in claim 2, characterized in that: The defect assessment involves identifying suspected defect areas in the metal parts. When the signal amplitude of the ultrasonic A-scan waveform in a metal part exceeds N times the standard deviation of the mean background noise, it is identified as a suspected defect area of ​​the ultrasonic A-scan waveform. The point where the rate of change of the real part of the impedance in the eddy current impedance signal exceeds the rate of change threshold is taken as the suspected defect area in eddy current detection. Regions in X-ray images with gray values ​​below a gray threshold are considered as suspected defect areas in the X-ray images. By integrating the coordinates of suspected defect areas from multimodal data, the suspected defect areas of metal parts are obtained through spatial alignment.

4. The method for evaluating the quality of internal minute defects in metal parts as described in claim 3, characterized in that: The process of calculating the local signal-to-noise ratio of multimodal data in the suspected defect area of ​​a metal part and generating a reliability coefficient for the multimodal data specifically involves... Select a background region with the same size as the suspected defect region of the metal part outside the suspected defect region of the metal part, and calculate the noise power in the background region. Based on noise power, the local signal-to-noise ratio of the ultrasonic A-scan waveform, eddy current impedance signal, and X-ray image in the suspected defect area of ​​the metal part is calculated sequentially, and then normalized to obtain the reliability coefficient of the multimodal data.

5. The method for evaluating the quality of internal minute defects in metal parts as described in claim 4, characterized in that: The method of dynamically assigning weights to the credibility coefficients of multimodal data using the Peephole-LSTM model is specifically as follows: The confidence coefficients of ultrasound A-scan waveforms, eddy current impedance signals and X-ray images in multimodal data are constructed into an input sequence in chronological order and input into the Peephole-LSTM model. The Peephole-LSTM model directly accesses the cell state through forget gate, input gate and output gate to calculate temporal features and outputs the dynamic fusion weights of multimodal data. The dynamic fusion weights of the multimodal data include the dynamic fusion weights of ultrasound A-scan waveforms, eddy current impedance signals, and X-ray images.

6. The method for evaluating the quality of internal minute defects in industrial metal parts as described in claim 5, characterized in that: The process of generating a defect feature map of a metal part involves using dynamic fusion weights for ultrasonic A-scan waveforms, eddy current impedance signals, and X-ray images. These weights are applied sequentially to the signal amplitude of the ultrasonic A-scan waveform, the real part of the impedance signal, and the grayscale value in the X-ray image within the suspected defect area of ​​the metal part. The weighted multimodal data are then superimposed and fused at the pixel level in the spatial domain to generate the defect feature map of the metal part.

7. The method for evaluating the quality of internal minute defects in industrial metal parts as described in claim 6, characterized in that: The method involves performing local CT scanning based on the defect feature map of the metal part to obtain defect reconstruction data of the metal part. Specifically, Based on the defect feature map of metal parts, CT is used to perform local projection scanning of metal parts to obtain multi-angle X-ray projection data; A filtered back-projection algorithm is used to perform three-dimensional reconstruction of multi-angle X-ray projection data. During the reconstruction process, total variational regularization constraints are added to suppress artifacts and generate reconstructed data of defects in metal parts.

8. The method for evaluating the quality of internal minute defects in industrial metal parts as described in claim 7, characterized in that: Based on the reconstructed data of metal part defects, the equivalent diameter, maximum cross-sectional area, and volume of the defects are calculated to obtain the defect quantification parameters. Specifically, Extract the defect voxel distribution matrix from the defect reconstruction data of metal parts; Based on the defect voxel distribution matrix, the maximum projected area of ​​all cross sections in the defect of the metal part is calculated as the maximum cross-sectional area of ​​the defect, and the equivalent diameter of the defect is calculated based on the maximum cross-sectional area of ​​the defect. The defect volume is calculated by spatial integration of the defect voxel distribution matrix. The equivalent diameter of the defect, the maximum cross-sectional area of ​​the defect, and the volume of the defect are used as parameters for defect quantification.

9. The method for evaluating the quality of internal minute defects in metal parts as described in claim 8, characterized in that: The calculation of the defect quality score for metal parts based on defect quantification parameters specifically involves... Convolutional neural networks are used to classify the morphological features of metal part defect reconstruction data and obtain the defect type risk coefficient. After normalizing the defect quantification parameters, a weighted sum is calculated with the defect type risk coefficient to generate a defect quality score for the metal part.

10. The method for evaluating the quality of internal minute defects in metal parts as described in claim 9, characterized in that: The process involves dividing metal parts into quality grade ranges based on defect quality scores, comparing these ranges, and obtaining the metal parts quality evaluation results. Specifically... Based on the defect quality score of metal parts, Y quality level intervals are divided. Each quality level interval corresponds to a different quality evaluation result. The defect quality score of metal parts is mapped to the corresponding quality level interval, and the corresponding quality evaluation result of metal parts is output.

Images