CN-121994921-A - Metal structural member defect detection method based on deep learning

CN121994921ACN 121994921 ACN121994921 ACN 121994921ACN-121994921-A

Abstract

The invention discloses a method for detecting defects of a metal structure based on deep learning, which relates to the technical field of defect detection, and comprises the steps of firstly configuring detection parameters of a phased array probe according to characteristic data of the metal structure, further obtaining original full matrix data of the metal structure, then preprocessing the original full matrix data of the metal structure, measuring actual sound velocity of the metal structure, obtaining sound velocity correction full-focus images of the metal structure based on the actual sound velocity of the metal structure and combining acoustic characteristic data of the metal structure, then adjusting sound beam incidence angles of the phased array probe, generating sound velocity correction full-focus images of a plurality of sound beam incidence angles of the metal structure, constructing a three-dimensional data stack of the metal structure, finally fusing the images, generating a comprehensive detection image of the metal structure, and carrying out defect detection according to the comprehensive detection image of the metal structure, thereby realizing high-precision and high-reliability detection of the defects of the metal structure.

Inventors

LI CHENGQUAN

Assignees

拓腾(镇江)工业有限公司

Dates

Publication Date: 20260508
Application Date: 20260130

Claims (10)

1. The method for detecting the defects of the metal structural part based on deep learning is characterized by comprising the following steps of: s1, configuring detection parameters of a phased array probe according to characteristic data of a metal structural member, and acquiring original full-matrix data of the metal structural member in a full-matrix capturing mode by using the phased array probe after parameter configuration; S2, preprocessing original full matrix data of the metal structural member to obtain acoustic characteristic data of the metal structural member, and further determining actual sound velocity of the metal structural member; S3, based on the actual sound velocity of the metal structural member, combining acoustic characteristic data of the metal structural member to obtain a sound velocity correction full-focusing image of the metal structural member; S4, adjusting the incidence angles of the sound beams of the phased array probe, generating sound velocity correction full-focusing images of a plurality of incidence angles of the sound beams of the metal structural member, and further constructing a three-dimensional data stack of the metal structural member; S5, fusing the images in the three-dimensional data stack of the metal structural member to generate a comprehensive detection image of the metal structural member, and further performing defect detection according to the comprehensive detection image of the metal structural member.
2. The method for detecting defects of a metal structural member based on deep learning of claim 1, wherein the method for configuring detection parameters of a phased array probe according to characteristic data of the metal structural member comprises the following specific steps: The characteristic data of the metal structural member comprises the material type, thickness, minimum resolution unit size, width of the area to be detected and theoretical sound velocity of the metal structural member; the detection parameters of the phased array probe comprise the number of array elements of the phased array probe and the pulse repetition frequency of the pulse generator; setting the number of array elements of the phased array probe according to the minimum resolution unit size of the metal structural member and the width of the area to be detected; the pulse repetition frequency range of the pulse generator is determined based on the theoretical sound velocity and thickness of the metal structural member.
3. The method for detecting the defects of the metal structural part based on the deep learning of claim 1, wherein the method for preprocessing the original full matrix data of the metal structural part is characterized by comprising the following specific steps: Carrying out signal band-pass filtering, direct current offset correction, envelope detection and dynamic range compression on all A-type display signals in original full-matrix data of the metal structural member to obtain acoustic characteristic data of the metal structural member; The acoustic characteristic data of the metal structural part comprise all the preprocessed A-type display signals and envelope signals corresponding to all the A-type display signals.
4. The method for detecting defects of a metal structural member based on deep learning of claim 1, wherein the method further comprises the following specific steps of: extracting and identifying a reference echo of the metal structural member in the acoustic characteristic data to obtain a real propagation duration data set of the reference echo of the metal structural member; obtaining a sound velocity search interval according to the material type of the metal structural member and theoretical sound velocity analysis; based on the real propagation time length data set and the sound speed search interval of the reference echo of the metal structural member, the actual sound speed of the metal structural member is obtained by adopting a step-by-step optimization method.
5. The method for detecting defects of a metal structural member based on deep learning of claim 4, wherein the sound velocity search interval is obtained by analyzing according to a material type and a theoretical sound velocity of the metal structural member, and comprises the following specific processes: Obtaining a sound velocity adjustment proportionality coefficient of the metal structural member according to the material type of the metal structural member, extracting the sound velocity adjustment proportionality coefficient of the metal structural member corresponding to the material type of the metal structural member from a defect detection database, and recording the product of the theoretical sound velocity of the metal structural member and the sound velocity adjustment proportionality coefficient as a sound velocity adjustment value; Setting a minimum value of sound velocity And the maximum value of sound velocity Taking the theoretical sound velocity of the metal structural member as a reference, and taking the sound velocity minimum value as the sound velocity minimum value obtained by subtracting the sound velocity regulating value from the theoretical sound velocity of the metal structural member The theoretical sound velocity of the metal structural part plus the sound velocity regulating value is taken as the maximum value of the sound velocity Thereby constructing sound velocity search interval 。
6. The method for detecting the defects of the metal structural member based on the deep learning of claim 1, wherein the step-by-step optimization method is used for obtaining the actual sound velocity of the metal structural member, and the specific process is as follows: Setting a search step length according to the sound velocity search interval, setting the product of the width of the sound velocity search interval and a set first fixed proportional coefficient as a first search step length, and setting the product of the first search step length and a set second fixed proportional coefficient as a second search step length; And (3) analyzing in a sound velocity searching interval by adopting a step optimization method, wherein in the first stage, the minimum value of sound velocity in the sound velocity searching interval is taken as a starting point, traversing searching is carried out by adopting a set first searching step length, for each candidate sound velocity, the candidate sound velocity is input into a sound velocity optimization objective function to obtain an objective function value, when the relative change rate of the objective function value is smaller than a set first convergence threshold value in a plurality of iterations, judging that the sound velocity is close to an optimal solution, then transferring to a second stage, taking the candidate sound velocity found in the first stage as a starting point, carrying out fine searching by adopting a set second searching step length, when the change rate of the objective function value is smaller than a set second convergence threshold value, terminating the iteration process, and determining the candidate sound velocity when the iteration process is terminated in the second stage as the actual sound velocity of the metal structural member.
7. The method for detecting defects of a metal structural member based on deep learning of claim 1, wherein the method for obtaining sound velocity correction full-focus images of the metal structural member comprises the following specific steps: Setting pixel point coordinates P (x, z) and A-type display signals (i, j), calculating the sound path time of sound waves from a transmitting array element i to the pixel point P (x, z) and then to a receiving array element j based on the actual sound velocity of a metal structural member for the pixel point P (x, z) in an imaging area of the metal structural member, repeatedly carrying out the calculation on all the A-type display signals (i, j), and generating a sound path time data set for the pixel point P (x, z); Step a2, extracting envelope signals corresponding to all A-type display signals (i, j) from acoustic feature data, finding the sound path time calculated in the step a1 on a time axis of the envelope signals, and reading out signal envelope amplitude values at the sound path time, wherein the operation is repeatedly executed on all A-type display signals (i, j) to extract a signal envelope amplitude data set corresponding to pixel points P (x, z); step a3, carrying out incoherent superposition on signal envelope amplitudes in the signal envelope amplitude data set corresponding to the pixel point P (x, z) to obtain a total amplitude A (x, z) of the pixel point P (x, z); and a4, repeating the steps a1 to a3 for each pixel point of the imaging area of the metal structural member, and finally synthesizing a sound velocity correction full-focusing image of the metal structural member.
8. The method for detecting the defects of the metal structural part based on the deep learning of claim 1, wherein the method is characterized in that a three-dimensional data stack of the metal structural part is constructed and obtained by the following specific processes: Presetting a group of discrete sound beam incidence angle sets, repeating the steps S1 to S3 based on the actual sound velocity of the metal structural member for each preset sound beam incidence angle in the sound beam incidence angle sets to obtain sound velocity correction full-focusing images corresponding to each preset sound beam incidence angle, and further obtaining a three-dimensional data stack of the metal structural member.
9. The method for detecting the defects of the metal structural part based on the deep learning of claim 1, wherein the method for fusing the images in the three-dimensional data stack of the metal structural part is characterized by comprising the following specific steps: And carrying out pixel-level maximum value fusion on sound velocity correction full-focus images corresponding to all preset sound beam incidence angles in the three-dimensional data stack, and carrying out image preprocessing after carrying out pixel-level maximum value fusion to obtain a comprehensive detection image of the metal structural member.
10. The method for detecting defects of a metal structural member based on deep learning according to claim 1, wherein the defect detection is performed according to a comprehensive detection image of the metal structural member, and the specific process is as follows: And inputting the comprehensive detection image of the metal structural member into a deep learning defect detection model, and automatically outputting defect information of the metal structural member by the model.

Description

Metal structural member defect detection method based on deep learning Technical Field The invention relates to the technical field of defect detection, in particular to a method for detecting defects of a metal structural member based on deep learning. Background In the field of high-end equipment manufacturing, such as precision instruments, etc., the internal quality of a metal structural member is used as a core bearing and functional component, which directly determines the performance, safety and service life of the whole equipment, microscopic defects such as air holes, cracks, inclusions, unfused and the like are inevitably generated in the processes of manufacturing, such as casting, welding, additive manufacturing and using, and under the background, an ultrasonic detection technology capable of detecting the internal defects in a nondestructive mode with high sensitivity becomes an indispensable key link for guaranteeing the quality. The prior art discloses an object defect detection method, device, equipment and storage medium based on ultrasonic waves, wherein the method comprises the steps of determining the starting time and the ending time corresponding to a plurality of bottom waves of an object to be detected, wherein the bottom waves refer to ultrasonic signals reflected from the bottom of the object to be detected when the object to be detected is subjected to ultrasonic detection, calculating the multiple superposition wave width of the object to be detected according to the starting time, the ending time and a preset signal amplitude threshold, and judging whether the object to be detected has defects according to the magnitude relation between the multiple superposition wave width and the preset value. A method, equipment and medium for detecting near-surface defects of a metal material, which are disclosed in the patent publication No. CN117554493B, relate to the field of ultrasonic nondestructive detection, and are characterized in that a controller in a near-surface defect detection system of the metal material is used for determining the detection result of the near-surface defects of the metal material to be detected according to the secondary echo of oblique incidence ultrasonic waves received by a phased array probe, a test piece is arranged between an upper row magnet and a lower row magnet of a gradient magnetic field device, the phased array probe is arranged on the test piece, the gradient magnetic field device is used for providing a gradient magnetic field for the test piece and assisting in enhancing a meandering coil or phased array sub-aperture oblique incidence excitation energy in a set range, the detection method comprises the steps of setting the instrument gate time according to thickness and time and capturing the secondary reflection echo of the metal material to be detected by utilizing a full matrix capturing technology, and a secondary reflection full-focusing algorithm is adopted for each target pixel point in an imaging area of the metal material to be detected according to a full-focusing image. In combination with the above scheme, it is found that in the technical field of defect detection of a metal structural member of a precise instrument, only conventional time domain or amplitude analysis is generally performed on a received ultrasonic echo signal, however, in the actual detection, when ultrasonic waves propagate in a metal material with an anisotropic grain structure, such as a specific titanium alloy and an aluminum alloy, the propagation speed of the ultrasonic waves changes along with the direction, so that millimeter-level deviation occurs in defect positioning, and acoustic beam bending and path deviation are caused, so that detection images are blurred and the signal-to-noise ratio is reduced, and the deviation can directly lead to omission detection of key defects such as micron-level pores, inclusions or fatigue cracks, so that the accuracy of the overall defect detection result is seriously affected, and the severe requirement of the precise instrument field on quality zero error cannot be met. Disclosure of Invention Aiming at the defects of the prior art, the invention provides a metal structural member defect detection method based on deep learning, which can effectively solve the problems related to the background art. The method for detecting the defects of the metal structural member based on deep learning comprises the following steps of S1, configuring detection parameters of a phased array probe according to characteristic data of the metal structural member, and acquiring original full-matrix data of the metal structural member in a full-matrix capturing mode by using the phased array probe after parameter configuration. S2, preprocessing the original full matrix data of the metal structural member to obtain acoustic characteristic data of the metal structural member, and further determining the actual sound velocity of the metal