A method for detecting a rail weld and an imaging detector

Abstract

The application provides a rail weld joint detection method and an imaging detector. It relates to the technical field of nondestructive testing. The detection method specifically comprises the following steps: collecting a two-dimensional surface color image matrix and a multi-channel one-dimensional ultrasonic echo signal array, encapsulating them into a multi-modal data frame, and distributing them to a first-in-first-out buffer queue and an asynchronous latest frame covering cache area; extracting the two-dimensional surface color image matrix in the buffer queue, generating a two-dimensional semantic mask matrix, and extracting a surface three-dimensional normal vector; updating the absolute coordinates of the ultrasonic probe after displacement compensation, performing sound path space reverse calculation on the multi-channel one-dimensional ultrasonic echo signal array in combination with the three-dimensional normal vector, reconstructing a three-dimensional voxel storage matrix and an initial C-scan two-dimensional plan view; performing ultrasonic shallow artifact suppression, blind area crack compensation and high-risk damage confirmation on the initial C-scan two-dimensional plan view based on the two-dimensional semantic mask matrix, normalized gray gradient amplitude and local morphological linearity, and reversely correcting the C-scan atlas matrix.

Description

Technical Field

[0001] This application relates to the field of nondestructive testing technology, and more specifically, to a method for inspecting rail welds and an imaging inspection instrument. Background Technology

[0002] With the rapid development of modern transportation and industrial infrastructure, the operating load on steel components, especially railway rails, is increasing daily. As a weak point in the structural components, the welding quality and fatigue state of rail welds directly affect the safety and stability of the overall equipment or train operation. Due to the long-term reciprocating pressure of train wheels and the influence of complex external climatic environments, rail weld areas are highly susceptible to various internal volumetric defects such as fatigue cracks, slag inclusions, and porosity, as well as area-type defects such as surface scratches and spalling. Therefore, in the daily inspection scenarios of railway maintenance sections, high-precision and high-reliability non-destructive testing of rail welds is a crucial step in preventing serious accidents such as rail breaks.

[0003] Currently, on-site non-destructive testing of rail welds mainly relies on ultrasonic testing technology. Traditional on-site testing operations often use pulse-echo A-mode ultrasonic detectors, with operators holding the probe and scanning the surfaces such as the rail head, rail web, and rail base. Because the ultrasonic signal is displayed only as a one-dimensional time-voltage amplitude waveform, the test results heavily depend on the operator's on-site experience and concentration. This is not only cumbersome and inefficient, but also prone to false positives or missed detections due to visual fatigue during long, high-intensity continuous operations. To overcome the lack of intuitiveness in one-dimensional waveform displays, the industry has gradually introduced B-mode and C-mode ultrasonic imaging technologies. This technology, by combining data from the position encoder on the mechanical scanning device, transforms the one-dimensional echo signal into spatial coordinates, reconstructing a two-dimensional or three-dimensional cross-sectional image reflecting the internal structure of the weld, thus improving the visualization of defects to a certain extent.

[0004] However, in actual outdoor flaw detection scenarios, the aforementioned ultrasonic imaging technology still faces extremely severe environmental interference challenges. Firstly, the efficient propagation of ultrasonic waves requires extremely high acoustic coupling between the probe and the detection surface. However, the surface of rails in the field often exhibits varying degrees of corrosion, oil stains, water damage, or irregular mechanical wear. These surface abnormalities or rough textures significantly alter the actual incident and refraction angles of the ultrasonic waves, causing beam scattering or waveform conversion. This interference directly manifests in the ultrasonic imaging system as the generation of a large number of clutter and artifacts unrelated to the actual internal structure. When these abnormal echoes caused by surface interference are captured by the system and mapped into the ultrasonic image, they typically appear as suspected internal defects, resulting in a high false positive alarm rate. Simultaneously, ultrasonic flaw detection inevitably has a near-surface blind zone, making it extremely insensitive to minute cracks located precisely within this blind zone or extending from the surface into shallow layers.

[0005] On the other hand, existing ultrasonic imaging systems heavily rely on spatial displacement data provided by mechanical encoders to achieve spatial arrangement and reconstruction of ultrasonic echoes. When oil or ice is present on the weld surface, the rollers of the scanning frame inevitably slip relative to each other. This mechanical slippage directly disrupts the strict mapping relationship between the ultrasonic timing signal and the actual spatial coordinates, resulting in varying degrees of stretching, compression, or local spatial misalignment in the final ultrasonic B-scan and C-scan images, further reducing the accuracy of defect spatial localization and quantification.

[0006] In recent years, machine vision technology has been gradually applied to track surface inspection. Visual inspection equipment can accurately capture the macroscopic morphology and color features of the surface of the object being inspected, such as rust, oil stains, and macroscopic surface cracks. However, the optical characteristics of machine vision mean it cannot penetrate metallic media and cannot detect deep-seated defects within welds. Under the existing inspection framework, ultrasonic testing and visual inspection typically operate as two independent systems, with the acquired data being fragmented in both time and space. When a suspected defect area appears in the ultrasonic imaging spectrum, current technology lacks effective means to correlate it with the surface condition in real time. The inspection system cannot autonomously determine whether the abnormal ultrasonic signal originates from a real deep metal defect or is merely an acoustic artifact caused by surface corrosion or oil stains. This cross-modal information silo situation prevents existing testing equipment from effectively verifying and eliminating interference in complex and harsh field conditions, resulting in a high false alarm rate and severely hindering the improvement of intelligent flaw detection in rail welds. Summary of the Invention

[0007] This invention provides a method for inspecting rail welds, the method comprising: The synchronously acquired two-dimensional surface color image matrix and multi-channel one-dimensional ultrasonic echo signal array are encapsulated into multimodal data frames and distributed to the first-in-first-out buffer queue and the asynchronous latest frame overlay buffer, respectively. Extract the two-dimensional surface color image matrix from the first-in-first-out buffer queue, generate a two-dimensional semantic mask matrix, and extract the three-dimensional surface normal vectors. The absolute coordinates of the ultrasonic probe are updated by displacement compensation. The path space of the multi-channel one-dimensional ultrasonic echo signal array is calculated in reverse by combining the three-dimensional normal vector, and the three-dimensional voxel storage matrix and the initial C-scan two-dimensional top view are reconstructed. Based on the two-dimensional semantic mask matrix, normalized gray-level gradient magnitude and local morphological linearity, ultrasonic shallow artifact suppression, blind zone crack compensation and high-risk damage confirmation are performed on the initial C-scan two-dimensional top view, and the reverse-corrected C-scan atlas matrix is ​​output.

[0008] Extracting the three-dimensional normal vector of the surface specifically includes: converting the two-dimensional surface color image matrix into a single-channel grayscale matrix, extracting the effective area dominated by diffuse reflection and calculating the normalized reflection intensity, limiting the normalized reflection intensity of overexposed pixels caused by specular highlights to 1; calculating the beam incident angle using the normalized reflection intensity, and calculating the surface tilt azimuth using the two-dimensional spatial gradient of the image; combining the beam incident angle and the surface tilt azimuth to reconstruct the three-dimensional normal vector representing the actual surface.

[0009] The absolute coordinates of the ultrasound probe are updated through visual odometry absolute displacement compensation. Specifically, this involves: defining a reference template matrix in the central region of the previous frame image, performing pixel-level search and matching in the current frame, and calculating the lateral pixel translation of the image. Combined with the camera's horizontal pixel size Calculate the actual displacement increment between two trigger cycles. : Update the current number The true absolute coordinates of the ultrasonic probe in the longitudinal direction of the rail during each trigger cycle. : in, This represents the true absolute coordinates of the ultrasonic probe in the longitudinal direction of the rail in the previous frame.

[0010] Inverse path space calculation of a multi-channel one-dimensional ultrasonic echo signal array is performed by combining three-dimensional normal vectors. Specifically, this includes: obtaining the surface three-dimensional normal vector at the incident point through reverse addressing. Calculate the unit vector of the incident direction of the downward-incident probe. With surface three-dimensional normal vector cosine of the included angle ; Calculate the unit vector of the refraction direction of the ultrasonic wave transmitted into the interior of the rail. : in, The ratio of the velocity of sound in the transmitting medium to the velocity of sound in the incident medium; Combining true absolute coordinates with the unit vector of ultrasonic refraction direction Solve for the absolute spatial coordinate vectors of each time-series sampling point in the three-dimensional coordinate system of the multi-channel one-dimensional ultrasonic echo signal array.

[0011] The reconstruction generates a 3D voxel storage matrix and an initial C-scan 2D top view, specifically including: The signal intensity of the sampling points in the multi-channel one-dimensional ultrasound echo signal array is updated to the voxel cells of the corresponding three-dimensional voxel storage matrix using a strategy of overwriting the maximum value. The three-dimensional voxel storage matrix is ​​projected with the maximum amplitude along the depth direction to generate the signal intensity of the initial C-scan two-dimensional top view, and the exact depth corresponding to the signal intensity is retrieved. The spatial mapping model is used to calculate the corresponding pixel points of the two-dimensional grid points in the current visual frame to extract the two-dimensional semantic mask labels, normalized gray-level gradient magnitude, and local morphological linearity.

[0012] The extraction process of normalized gray-level gradient magnitude and local morphological linearity includes: using the horizontal and vertical Sobel gradient operators. and Calculate the normalized grayscale gradient magnitude : in, The extreme values ​​of the calibrated gray-scale gradient response; Construct a structure tensor matrix in the neighborhood of each pixel and compute its two non-negative eigenvalues. and And set Calculate local morphological linearity : in, To prevent extremely small constants with a denominator of zero.

[0013] Perform ultrasound shallow artifact suppression on the initial C-scan 2D top view, specifically including: When the signal strength of the initial C-scan 2D top view When the amplitude exceeds the effective alarm threshold and the label representation of the two-dimensional semantic mask matrix shows severe corrosion or oil contamination, the amplitude is determined based on the exact depth. Constructing empirical deep confidence weights : in, The confidence depth of the calibrated features; Suppression correction is applied to the initial C-scan 2D top view, and the corrected signal intensity is output in the inverse corrected C-scan spectral matrix. : The execution of blind zone crack compensation and high-risk nuclear damage resonance confirmation specifically includes: When the two-dimensional semantic mask matrix label represents a clean surface medium, and the normalized gray-level gradient amplitude is greater than the preset crack threshold and the local morphological linearity is greater than the preset linearity threshold, the three-dimensional voxels between the ultrasonic near-surface blind zone limit depth and the maximum depth of the rail bottom surface are extracted, and the local maximum amplitude in the deep effective detection area is calculated. If the amplitude of the local maximum value is less than or equal to the effective amplitude threshold of the alarm, a fixed artificial warning amplitude is injected into the reverse correction C-scan spectrum matrix to perform blind spot missed detection compensation. If the amplitude of the local maximum exceeds the effective amplitude threshold of the alarm, the amplitude of the reverse-corrected C-scan spectrum matrix will be saturated at the top to perform high-risk resonance confirmation of penetrating nuclear damage.

[0014] Generating a two-dimensional semantic mask matrix specifically includes: The two-dimensional surface color image matrix is ​​converted to a color space that includes hue, saturation and lightness components, and the cyclic hue difference of the pixels is calculated by combining the periodic wrap-around characteristics of hue. A feature vector is constructed using the cyclic hue difference, saturation, and lightness components; the Mahalanobis distance of the feature vector belonging to a pre-calibrated Gaussian distribution model of various surface anomalies is calculated; When the Mahalanobis distance is less than the preset feature confidence threshold, a label representing the abnormal medium category is assigned to the corresponding pixel according to the minimum distance judgment principle to generate a two-dimensional semantic mask matrix.

[0015] The present invention also provides a rail weld imaging inspection instrument, comprising: The scanning device is equipped with a three-axis detection linkage mechanism and a constant pressure coupling spring mechanism, and an incremental encoder that outputs position pulse signals is mechanically connected to the transmission wheel shaft on the same axis. An ultrasonic probe assembly, mounted on a triaxial detection linkage mechanism, includes probes that simultaneously execute a transmit-receive mode and a self-transmit and self-receive mode; The image acquisition module is fixedly mounted on the main frame of the scanning device, with its optical lens facing the detection surface and maintaining a pre-calibrated rigid physical relative position with the ultrasonic probe group in space; The ultrasound acquisition module has independent parallel high-speed signal acquisition channels, which are connected to the ultrasound probe group respectively; The embedded data processing platform is connected to the incremental encoder, image acquisition module and ultrasonic acquisition module respectively. The embedded data processing platform is equipped with a first-in-first-out buffer queue and an asynchronous latest frame overlay buffer area, and is used to execute the rail weld detection method.

[0016] This application provides a rail weld inspection method and imaging inspection instrument, which is suitable for the complex and harsh working conditions of on-site flaw detection in railway engineering sections. By integrating machine vision with multimodal deep fusion, the accuracy, reliability and robustness of detecting internal defects and surface damage in rail welds are improved.

[0017] To address the unavoidable fluctuations in movement speed and differences in response delays between heterogeneous sensors during manual on-site pushing of the scanning device, this application introduces an adaptive spatial synchronization triggering mechanism based on local motion compensation, eliminating the misalignment problem between the visual image and the ultrasonic beam during dynamic scanning. Based on this, the system constructs a dual-path asynchronous distribution and caching architecture. This architecture not only ensures the rigorous spatiotemporal topological coherence of the data required by the background 3D reconstruction algorithm through a first-in-first-out queue, but also prevents the memory accumulation impact of high-frequency acquisition bursts of data on the low-frequency display terminal through an asynchronous latest frame overwrite mechanism, ensuring that on-site operators can obtain extremely low-latency and lag-free real-time flaw detection feedback during human-computer interaction.

[0018] When faced with situations where mechanical encoders slip due to rail surface oil or ice, a novel approach is taken to construct a visual odometer using the high-frequency overlap features of consecutive frame images. The absolute displacement coordinates are reconstructed using the rigid translation features of optical flow, eliminating longitudinal stretching distortion caused by mechanical ranging failure. Simultaneously, the true three-dimensional surface normal vector is extracted to address the irregular mechanical wear on the rail surface, and the inter-refractive direction is introduced for spatial dynamic inverse calculation. Adaptive sound path compensation based on prior information of the visual true boundary corrects the beam deflection error caused by surface wear, achieving zero-distortion three-dimensional voxel mapping and cross-sectional reconstruction of ultrasonic signals within the rail weld.

[0019] More importantly, this application constructs a comprehensive multimodal image conflict arbitration and reverse correction mechanism, breaking through the perception limit of a single ultrasonic mode. Addressing the problem that common rust patches and oil stains on rail surfaces easily generate strong interface clutter and lead to false positives in ultrasonic testing, the system utilizes visual semantic features to adaptively suppress the depth confidence of the ultrasonic reconstruction matrix, removing shallow acoustic artifacts caused by non-defective media. To address the common problem of near-surface blind zone omissions in ultrasonic flaw detection, the system combines visually extracted high-frequency gradients and local morphological features to identify surface micro-fatigue cracks with high confidence and injects warning compensation within the blind zone. When the visually represented crack and deep ultrasonic high-echo occur at the same spatial location, generating cross-modal resonance, the system can autonomously confirm penetrating fatigue defects that pose a high risk of rail breakage. This mechanism enables the system to possess intelligent decision-making capabilities that eliminate interference from complex working conditions, significantly reducing the false negative and false positive rates in on-site flaw detection. Attached Figure Description

[0020] Figure 1 This is a flowchart of the rail weld inspection process of the present invention; Figure 2 This is a comparison diagram of displacement compensation data when the mechanical encoder of the present invention slips. Figure 3 This is the weighting function curve for suppressing superficial ultrasound artifacts in this invention; Figure 4 This is a geometric model of the surface three-dimensional normal vector extraction and grayscale mapping of the present invention. Detailed Implementation

[0021] This embodiment provides a rail weld imaging inspection instrument, mainly used in industrial non-destructive testing, preferably for various types of railway rails, including but not limited to mainstream 60-gauge and 75-gauge models, for simultaneous full-section flaw detection of welds. This imaging inspection instrument, through hardware module arrangement and structural design, achieves multimodal simultaneous acquisition of ultrasonic echo signals and surface visual images.

[0022] Specifically, the rail weld imaging inspection instrument of this embodiment includes: a scanning device, an ultrasonic probe group, an image acquisition module, an ultrasonic acquisition module, an embedded data processing platform, and a human-computer interaction display terminal.

[0023] The scanning device forms the basis for the system's mechanical displacement and sensor mounting. The main body of the scanning device is preferably made of aluminum alloy to reduce overall weight, while the guide wheels and probe housings that contact the rail are preferably made of steel to prevent jamming during the scanning process. The scanning device incorporates a three-axis detection linkage mechanism, which can synchronously drive the entire detection assembly to move smoothly along the rail via a handwheel. An internal constant-pressure coupling spring mechanism is installed in the scanning device. Through the mechanical deformation of the spring, the mounted ultrasonic probe assembly maintains a constant downward pressure of 5-10N as it slides along the detection surface, thus ensuring the stability of the ultrasonic coupling state. Furthermore, a high-precision pulse-type incremental encoder is coaxially mechanically connected to the drive shaft of the scanning device. This encoder has a minimum scanning increment resolution of 0.1mm and is used to output real-time hardware position pulse signals reflecting the system's absolute displacement on the rail.

[0024] The ultrasonic probe assembly is fixedly mounted on the triaxial detection linkage mechanism of the scanning device, capable of covering the core cross-sectional area of ​​the rail weld in a single operation. The ultrasonic probe assembly preferably includes three single-crystal probes and three pairs of tandem (K-type) probes. Regarding specific parameter selection, differentiated probe configurations are adopted based on the acoustic thickness and defect orientation of different parts of the rail: a probe with a 50-degree tilt angle and a center frequency of 2.5MHz is preferably used for detecting the rail head area; a probe with a 37-degree tilt angle and a center frequency of 4MHz is used for detecting the rail web area; and a probe with a 41-degree tilt angle and a center frequency of 4MHz is used for detecting the rail bottom area, which has complex bottom surface reflections. In terms of hardware circuitry, each pair of probes is connected to an independent transmit / receive control loop, supporting simultaneous execution of a one-transmit-one-receive mode and a self-transmitting and self-receiving mode. The one-transmit-one-receive mode captures area-type defect echoes through spatial sound field convergence, while the self-transmitting and self-receiving mode captures volumetric defect echoes using the original path reflection.

[0025] The image acquisition module preferably uses an industrial-grade color camera, with uniform supplementary lighting sources arranged around the camera. This image acquisition module is fixedly mounted on the main frame of the scanning device, with its optical lens pointing downwards towards the rail surface, and maintaining a pre-calibrated rigid relative position with the ultrasonic probe group in three-dimensional space. The hardware trigger control terminal of the image acquisition module is electrically connected to the pulse output terminal of the incremental encoder, used to synchronously acquire high-resolution RGB color images of the rail surface under specific displacement increments.

[0026] The ultrasonic acquisition module is configured as a high-speed signal acquisition circuit with eight independent parallel channels. The analog section of the module includes an ultrasonic signal preprocessing channel, a multiplexer, a controllable gain amplifier, and a hardware detector buffer; the digital section includes a high-speed analog-to-digital converter (A / D) and timing logic control circuitry. The high-speed A / D converter has a sampling frequency set to 120MHz and a bandwidth configured as a wideband mode of 0.4–15MHz. The pulse transmitter supports the emission of negative spike pulses, and the pulse width is continuously adjustable within the range of 50–1000ns. The system's ultrasonic repetition trigger frequency is adjustable within the range of 100–1000Hz. The ultrasonic acquisition module communicates with the embedded data processing platform via a high-speed USB bus, enabling non-blocking transmission of parallel data from the eight flaw detection channels.

[0027] The embedded data processing platform, serving as the central hub for data computation and coordination control of the device, employs a low-power embedded industrial computer system architecture. This platform preferably features an Intel ATOM 270 core processor, an integrated solid-state drive (e.g., a 64GB SSD), and RAM, running Windows XPE or an equivalent real-time embedded operating system. In terms of hardware memory address space allocation, the data processing platform has dedicated data buffer queue memory areas, including a double-buffered first-in-first-out (FIFO) queue memory area for lossless reception of multi-channel raw ultrasound data, encoder position data, and RGB image data, and a last-in-first-out (FILO) queue memory area for sampling and displaying the foreground interface. This hardware-level memory block design ensures timing integrity during multi-threaded concurrent processing.

[0028] The human-machine interface display terminal is connected to the embedded data processing platform via an LVDS or VGA video bus, preferably configured as a 10.4-inch, 1024×768 pixel resolution high-brightness true-color LCD touchscreen for real-time rendering of multimodal flaw detection maps pushed from the platform's internal dual-buffered video memory. Simultaneously, the instrument's main panel is equipped with a full-duplex / half-duplex RJ45 network communication interface supporting the IEEE 802.3 standard and the TCP / IP network layer transmission protocol, used to transmit raw flaw detection data or generated reports back to the HMIS monitoring network in real time. The overall power supply of the equipment preferably uses a high-capacity lithium-ion battery (e.g., 6600mAh) to meet the needs of continuous mobile operation in outdoor environments without external mains power.

[0029] This embodiment, based on the hardware architecture, further provides a data processing method for a rail weld imaging inspection instrument. During on-site rail flaw detection, operators push the scanning device along the rail. Due to uneven rail surface friction and variations in manual pushing force, fluctuations in the scanning device's speed are inevitable. Simultaneously, the embedded data processing platform needs to balance multi-channel ultrasonic 3D computation with real-time UI rendering. How to ensure spatial alignment between the 2D optical image and the high-frequency 1D ultrasonic signal under such dynamic conditions with speed fluctuations, and how to solve data read / write blocking and display latency issues, are the technical problems this embodiment aims to address.

[0030] Therefore, the rail weld data processing method of this application includes the following steps: S1: synchronous acquisition and dual-path distribution caching of multimodal heterogeneous data.

[0031] This step primarily addresses the spatial alignment issue of heterogeneous sensors during dynamic scanning, as well as the asynchronous distribution problem of embedded system buses when facing high-frequency burst data. Specifically, it includes the following steps: S11: Adaptive spatial synchronization triggered acquisition based on local first-order motion compensation: In rail flaw detection, the acoustic response time of the ultrasonic probe array is extremely short, while the image acquisition module uses global shutter exposure, and the scanning device continues to move along the rail during its exposure integration time. If a fixed hardware pulse is used to trigger both the camera and the ultrasonic probe at the same moment, the optical field of view at the exposure center will deviate from the actual incident point of the ultrasonic beam, resulting in mode misalignment.

[0032] To address the aforementioned issues, this application introduces an advance triggering mechanism based on local first-order motion compensation. The direction of the scanning device's movement along the rail is defined as the positive displacement direction. Regarding hardware installation constraints, the optical center axis of the image acquisition module is defined to lead the central acoustic beam axis of the ultrasonic probe assembly along the positive displacement direction, and the mechanical installation offset distance between the two is calibrated as follows: (rice, ,and ).

[0033] The embedded data processing platform reads the incremental encoder signal in real time and calculates the instantaneous speed of the scanning device along the longitudinal direction of the rail using differential calculation. (meters per second) ). Let the system be set to the first The absolute coordinates of the target ultrasonic sampling points in the longitudinal coordinate system of the rail are: (meter, m).

[0034] Considering the inherent response delay time of the hardware triggering of the image acquisition module (Second, ), and exposure points time (Second, ).because and The time window is extremely short (on the order of milliseconds), within which the speed fluctuation of the scanning device is negligible. Therefore, this step uses a local first-order uniform velocity model for engineering approximation compensation, while assuming uniform shutter weights, and equating the spatial centroid of exposure motion blur to the midpoint of time. The embedded data processing platform outputs independently calculated pre-visual trigger spatial coordinates to the image acquisition module. : The second term on the right side of the equation is the bias compensation. The third item is dynamic displacement compensation consisting of the product of velocity and time window. ).

[0035] because And during the scanning advance Therefore, there is When the scanning device advances to the preceding coordinate... At that time, the system triggers the image acquisition module; when the position continues to advance to reach... At that time, the system triggers the ultrasound acquisition module. Thus, the system obtains ultrasound data at the same coordinates. Strictly aligned two-dimensional surface color image matrix With multi-channel one-dimensional ultrasound echo signal array .

[0036] S12: Ordered encapsulation of multimodal data frames: To ensure the correlation of cross-modal data, the embedded data processing platform will use coordinates Two-dimensional surface color image matrix acquired at the location Multi-channel one-dimensional ultrasonic echo signal array Location coordinates and the system absolute timestamp generated by the system's internal clock. According to the preset structure protocol, it is encapsulated into multimodal data frames in an orderly manner. Its data structure is represented as ordered tuples: Each complete spatial synchronization trigger cycle generates a discrete multimodal data frame with internal spatiotemporal constraints. .

[0037] S13: Dual-path asynchronous dispatch cache based on differentiated lifecycles: Generate multimodal data frames Subsequently, the data flow needs to be scheduled. 3D reconstruction algorithms require a coherent spatial sequence without dropped frames; however, human-computer interaction display terminals are limited by the LCD hardware refresh rate (e.g., 60Hz), and only need to obtain the latest frame from the operator's current pushing position to assist in on-site judgment. Any backlog of historical transition frames will cause display delays and misjudgments in the UI interface.

[0038] To address this, the embedded data processing platform allocates two independent buffer areas in memory to construct the main processing flow and the UI observation flow: Construct a main processing flow with a first-in, first-out (FIFO) queue mechanism: All generated multimodal data frames... Fully push the first-in-first-out (FIFO) buffer queue, denoted as . Enqueuing attributes follow a monotonically increasing timestamp, i.e.: like First press in Press in again .

[0039] High-time-consuming background processing threads (such as image semantic extraction and acoustic path coordinate space transformation) have exclusive access. The FIFO mechanism ensures that the order in which the algorithm retrieves data is strictly isomorphic to the sliding trajectory of the scanning device on the rail, thus guaranteeing the topological integrity of the 3D voxel reconstruction.

[0040] To construct a UI observation stream, an asynchronous latest frame overlay caching mechanism is implemented: To eliminate logical redundancy in the stack structure during fetch-and-clear mode and to completely block the memory accumulation impact of high-frequency acquisition on low-frequency display, the system configures an asynchronous latest frame overlay cache for the UI thread, reserving only a single frame space, denoted as... .

[0041] In Press in At the same time, the acquisition thread uses a lock-free concurrency mechanism to... The copy is directly overwritten to In China, we have always maintained The data within is the latest frame in a temporal sense.

[0042] The foreground UI rendering thread is limited by the hardware display timer, at fixed time intervals. (Second, The UI rendering thread is periodically woken up. Each time it is woken up, the UI rendering thread executes the following extraction logic, directly reading... The data frame resident in the middle is used as the current rendering frame. : UI rendering thread extraction The internal image matrix and ultrasonic wave pattern are used to refresh the screen. This asynchronous single-slot overwrite mechanism is based on the visual requirement of on-site workers to "only observe the real-time working conditions directly below the probe," reducing the rendering overhead of historical data and achieving extremely low latency and zero lag in the human-machine interface.

[0043] In real railway maintenance inspections, the assumption that the rail surface is an ideally flat, rigid body with a constant acoustic coupling state often fails. The rail surface commonly faces two major interference factors: first, surface coatings, such as brownish-red rust patches or black oil stains, which can cause severe scattering of ultrasonic waves or acoustic impedance mismatch, resulting in acoustic artifacts similar to real deep nuclear damage during ultrasonic acquisition; second, irregular mechanical wear and chipping on the surface, which can cause the actual three-dimensional normal vector of ultrasonic waves incident on the rail surface to deflect, leading to distortion in the spatial coordinate calculation of internal defects.

[0044] S2: Visual modality feature extraction and spatial coordinate calibration: This step is performed by a background processing thread in the embedded data processing platform, through the queue. Two-dimensional surface color image matrix The process involves analyzing and extracting prior information about the rail surface condition, and establishing a precise spatial mapping between visual and ultrasonic measurements. Specifically, this includes the following steps: S21: Surface color and semantic feature extraction based on cyclic Mahalanobis distance: To eliminate the interference of light intensity fluctuations on color recognition during on-site inspections, the system uses a two-dimensional surface color image matrix in the RGB color space. Convert to the HSV (Hue, Saturation, Lightness) color space. (Hue is taken into consideration.) have The periodic orbital characteristics (i.e.) and Visually highly similar, but with a very large Euclidean range, a cyclic minimum difference mapping is introduced for the hue components.

[0045] set up The position of any pixel in the middle is (pixels, Taking rust feature extraction as an example, let the rust reference color hue angle pre-calibrated by the system be... Calculate the cyclic hue difference value of this pixel. : Construct the corrected feature vector Its definition is: A multivariate Gaussian distribution model of rust and oil stains in HSV space is pre-established offline. Let the prior mean vector of the rust category be... The inverse of the prior covariance matrix is Calculate the current feature vector pixel by pixel. Mahalanobis distance for corrosive media : Similarly, the Mahalanobis distance belonging to the oil pollution category is calculated. .

[0046] Set a uniform feature confidence threshold Generate a match based on the minimum distance principle. Two-dimensional semantic mask matrix with consistent size : Among them, the label Characterized by severe corrosion, label Characterizing oil stains, labels The medium is normal. When the boundary is equal (the distances of the two types are equal and both are less than the threshold), it is preferentially classified as oil to conservatively suppress ultrasonic emission.

[0047] S22: Surface 3D Normal Vector Extraction: A single scalar tilt angle is insufficient to support complete three-dimensional Snell refraction compensation. This step utilizes a near-coaxial directional light source configured in hardware to ensure field illumination while maintaining a monotonic mapping between image grayscale and surface tilt, thereby extracting the complete three-dimensional surface normal vector. .

[0048] Will Convert to a single-channel grayscale matrix And filter out The region of the effective metallic microfacet where diffuse reflection is dominant. Let the reference reflectance grayscale value be a flat, wear-free surface. Calculate the normalized reflection intensity. (Dimensionless): For overexposed pixels caused by specular highlights ( ), limiting it to That is, it is equivalent to a horizontal smooth surface with zero tilt angle.

[0049] Light beam incident angle (zenith angle) The calculation model is as follows: Simultaneously, the surface tilt azimuth angle is calculated using the two-dimensional spatial gradient of the image. : in, and These are the Sobel gradient operators for the horizontal and vertical directions of the pixels, respectively.

[0050] By combining the zenith angle and azimuth angle, a 3D normal vector model representing the actual surface at this pixel point is reconstructed: This generates a two-dimensional geometric normal vector matrix. It includes the complete tilt amplitude and orientation.

[0051] S23: Rigid calibration and mapping of visual-ultrasound spatial coordinates: The above and In pixels In the two-dimensional image plane based on the rail, it must be mapped to the three-dimensional coordinate system of the rail. ultrasonic echo signal Establish alignment relationships.

[0052] The origin of the coordinate system is the projection of the central acoustic beam incident point onto the rail surface at the moment of current triggering of the ultrasonic probe assembly. Let the absolute longitudinal coordinate of any target point on the rail in space be... (meters, m), with the horizontal absolute coordinate as (meters, m). The pixel size of the camera's image sensor is set to horizontal. With longitudinal ( The coordinates of the optical principal point are ( ).

[0053] Based on the motion compensation logic executed in step S11, since the early triggering mechanism has ensured that at the center of image exposure, the camera's optical principal point exactly crosses the current ultrasonic triggering reference coordinates. Therefore, vertical mapping should no longer use visual trigger coordinates. Instead, directly with This serves as the positioning point for the principal point of the image plane. Simultaneously, a mechanical mounting offset is introduced between the image acquisition module and the ultrasonic probe assembly along the transverse direction of the rail. (rice, ).

[0054] any point on the surface of the rail Projected onto the pixel coordinates of the current frame image The geometric mapping model is represented as: Through this mapping model, the data processing platform can calculate absolute position. When dealing with any ultrasound voxel point, directly reverse addressing is used to obtain the point's position in the visual frame. The corresponding semantic state With three-dimensional normal vector This solves the problem of spatiotemporal misalignment between modalities.

[0055] This embodiment further illustrates the spatial reconstruction process of ultrasonic echo data by the main processing flow.

[0056] In traditional ultrasonic flaw detection equipment, the spatial conversion from A-scan (one-dimensional waveform) to B-scan (two-dimensional longitudinal section) or C-scan (two-dimensional top view) is usually based on the assumption of absolute precision in mechanical encoder displacement and that the ultrasonic probe is in close contact with the rail surface and refracts at a constant nominal angle. However, in actual working conditions, oil stains and ice on the rail surface can easily cause microscopic slippage of the scanning device, severing the correlation between hardware pulses and displacement; at the same time, irregular surface wear can cause the normal vector of the incident point to deflect. If these dynamic errors are not eliminated, the reconstructed three-dimensional spatial coordinates of the defect will suffer severe longitudinal stretching and depth distortion.

[0057] Therefore, this embodiment provides an ultrasonic spatial adaptive reconstruction step in the data processing method, which uses extracted visual features to guide and correct the ultrasonic signal: S3: Reconstruction from A-scan to B / C-scan guided by visual features: This step is performed by the embedded data processing platform from the first-in-first-out buffer queue ( Extracting multimodal data frames sequentially from ) Perform the following steps: S31: Visual odometry absolute displacement compensation based on sequence image feature matching: To significantly suppress the cumulative longitudinal displacement error caused by mechanical roller slippage, this step utilizes a matrix of two consecutive surface color images. and The high-frequency overlap characteristics are used to introduce visual odometry to calculate the true displacement.

[0058] Suppose the system extracts the current frame At that time, the absolute coordinates of the ultrasonic probe group in the longitudinal direction of the rail in the previous frame have been confirmed as follows: (rice, The system is in A reference template matrix is ​​defined in the central region, and the normalized cross-correlation (NCC) algorithm is used in the current frame. Pixel-level search and matching are performed. Based on the constructed imaging geometry system, the horizontal direction of the image is... longitudinal axis and rail Since the axes are strictly parallel and corresponding, the horizontal pixel shift of the image calculated by the system represents the longitudinal displacement increment of the probe along the rail, denoted as... (pixels, ).

[0059] Combined with the horizontal pixel size of the camera (meters per pixel) ), calculate the displacement increment between two trigger cycles. : Then, update the current number. The true absolute coordinates of the probe group in the longitudinal direction of the rail during each trigger cycle : This step reconstructs the system's absolute displacement coordinate system through the continuous optical flow rigid translation feature, effectively absorbing and correcting the misalignment of the hardware trigger pulse in high-speed, high-frequency, short-baseline acquisition.

[0060] S32: Adaptive sound path space calculation based on three-dimensional normal vectors: Determining the absolute position Then, the one-dimensional ultrasonic echo signal array needs to be... Mapped to the interior of the rail. This step is based on the extracted true surface normal vector, assuming the nominal emission tilt angle of a specific ultrasonic probe is... In the system's global coordinate system (Z-axis pointing vertically downwards into the rail, X-axis along the longitudinal direction of the rail, Y-axis along the transverse direction), the unit vector of the ultrasonic wave's incident direction within the probe... Represented as: The data processing platform uses the absolute coordinates of the current ultrasound probe. Reverse addressing is used to obtain the surface three-dimensional normal vector at the incident point. The direction points upwards into the air, from The matrix is ​​provided.

[0061] Let the incident sound velocity of the ultrasonic wave within the probe wedge be... (meters per second) The velocity of sound transmitted through the metal material of the rail is ( Define the ratio of the sound velocity of the transmitting medium to that of the incident medium. : Downward incident wave vector With the upward normal vector The cosine of the included angle is calculated as follows: The unit vector of the refraction direction of ultrasonic waves transmitted into the interior of the rail for: If the criterion within the square root If the signal is zero, it indicates that the sound wave undergoes total reflection at the current normal vector interface and cannot enter the rail. The system then marks the ultrasonic data at the current coordinate point as invalid.

[0062] For a one-dimensional array The first in There are 1 time-series sampling points, and the system sampling frequency is 1. (hertz, The round-trip inherent acoustic path delay time inside the probe wedge is (Second, The sampling point's round-trip flight time inside the rail. (dimensions are) )for: One-way radial acoustic path distance from the defect reflecting surface to the incident point ( The calculation is as follows: Combining the coordinates of the incident reference point and the refraction vector, this first... The absolute spatial coordinate vector of each sampling point in the three-dimensional coordinate system ( The solution is: S33: Cross-modal label fusion and 3D voxel matrix reconstruction: The system establishes a uniformly discretized three-dimensional voxel storage matrix in space. The data processing platform will use arrays The Middle Signal strength at each sampling point (Using percentage amplitude), the strategy of overwriting the maximum value is adopted to update the corresponding absolute coordinates. In the voxel unit.

[0063] During this process, the extracted two-dimensional semantic mask matrix is ​​invoked synchronously. If the current ultrasonic incident reference point is determined to be disturbed in visual features, such as a mask label indicating rust or oil contamination at that location, the system updates the voxel matrix. At that time, surface visual interference identifiers will be injected into all spatial voxel properties derived from the incident point.

[0064] The data processing platform performs maximum amplitude projection: along The axial direction projects voxels into a precisely corrected form. Planar longitudinal section view (B-scan view); along axial projection is Planar top view (C-scan base image). Through the dual intervention of visual displacement correction and three-dimensional normal vector refraction, the system reconstructs a flaw detection spatial map that removes mechanical slippage and wear distortion.

[0065] This embodiment further provides a C-scan reverse correction method based on multimodal image collision detection.

[0066] By employing both visual displacement correction and three-dimensional normal vector refraction, a three-dimensional voxel storage matrix with accurate spatial coordinates was reconstructed. And the initial C-scan two-dimensional top view. However, in actual rail flaw detection, ultrasonic imaging often faces two inherent modal misleading factors: one is shallow acoustic artifacts (false positives). When there are rust patches or oil stains on the rail surface, the ultrasonic waves will generate strong interface clutter reflections at the interface between the wedge and the rough medium. During reconstruction, these clutters will appear as abnormally high reflection amplitudes in the shallow depth range of the C-scan image, which are easily misjudged as internal nuclear defects. The second is near-surface blind zone missed detection (false negatives). Due to the blind zone effect of the ultrasonic emission initial pulse, the system can only detect the shallowest layer of the rail (such as... Microscopic cracks (at a certain depth) are extremely insensitive. When early micro-fatigue cracks occur on the rail surface, the initial ultrasonic C-scan may show no defects, thus missing the core hidden danger that could lead to rail breakage.

[0067] To overcome the limitations of a single ultrasound modality, this embodiment provides a core multimodal image conflict detection and C-scan reverse correction step. A mutually exclusive decision tree rule is designed, utilizing visually extracted surface realities to determine and correct the initial ultrasound C-scan atlas. S4: Multimodal Image Collision Detection and C-Scan Reverse Correction This step is performed by the embedded data processing platform after completing the current multimodal data frame. After the 3D voxel reconstruction, the process involves pixel-by-pixel conflict arbitration of heterogeneous data based on a unified absolute coordinate system, specifically including the following steps: S41: Homologous Alignment Query and Morphological Extraction of Cross-Modal Heterogeneous Features: The system defines the reconstructed initial C-scan 2D top view as a discrete matrix. ,in The position of the two-dimensional grid in the absolute coordinate system (meters). The matrix element value is the depth direction of that grid point. Maximum signal strength percentage amplitude on axis (unitless, range) Simultaneously, the system stores data in a three-dimensional voxel matrix. (The amplitude of its stored voxel units is also specified as) In the above, the exact depth corresponding to the maximum amplitude value is retrieved and denoted as . ( ).

[0068] For grid points Using the constructed spatial mapping model, the corresponding pixel in the current visual frame is calculated. And extract the following three types of visual ground truth features: visual semantic mask labels: directly called (Values) Dimensionless normalized grayscale gradient magnitude Using the horizontal and vertical gradient operators, the computational model is as follows: in, The extreme values ​​of the calibrated gray-scale gradient response. Local morphological linearity High gradient amplitudes can be caused by fatigue cracks, or they can be harmless isolated scratches or irregular wear. The system at the pixel level... Construct a structure tensor matrix within the neighborhood and compute its two non-negative eigenvalues. and (and set) Define the dimensionless linearity index: When the feature exhibits a distinctly narrow linear shape, such as a real crack: ; When the features are isotropic spots or ordinary scratches: .

[0069] Introducing a minimum constant To prevent the denominator from being zero.

[0070] S42: Suppression of shallow ultrasound artifacts based on visual semantic masks: This step has the highest decision priority and is specifically designed to eliminate false positives in ultrasound caused by surface rust and prevent visual features from being misjudged in rusted areas.

[0071] When the grid point initially alarms ( , (The effective amplitude threshold for alarm), and visual semantic judgment of surface medium abnormality (i.e. Perform this step when ( ).

[0072] Under this condition, considering that clutter likelihood decreases with increasing depth, the system is constructed based on the actual depth. Empirical depth confidence weight : in, The confidence depth of the calibrated features ( ).

[0073] Based on this, the system performs suppression correction on the initial C-scan image: When the anomaly is located in a very shallow layer ( )hour, Artifacts are effectively eliminated; when anomalies are deep, Preserve the true internal signal. If this step is triggered, output directly. This prevents the process from proceeding to the next step, thus avoiding the logical conflict of mistaking the high gradient at the edge of rust or oil stains for cracks.

[0074] S43: Blind zone crack compensation supported by high-frequency visual features: This step is specifically designed to address missed detections (false negatives) in near-surface blind zones and to mark high-risk lesions during multimodal resonance.

[0075] Only when the grid point does not trigger priority one (i.e. (The surface medium is clean), while satisfying a high gradient ( ) and high linearity ( This is performed under a dual threshold. This condition ensures that the vision captures genuine surface fatigue cracks, rather than ordinary scratches or interference.

[0076] After triggering, the system extracts the limiting depth of the near-surface blind zone of the ultrasound. ( (to the maximum depth of the bottom surface of the rail) Calculate the amplitude of local maxima within the effective detection zone of the deep layer using three-dimensional voxels. : Utilizing this deep peak characteristic and peak alarm threshold Perform an objective comparison at the same scale: Blind spot detection compensation: If This indicates that no abnormalities were found in the deeper layers, and the crack was confined to the ultrasonic blind zone. The superficial layer within. Due to missed detection by ultrasound, the system injects a fixed artificial warning amplitude into the C-scan matrix. (e.g., retrieving parameters) ): A single-mode shallow crack marker is applied to the grid point.

[0077] Confirmation of high risk for penetrating injury: If This indicates that while a high-linearity crack appeared on the surface, an abnormally strong echo also appeared in the deeper layer outside the blind zone directly below it. This cross-modal spatial pattern confirms that the crack has maliciously propagated deep into the rail, constituting the most dangerous penetrating damage. The system directly sets the amplitude at this point to the maximum. And trigger the highest level of multimodal penetrating nuclear damage alarm for that grid point.

[0078] For ordinary grid points that do not trigger any conditions S42 and S43, the system directly inherits the original mapping, i.e. .

[0079] The embedded data processing platform outputs the final reverse-corrected C-scan spectral matrix. This mechanism, through a mutually exclusive decision tree, not only resolves logical conflicts but also fully combines the geometric sensitivity of vision to micro-cracks with the penetrating power of ultrasound to internal depth, ultimately achieving highly reliable intelligent flaw detection that eliminates artifacts and compensates for blind spots.

[0080] This embodiment further provides a final method for multimodal graph rendering and human-computer interaction display.

[0081] In continuous on-site scanning operations of rail welds, the high-frequency emission of the ultrasonic probe group and the background 3D reconstruction algorithm typically update image data in system memory at extremely high frequencies (e.g., 500Hz or higher). However, the hardware refresh rate of the human-computer interaction display terminal (LCD touch screen) is usually low (e.g., 60Hz). If the background processing thread is allowed to directly drive and rewrite the foreground memory of the display terminal after each generation of a new voxel or extraction of a new image frame, the high-frequency data stream will cause a severe read-write conflict with the low-frequency screen refresh cycle. Visually, this manifests as severe screen flickering, screen tearing, and UI interface lag, greatly affecting the operator's observation and judgment of real-time defect signals.

[0082] To resolve the asynchronous conflict between high-frequency data processing and low-frequency hardware display, and to smoothly present the reverse-corrected multimodal graph to the operator, this embodiment provides the final rendering step in the data processing method: This step is independently handled by the UI rendering thread in the embedded data processing platform. By constructing an isolated drawing workspace in memory, complete decoupling of the underlying data flow from the foreground display flow is achieved. Specifically, it includes the following implementation steps: The embedded data processing platform allocates two graphics buffer blocks with independent addresses and the same capacity in the system's graphics memory or main memory area: one is defined as the back buffer, and the other is defined as the front buffer, which is directly connected to the display hardware.

[0083] The foreground memory is write-protected, and its data is scanned in read-only mode by the display hardware controller at a fixed refresh rate (e.g., 60Hz) and mapped onto the LCD screen pixels. The background buffer, on the other hand, is fully open to memory read and write operations and serves as a high-frequency virtual canvas for background processing threads and drawing functions.

[0084] As the scanning device advances, the data processing platform plans a multi-window UI layout within the background buffer and silently renders various deeply processed multimodal data at high frequency (synchronized with the sensor triggering frequency) within this background buffer. The specific content of the composite rendering includes: In the core main view area of ​​the screen, the final output reverse-corrected C-scan atlas matrix (top view) and B-scan base map (longitudinal section view) are filled with two-dimensional images. During the drawing process, different color bands are used to map the ultrasonic amplitude values, and special grid points that trigger high-risk injuries or shallow cracks in blind areas are overlaid with highlighted warning borders or specific colors.

[0085] To allow operators to intuitively grasp the current instantaneous state directly below the probe, the system retrieves the constructed "UI observation stream latest frame overlay buffer". From this buffer, it extracts the original A-scan waveform segment with the latest absolute timestamp and the corresponding color camera footage, and plots both of them in a preset independent floating monitoring window in the background buffer.

[0086] Throughout the drawing process, since all pixel rewriting operations are completed through memory copying within a hidden background buffer without generating any direct driving signals to the screen, the system can handle data refresh and redrawing at the hundreds of hertz level without any hindrance, completely eliminating screen flicker and tearing.

[0087] The front-end UI rendering thread is configured with an independent hardware-level high-precision timer. The trigger frequency of this timer is locked to be consistent with the refresh rate of the LCD touch screen (e.g., fixed at a 60Hz time interval).

[0088] Whenever a timer triggers and wakes up the UI rendering thread, the thread immediately suspends the brief write operation to the background buffer (using a mutex mechanism), and then calls the underlying block memory copy interface to push the complete composite graph that has been drawn in the background buffer into the foreground video memory all at once. After the push is complete, the write lock on the background buffer is immediately released, allowing the background to continue the next round of high-frequency silent drawing.

[0089] Finally, it should be noted that those skilled in the art will understand that all or part of the processes in the above embodiments can be implemented by a computer program instructing related hardware. The program can be stored in a computer-readable storage medium, and when executed, it can include the processes of the embodiments of the above methods. The storage medium can be a magnetic disk, optical disk, read-only memory (ROM), or random access memory (RAM), etc. The above computer program embodiments can achieve the same or similar effects as any of the corresponding foregoing method embodiments.

[0090] Furthermore, typically, the devices and equipment disclosed in the embodiments of this invention can be various electronic terminal devices, such as mobile phones, personal digital assistants (PDAs), tablet computers (PADs), smart TVs, etc., or they can be large terminal devices, such as servers. Therefore, the scope of protection disclosed in the embodiments of this invention should not be limited to a specific type of device or equipment. The client disclosed in the embodiments of this invention can be applied to any of the above-mentioned electronic terminal devices in the form of electronic hardware, computer software, or a combination of both. Furthermore, the method disclosed in the embodiments of the present invention can also be implemented as a computer program executed by a CPU, which may be stored in a computer-readable storage medium. When the computer program is executed by the CPU, it performs the functions defined in the method disclosed in the embodiments of the present invention. Furthermore, the above-described method steps and system units can also be implemented using a controller and a computer-readable storage medium for storing a computer program that enables the controller to perform the functions of the above-described steps or units. Furthermore, it should be understood that the computer-readable storage medium (e.g., memory) described herein can be volatile memory or non-volatile memory, or may include both volatile and non-volatile memory. By way of example, and not limitation, non-volatile memory may include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), or flash memory. Volatile memory may include random access memory (RAM), which can act as external cache memory. By way of example, and not limitation, RAM may be available in various forms, such as synchronous RAM (DRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), dual data rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), synchronous link DRAM (SLDRAM), and direct Rambus RAM (DRRAM). The storage devices disclosed herein are intended to include, but are not limited to, these and other suitable types of memory. In this specification, the same or similar parts between the various embodiments can be referred to mutually. Each embodiment focuses on describing the differences from other embodiments. In particular, the descriptions of the embodiments described later are relatively simple, and relevant parts can be referred to the descriptions of the foregoing embodiments.

Claims

1. A method for inspecting rail welds, characterized in that, The method includes: The synchronously acquired two-dimensional surface color image matrix and multi-channel one-dimensional ultrasonic echo signal array are encapsulated into multimodal data frames and distributed to the first-in-first-out buffer queue and the asynchronous latest frame overlay buffer, respectively. Extract the two-dimensional surface color image matrix from the first-in-first-out buffer queue, generate a two-dimensional semantic mask matrix, and extract the three-dimensional surface normal vectors. The absolute coordinates of the ultrasonic probe are updated by displacement compensation. The path space of the multi-channel one-dimensional ultrasonic echo signal array is calculated in reverse by combining the three-dimensional normal vector, and the three-dimensional voxel storage matrix and the initial C-scan two-dimensional top view are reconstructed. Based on the two-dimensional semantic mask matrix, normalized gray-level gradient magnitude and local morphological linearity, ultrasonic shallow artifact suppression, blind zone crack compensation and high-risk damage confirmation are performed on the initial C-scan two-dimensional top view, and the reverse-corrected C-scan atlas matrix is ​​output.

2. The method for inspecting rail welds according to claim 1, characterized in that, Extracting the three-dimensional normal vector of the surface specifically includes: converting the two-dimensional surface color image matrix into a single-channel grayscale matrix, extracting the effective area dominated by diffuse reflection and calculating the normalized reflection intensity, limiting the normalized reflection intensity of overexposed pixels caused by specular highlights to 1; calculating the beam incident angle using the normalized reflection intensity, and calculating the surface tilt azimuth using the two-dimensional spatial gradient of the image; combining the beam incident angle and the surface tilt azimuth to reconstruct the three-dimensional normal vector representing the actual surface.

3. The method for inspecting rail welds according to claim 1, characterized in that, The absolute coordinates of the ultrasound probe are updated through visual odometry absolute displacement compensation. Specifically, this involves: defining a reference template matrix in the central region of the previous frame image, performing pixel-level search and matching in the current frame, and calculating the lateral pixel translation of the image. Combined with the camera's horizontal pixel size Calculate the actual displacement increment between two trigger cycles. : ； Update the current number The true absolute coordinates of the ultrasonic probe in the longitudinal direction of the rail during each trigger cycle. : ； in, This represents the true absolute coordinates of the ultrasonic probe in the longitudinal direction of the rail in the previous frame.

4. The method for inspecting rail welds according to claim 1, characterized in that, Inverse path space calculation of a multi-channel one-dimensional ultrasonic echo signal array is performed by combining three-dimensional normal vectors. Specifically, this includes: obtaining the surface three-dimensional normal vector at the incident point through reverse addressing. Calculate the unit vector of the incident direction of the downward-incident probe. With surface three-dimensional normal vector cosine of the included angle ; Calculate the unit vector of the refraction direction of the ultrasonic wave transmitted into the interior of the rail. : ； in, The ratio of the velocity of sound in the transmitting medium to the velocity of sound in the incident medium; Combining true absolute coordinates with the unit vector of ultrasonic refraction direction Solve for the absolute spatial coordinate vectors of each time-series sampling point in the three-dimensional coordinate system of the multi-channel one-dimensional ultrasonic echo signal array.

5. The method for inspecting rail welds according to claim 1, characterized in that, The reconstruction generates a 3D voxel storage matrix and an initial C-scan 2D top view, specifically including: The signal intensity of the sampling points in the multi-channel one-dimensional ultrasound echo signal array is updated to the voxel cells of the corresponding three-dimensional voxel storage matrix using a strategy of overwriting the maximum value. The three-dimensional voxel storage matrix is ​​projected with the maximum amplitude along the depth direction to generate the signal intensity of the initial C-scan two-dimensional top view, and the exact depth corresponding to the signal intensity is retrieved. The spatial mapping model is used to calculate the corresponding pixel points of the two-dimensional grid points in the current visual frame to extract the two-dimensional semantic mask labels, normalized gray-level gradient magnitude, and local morphological linearity.

6. The method for inspecting rail welds according to claim 1, characterized in that, The extraction process of normalized gray-level gradient magnitude and local morphological linearity includes: using the horizontal and vertical Sobel gradient operators. and Calculate the normalized grayscale gradient magnitude : ； in, The extreme values ​​of the calibrated gray-scale gradient response; Construct a structure tensor matrix within the neighborhood of each pixel and compute its two non-negative eigenvalues. and And set Calculate local morphological linearity : ； in, To prevent extremely small constants with a denominator of zero.

7. The method for inspecting rail welds according to claim 1, characterized in that, Perform ultrasound shallow artifact suppression on the initial C-scan 2D top view, specifically including: When the signal strength of the initial C-scan 2D top view When the amplitude exceeds the effective alarm threshold and the label representation of the two-dimensional semantic mask matrix shows severe corrosion or oil contamination, the amplitude is determined based on the exact depth. Constructing empirical deep confidence weights : ； in, The confidence depth of the calibrated features; Suppression correction is applied to the initial C-scan 2D top view, and the corrected signal intensity is output in the inverse corrected C-scan spectrum matrix. : 。 8. The method for inspecting rail welds according to claim 1, characterized in that, The implementation of blind zone crack compensation and high-risk nuclear damage resonance confirmation specifically includes: When the two-dimensional semantic mask matrix label represents a clean surface medium, and the normalized gray-level gradient amplitude is greater than the preset crack threshold and the local morphological linearity is greater than the preset linearity threshold, the three-dimensional voxels between the ultrasonic near-surface blind zone limit depth and the maximum depth of the rail bottom surface are extracted, and the local maximum amplitude in the deep effective detection area is calculated. If the amplitude of the local maximum value is less than or equal to the effective amplitude threshold of the alarm, a fixed artificial warning amplitude is injected into the reverse correction C-scan spectrum matrix to perform blind spot missed detection compensation. If the amplitude of the local maximum exceeds the effective amplitude threshold of the alarm, the amplitude of the reverse-corrected C-scan spectrum matrix will be saturated at the top to perform high-risk resonance confirmation of penetrating nuclear damage.

9. The method for inspecting rail welds according to claim 1, characterized in that, Generating a two-dimensional semantic mask matrix specifically includes: The two-dimensional surface color image matrix is ​​converted to a color space that includes hue, saturation and lightness components, and the cyclic hue difference of the pixels is calculated by combining the periodic wrap-around characteristics of hue. A feature vector is constructed using the cyclic hue difference, saturation, and lightness components; the Mahalanobis distance of the feature vector belonging to a pre-calibrated Gaussian distribution model of various surface anomalies is calculated; When the Mahalanobis distance is less than the preset feature confidence threshold, a label representing the abnormal medium category is assigned to the corresponding pixel according to the minimum distance judgment principle to generate a two-dimensional semantic mask matrix.

10. A rail weld seam imaging inspection instrument, characterized in that, include: The scanning device is equipped with a three-axis detection linkage mechanism and a constant pressure coupling spring mechanism, and an incremental encoder that outputs position pulse signals is mechanically connected to the transmission wheel shaft on the same axis. An ultrasonic probe assembly, mounted on a triaxial detection linkage mechanism, includes probes that simultaneously execute a transmit-receive mode and a self-transmit and self-receive mode; The image acquisition module is fixedly mounted on the main frame of the scanning device, with its optical lens facing the detection surface and maintaining a pre-calibrated rigid physical relative position with the ultrasonic probe group in space; The ultrasound acquisition module has independent parallel high-speed signal acquisition channels, which are connected to the ultrasound probe group respectively; An embedded data processing platform is communicatively connected to an incremental encoder, an image acquisition module, and an ultrasonic acquisition module. The embedded data processing platform is configured with a first-in-first-out buffer queue and an asynchronous latest frame overlay buffer. The embedded data processing platform is used to execute the rail weld inspection method as described in any one of claims 1 to 9.

Images