An intelligent detection method and system for prefabricated building quality

Abstract

The present invention provides an intelligent detection method and system for prefabricated building quality, relating to the field of data processing technology. The method comprises: using a deep residual network to perform feature fusion analysis on ultrasonic echo signals to obtain a eigenvector corresponding to each signal; identifying welding defect types and labeling a set of defect spatial coordinates based on the eigenvectors and a time-frequency domain joint threshold segmentation algorithm to form a defect coordinate set containing four-dimensional information. The welding defect types include pores, lack of fusion, and cracks; calculating dynamic offset parameters of the defect coordinate set relative to the theoretical axis of the cantilever beam in a defect-reference fusion model; the parameters include the defect depth gradient, the lateral offset vector, and the stress wave attenuation coefficient; and inputting the dynamic offset parameters into a pretrained welding quality assessment model to predict the structural life. The present invention can more accurately assess the remaining useful life of prefabricated building structures.

Description

Technical Field

[0001] The present invention relates to the field of data processing technology, and in particular to an intelligent detection method and system for prefabricated building quality. Background Art

[0002] Prefabricated buildings are constructed from numerous prefabricated components assembled on-site. The quality of the connections between these components plays a decisive role in the safety and stability of the overall structure. Currently, traditional methods for inspecting the quality of prefabricated buildings have some limitations.

[0003] For example, during construction of a prefabricated residential complex, quality issues were discovered at the weld joints between some cantilever beam ends and balcony slabs. Traditional testing methods initially identified possible welding defects, but were unable to accurately determine the type, location, and severity of the defects. Further detailed testing confirmed the presence of a lack of fusion defect at the weld joints.

[0004] Lack of fusion primarily refers to the lack of fusion between the filler metal and the base metal. The main causes of this defect can be unclean grooves, excessive welding speeds, low welding currents, and improper electrode angles. In this case, investigation and analysis revealed that improper welding procedures and excessive welding speeds resulted in incomplete fusion in some areas. This lack of fusion severely impacted the strength and integrity of the weld joint, reducing the structure's bearing capacity and posing a significant risk to building safety. If such defects are not promptly identified and properly addressed, over time and as the building structure undergoes various loads, the affected area is highly likely to fracture, leading to serious safety incidents. Summary of the Invention

[0005] The technical problem to be solved by the present invention is to provide an intelligent detection method and system for the quality of prefabricated buildings, which can more accurately evaluate the remaining service life of prefabricated building structures.

[0006] In order to solve the above technical problems, the technical solutions of the present invention are as follows:

[0007] In a first aspect, a method for intelligently detecting the quality of prefabricated buildings is provided, the method comprising:

[0008] Collect ultrasonic echo signals from the welding nodes between the cantilever beam end and the balcony slab;

[0009] A reference spatial grid model is constructed based on the theoretical installation coordinates of the embedded parts. A three-dimensional defect distribution map of the welding interface is generated using the time-of-flight inversion algorithm of the ultrasonic echo signal. This map is then spatially aligned with the reference grid model to obtain a defect-reference fusion model.

[0010] Based on the spatial coordinate mapping relationship of the defect-reference fusion model, a deep residual network is used to perform feature fusion analysis on the ultrasonic echo signals to obtain the eigenvector corresponding to each signal. Based on the eigenvector and a joint threshold segmentation algorithm in the time and frequency domains, the welding defect types are identified and the defect spatial coordinate sets are marked to form a defect coordinate set containing four-dimensional information. The welding defect types include porosity, lack of fusion, and cracks.

[0011] Calculating dynamic offset parameters of the defect coordinate set relative to the theoretical axis of the cantilever beam in the defect-reference fusion model, the parameters including defect depth gradient, lateral offset vector and stress wave attenuation coefficient;

[0012] The dynamic offset parameters are input into a pre-trained weld quality assessment model to predict the structural life.

[0013] In a second aspect, an intelligent detection system for prefabricated building quality is provided, comprising:

[0014] An acquisition module is used to collect ultrasonic echo signals from the welding nodes between the cantilever beam end and the balcony slab;

[0015] The fusion module is used to construct a reference spatial grid model based on the theoretical installation coordinates of the embedded parts, generate a three-dimensional defect distribution map of the welding interface through the time-of-flight inversion algorithm of the ultrasonic echo signal, and perform spatial registration with the reference grid model to obtain a defect-reference fusion model;

[0016] A processing module is used to perform feature fusion analysis on ultrasonic echo signals using a deep residual network based on the spatial coordinate mapping relationship of the defect-reference fusion model to obtain a feature vector corresponding to each signal. The module then identifies the welding defect type and labels the defect spatial coordinate set based on the feature vector and a joint threshold segmentation algorithm in the time and frequency domains to form a defect coordinate set containing four-dimensional information. The welding defect types include porosity, lack of fusion, and cracks.

[0017] A calculation module, configured to calculate dynamic offset parameters of a defect coordinate set relative to a theoretical axis of the cantilever beam in a defect-reference fusion model, the parameters including a defect depth gradient, a lateral offset vector, and a stress wave attenuation coefficient;

[0018] A prediction module is used to input dynamic offset parameters into a pre-trained weld quality assessment model to predict the structure life.

[0019] The above solution of the present invention includes at least the following beneficial effects.

[0020] By collecting ultrasonic echo signals from the weld joint between the cantilever beam end and the balcony slab and applying a time-of-flight inversion algorithm to generate a three-dimensional defect distribution map of the weld interface, this method can accurately locate the position of welding defects in three-dimensional space. Compared with traditional detection methods, this method greatly improves the accuracy of defect detection and can effectively identify tiny defects, avoiding safety hazards caused by missed defects. For example, this method can clearly display the location and size of millimeter-level porosity defects, which were previously difficult to detect.

[0021] Using the theoretical installation coordinates of embedded components to construct a reference spatial grid model, spatially registering it with the 3D defect distribution map to obtain a fusion model. This model is then used for feature fusion analysis using a deep residual network. This series of operations automates and intelligentizes the inspection process. Compared to manual inspection, this significantly shortens inspection time and improves efficiency. It can complete the inspection of numerous welded joints in a short period of time, meeting the demands of rapid prefabricated construction. In large-scale prefabricated construction projects, it can save several times the inspection time and cost.

[0022] It can not only identify the types of welding defects, such as pores, lack of fusion and cracks, but also mark the defect spatial coordinate set to form a defect coordinate set containing four-dimensional information, and calculate dynamic offset parameters such as defect depth gradient, lateral offset vector and stress wave attenuation coefficient. These parameters provide rich data support for in-depth analysis of the impact of defects on structural performance, help to fully understand the characteristics of welding defects and their potential harm to the structure, and thus formulate more targeted repair plans. For example, the defect depth gradient can be used to intuitively judge the development trend of the defect in the depth direction.

[0023] By inputting dynamic offset parameters into a pre-trained welding quality assessment model to predict structural lifespan, this method changes the previous approach of determining structural lifespan based solely on experience or simple estimates. Based on actual inspection data and scientific model predictions, this method can more accurately assess the remaining useful life of prefabricated building structures. For example, for a prefabricated building structure with welding defects, predicting its remaining useful life using this method can provide a key reference for subsequent maintenance decisions, avoiding unnecessary maintenance work too early or too late, saving costs while ensuring building safety. BRIEF DESCRIPTION OF THE DRAWINGS

[0024] Figure 1 It is a flow chart of an intelligent detection method for prefabricated building quality provided by an embodiment of the present invention.

[0025] Figure 2 It is a schematic diagram of an intelligent detection system for prefabricated building quality provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0026] Exemplary embodiments of the present disclosure will be described in more detail below with reference to the accompanying drawings. Although exemplary embodiments of the present disclosure are shown in the accompanying drawings, it should be understood that the present disclosure can be implemented in various forms and should not be limited by the embodiments set forth herein. Rather, these embodiments are provided to enable a more thorough understanding of the present disclosure and to fully convey the scope of the present disclosure to those skilled in the art.

[0027] like Figure 1 As shown, an embodiment of the present invention provides an intelligent detection method for prefabricated building quality, the method comprising the following steps:

[0028] Collect ultrasonic echo signals from the welding nodes between the cantilever beam end and the balcony slab;

[0029] A reference spatial grid model is constructed based on the theoretical installation coordinates of the embedded parts. A three-dimensional defect distribution map of the welding interface is generated using the time-of-flight inversion algorithm of the ultrasonic echo signal. This map is then spatially aligned with the reference grid model to obtain a defect-reference fusion model.

[0030] Based on the spatial coordinate mapping relationship of the defect-reference fusion model, a deep residual network is used to perform feature fusion analysis on the ultrasonic echo signals to obtain the eigenvector corresponding to each signal. Based on the eigenvector and a joint threshold segmentation algorithm in the time and frequency domains, the welding defect types are identified and the defect spatial coordinate sets are marked to form a defect coordinate set containing four-dimensional information. The welding defect types include porosity, lack of fusion, and cracks.

[0031] Calculating dynamic offset parameters of the defect coordinate set relative to the theoretical axis of the cantilever beam in the defect-reference fusion model, the parameters including defect depth gradient, lateral offset vector and stress wave attenuation coefficient;

[0032] The dynamic offset parameters are input into a pre-trained weld quality assessment model to predict the structural life.

[0033] In an embodiment of the present invention, a three-dimensional defect distribution map of the welding interface is constructed by ultrasonic echo signal acquisition and time-of-flight inversion algorithm, and aligned with the reference spatial grid model, thereby achieving millimeter-level precision positioning of the defect position. Compared with traditional ultrasonic testing that can only provide two-dimensional cross-sectional information, this method can intuitively present the spatial distribution morphology of defects within the welding node (such as the three-dimensional volume of pores and the extension path of cracks). Combined with the reference model constructed based on the theoretical coordinates of embedded parts, it can accurately identify the offset error of the defect relative to the design axis, providing a quantitative basis for the spatial position of the defect for defect hazard assessment. For example, for a microcrack with a length of 2mm in a certain project, its specific coordinates in the X / Y / Z axis direction and the angle with the theoretical axis can be accurately marked to avoid structural safety hazards caused by misjudgment of the defect position.

[0034] By using a deep residual network (ResNet) to perform feature fusion analysis on ultrasonic signals and combining it with a time-frequency domain joint threshold segmentation algorithm, the problem of traditional manual interpretation relying on experience and being prone to omissions and misjudgments is resolved. The model can automatically extract the signal's multi-dimensional features in the time domain waveform, frequency domain energy distribution, and time-frequency joint domain (such as stress wave attenuation rate and frequency offset), enabling intelligent classification of three typical defects: pores, lack of fusion, and cracks. The recognition accuracy is improved by more than 30% compared to traditional methods. Taking lack of fusion defects as an example, the algorithm can effectively distinguish between local lack of fusion caused by insufficient welding current and interface separation caused by the oxide film of the parent material by analyzing the reflected energy mutation characteristics and phase offset characteristics of the signal at the defect interface, thus avoiding misjudging functional interfaces as defects.

[0035] By calculating dynamic parameters such as the defect depth gradient, lateral offset vector, and stress wave attenuation coefficient, a mapping relationship between the spatial location of the defect and the mechanical properties of the structure is established. The depth gradient reflects the tendency of the defect to extend along the thickness of the component (e.g., the rate at which a crack propagates toward the core load-bearing zone of a cantilever beam), the lateral offset vector quantifies the degree to which the defect deviates from the theoretical load-bearing axis, and the stress wave attenuation coefficient characterizes the defect's impact on the vibration transmission characteristics of the structure. During an actual engineering inspection, this parameter system revealed that the stress wave attenuation coefficient of a balcony slab welded joint exceeded a threshold. Combined with finite element simulation verification, it was confirmed that the defect caused a 15% decrease in joint stiffness, providing data support for timely reinforcement measures.

[0036] In a preferred embodiment of the present invention, collecting ultrasonic echo signals of the welding nodes between the cantilever beam ends and the balcony slabs includes:

[0037] A high-frequency piezoelectric ceramic ultrasonic probe with a center frequency of 5-10MHz is used to match the acoustic impedance characteristics of the welded joint base material (Q345B steel). The probe wafer has a diameter of φ6mm to balance resolution and penetration. To address the complex geometric curves of the cantilever beam and balcony slab weld joint, a custom curved wedge probe with an 80mm radius of curvature is used to ensure a coupling area of ≥90% with the test surface (weld reinforcement ≤3mm), reducing boundary reflection noise. A multi-channel ultrasonic data acquisition system supports 16 channels of simultaneous acquisition, a sampling rate of 200MHz, a dynamic range of 120dB, and an integrated high-precision positioning module (accuracy ±0.1mm). A laser rangefinder is used to calibrate the distance between the probe and the test surface in real time, ensuring a sound path measurement error of less than 0.5%.

[0038] Use sandblasting (abrasive is 80 mesh aluminum oxide) to remove the oxide scale and spatter on the surface of the welding joint, and control the roughness to Ra ≤ 6.3μm. Use anhydrous ethanol to wipe the test area to remove oil and dry it. Ensure that the coupling agent (silicone-based gel, acoustic impedance 1.5×10 6 Apply Rayl) evenly with a thickness of 0.2-0.5mm to avoid bubbles.

[0039] The theoretical installation coordinates of the embedded parts were collected using a total station (Leica TS60, angular measurement accuracy 0.5″). A right-handed Cartesian coordinate system was established with the cantilever beam axis as the Z axis. Reflective markers with a diameter of 5 mm were affixed to the inspection surface as a spatial positioning reference, and a millimeter-level precision inspection grid (grid spacing 5 mm × 5 mm) was constructed.

[0040] A zigzag linear scanning pattern was used, with a stepping interval of 2mm along the length of the weld (Z axis) and a 20mm area on each side of the weld (including the heat-affected zone) in the transverse direction (X axis). For complex areas of fillet welds, a 45° oblique incidence scan (probe tilt angle α = 30° / 60°) was added to stimulate shear and longitudinal wave modes, covering areas sensitive to lack of fusion defects. Acquisition parameter settings:

[0041] Excitation mode: pulse echo method, emission voltage 150V, pulse width 0.1μs;

[0042] Gain control: Initial gain 60dB, combined with real-time A-scan waveform dynamic adjustment to ensure that the bottom echo amplitude reaches 80% of the full screen;

[0043] Synchronous triggering: The encoder triggers the acquisition, and one group of signals (including 1024 sampling points) is synchronously collected every 0.5mm of movement;

[0044] Filter setting: 5-15MHz bandpass filter, suppressing low-frequency noise and high-frequency clutter.

[0045] During the acquisition process, the A-scan waveform is displayed in real time, with key monitoring: initial wave width (≤0.5μs) to determine coupling stability; defect wave amplitude fluctuation (≤5%) to avoid signal attenuation differences caused by probe pressure changes; bottom echo signal-to-noise ratio (≥20dB) to ensure effective penetration of the weld fusion zone; when the signal-to-noise ratio of three consecutive scanning points is less than 15dB, the surface reprocessing process is automatically triggered (re-apply coupling agent or adjust the probe angle). System calibration and calibration:

[0046] Using the CSK-IA standard test block (sound velocity 5900m / s), the instrument sound velocity parameter was corrected to 5880±10m / s (taking into account the influence of the field temperature of 25±2℃) by measuring the echo sound path of a 20mm / 50mm flat-bottom hole.

[0047] For probe delay calibration, a 10mm thick comparison test block is used to measure the bottom surface echo time difference and correct the probe front distance to within 0.3mm.

[0048] Sensitivity calibration increases the echo of a φ2mm long horizontal hole to 80% of the full screen, which serves as the benchmark sensitivity for defect detection. Data collection and time-space correlation recording:

[0049] Each detection point is synchronously recorded: the original time domain signal (1024 points / channel, including amplitude and time parameters); spatial coordinates (X, Y, Z, accuracy ±0.1mm); detection timestamp (accurate to milliseconds); environmental parameters (temperature 25±2°C, humidity 40±5%RH); a one-to-one mapping relationship is established between each A-scan signal and the spatial coordinates to generate a signal matrix file with geographic coordinates. At the same time, equipment parameters such as probe number and coupling agent type are recorded to form a traceable original detection data set. For ease of use, the ultrasonic echo signal needs to be preprocessed and standardized before specific application.

[0050] In a preferred embodiment of the present invention, a reference space grid model is constructed based on the theoretical installation coordinates of the embedded parts, including:

[0051] Extract the theoretical coordinates of embedded parts from architectural design drawings, including the coordinates of each embedded part in three-dimensional space. Based on the theoretical coordinates of the embedded parts, determine the spatial range of the grid to be constructed. This involves directly reading the three-dimensional coordinate information of each embedded part from the architectural design drawings (such as BIM models or CAD drawings). The coordinates are marked in millimeters and include the specific location of the embedded part's geometric center in space (X, Y, Z). For example, if the center coordinates of an embedded part are (1500, 800, 200) obtained from the drawing annotation, this means that the embedded part is 1500mm from the origin along the cantilever beam length (X-axis), 800mm along the balcony slab width (Y-axis), and 200mm along the weld interface thickness (Z-axis). If there are multiple embedded parts, record the coordinates of each embedded part one by one to form a coordinate set (e.g., embedded part A: 1200, 600, 180; embedded part B: 1800, 1000, 220, etc.). Determine the spatial range of the grid:

[0052] X-axis range: Traverse the X-coordinates of all embedded parts and find the minimum value (such as 1200mm) and the maximum value (such as 1800mm). The final X-axis range is determined from the minimum value to the maximum value, that is, 1200mm to 1800mm, to ensure that the distribution of all embedded parts in the cantilever beam length direction is covered.

[0053] Y-axis range: Similarly, extract the minimum value (such as 600mm) and maximum value (such as 1000mm) of the Y coordinate, and determine the Y-axis range as 600mm to 1000mm, covering the distribution of embedded parts in the width direction of the balcony slab.

[0054] Z-axis range: Considering the design thickness of the welding interface (such as the theoretical thickness is 100mm), with the average value of the embedded part's Z coordinate (such as 200mm) as the center, expand half of the thickness (±50mm) to both sides. The final Z-axis range is 150mm to 250mm to ensure that the full thickness area of the welding interface is included.

[0055] The grid cell size is determined based on the ultrasonic testing resolution and the spatial extent of the grid. A three-dimensional rectangular coordinate system is established within the testing area, using the theoretical coordinates of the embedded component as a reference. The origin is the geometric center of the embedded component, the X-axis is along the length of the cantilever beam, the Y-axis is along the width of the balcony slab, and the Z-axis is along the thickness of the weld interface. Specifically, based on the ultrasonic testing equipment's resolution (e.g., 0.5mm vertically and 1mm horizontally) and computational efficiency, a grid cell is selected as a cube with a side length of 5mm. This size is an integer multiple of the resolution, ensuring testing accuracy while avoiding data redundancy. For example, if the testing area extends 600mm in the X-axis (1800-1200 mm), the X-axis can be divided into 120 grid cells (600mm ÷ 5mm / cell). The geometric center of a single embedded component is used as the origin. (For multiple embedded components, the average coordinate of the geometric centers of all embedded components is used as the origin.) For example, if the coordinates of two embedded parts are (1200, 600, 180) and (1800, 1000, 220), then the X coordinate of the origin is (1200 + 1800) ÷ 2 = 1500 mm, the Y coordinate is (600 + 1000) ÷ 2 = 800 mm, and the Z coordinate is (180 + 220) ÷ 2 = 200 mm, that is, the origin is (1500, 800, 200). Coordinate axis definition:

[0056] X-axis: along the length of the cantilever beam, starting from the origin and pointing to the cantilever end of the cantilever beam as the positive direction (for example, from the side of the fixed end connected to the main structure to the side of the balcony slab at the cantilever end).

[0057] Y-axis: perpendicular to the X-axis, along the width of the balcony slab, centered at the origin, and extending to both sides of the balcony slab (for example, the left side is the negative direction and the right side is the positive direction).

[0058] Z axis: perpendicular to the welding interface, along the thickness direction, with the origin as the center, pointing from the cantilever beam base material to the balcony plate base material as the positive direction (for example, upward is positive and downward is negative).

[0059] Starting from the origin, grid lines are evenly divided along the X-axis, Y-axis, and Z-axis to form a regular cubic grid. Specifically, starting from the X coordinate of the origin (such as 1500mm), grid lines are drawn in the positive and negative directions at intervals of 5mm. For example, the positive direction of the X axis is 1505mm, 1510mm... until the maximum range of the X axis (1800mm); the negative direction is 1495mm, 1490mm... until the minimum range of the X axis (1200mm). Each grid line represents a position node along the length of the cantilever beam, and the spacing between adjacent grid lines is 5mm. Division in the Y-axis direction:

[0060] Centered on the Y coordinate of the origin (e.g. 800mm), grid lines are divided at 5mm intervals on both sides of the balcony slab. For example, the positive direction of the Y axis is 805mm, 810mm... to the maximum Y coordinate of 1000mm; the negative direction is 795mm, 790mm... to the minimum Y coordinate of 600mm, covering the entire detection area along the width of the balcony slab. Division in the Z axis direction:

[0061] Along the thickness of the weld interface, with the origin Z coordinate (e.g., 200mm) as the center, grid lines are divided at 5mm intervals upward (toward the balcony slab base material) and downward (toward the cantilever beam base material). For example, the positive Z-axis direction is 205mm, 210mm, and so on to 250mm; the negative Z-axis direction is 195mm, 190mm, and so on to 150mm, ensuring that the full thickness of the weld interface (100mm) is covered.

[0062] The grid lines of the X, Y, and Z axes intersect in space to form a cubic grid. Each intersection is a grid point, whose coordinates are formed by combining the coordinates of the grid lines of the three axes (for example, X = 1500mm, Y = 800mm, Z = 200mm is the grid point at the origin; X = 1505mm, Y = 805mm, Z = 205mm are adjacent grid points). All grid points form a regular three-dimensional grid matrix that covers the entire detection area.

[0063] The theoretical geometric features of the weld nodes are embedded in the cubic grid to form a reference spatial grid model containing the coordinates and theoretical geometric features of all grid points, including the theoretical position of the weld interface, the theoretical axis of the cantilever beam, and the theoretical outline of the embedded parts. Specifically, the theoretical range of the weld interface on the Z axis is determined according to the design drawings (for example, for a thickness of 100mm, Z = 200 ± 50mm). In the grid model, all grid points with Z coordinates between 150mm and 250mm are marked as the "weld interface area," and the theoretical thickness boundary of this area is recorded (Z = 250mm on the upper surface and Z = 150mm on the lower surface) as a basis for determining whether subsequent defects are located within the weld interface.

[0064] Mark the theoretical axis of the cantilever beam: The theoretical axis of the cantilever beam is a straight line passing through the origin and along the X-axis. This refers to the set of grid points (e.g., X = 1200, Y = 800, Z = 200; X = 1300, Y = 800, Z = 200) that satisfy the conditions Y = the Y coordinate of the origin (e.g., 800 mm) and Z = the Z coordinate of the origin (e.g., 200 mm). In the grid model, mark these points as the "theoretical axis of the cantilever beam" to serve as the reference line for calculating the defect's lateral offset vector (e.g., the distance the defect center deviates from the axis).

[0065] Theoretical outline of embedded parts:

[0066] Rectangular embedded part: If the design dimensions of the embedded part are 200mm long, 100mm wide, and 80mm high (centered at the origin), the range of its outline in the grid is:

[0067] X-axis: 1500-100mm to 1500+100mm (1400mm to 1600mm);

[0068] Y-axis: 800-50mm to 800+50mm (750mm to 850mm);

[0069] Z-axis: 200-40mm to 200+40mm (160mm to 240mm).

[0070] Mark the grid points within the above range as "embedded part outline".

[0071] Cylindrical embedded parts: If the embedded part has a diameter of 60mm and a height of 100mm, the radius of the bottom circle in the XY plane with the origin as the center is 30mm (that is, it meets (X-1500) 2 +(Y-800) 2 ≤30 2 ), the Z-axis range is 200-50mm to 200+50mm (150mm to 250mm), and the grid points that meet the conditions are marked as "embedded part outline".

[0072] The coordinates (X, Y, Z) of all grid points are associated with their corresponding theoretical geometric features (such as whether they belong to the weld interface, cantilever beam axis, or embedded component outline), forming a reference spatial grid model that contains spatial coordinates and engineering semantics. For example, in addition to recording the coordinates, each grid point is also accompanied by an attribute label (such as "weld interface - upper surface," "cantilever beam axis point," "embedded component internal point," etc.), providing a geometric reference for subsequent registration of the defect map with the reference model.

[0073] In an embodiment of the present invention, by extracting the theoretical coordinates of the embedded parts in the architectural design drawings (with an accuracy of ±0.5mm), a three-dimensional rectangular coordinate system is established with the geometric center of the embedded parts as the origin, so that the detection data and the design benchmark are spatially aligned at the millimeter level. Compared with traditional detection that relies on manual measurement of reference points (with an error of more than ±2mm), this method eliminates the coordinate system conversion error and ensures the absolute consistency of the defect spatial coordinates with the design drawings. For example, in a certain project, the theoretical coordinates of the embedded parts are (1000, 500, 300) mm. During actual detection, the spatial positioning of all ultrasonic signals is based on this origin, avoiding the misjudgment of the defect position due to benchmark drift. According to the ultrasonic detection resolution (0.5mm longitudinally and 1mm horizontally), a 5mm×5mm×5mm cubic grid unit is set, which not only meets the sampling density requirements of signal processing but also avoids redundant data. After embedding geometric features such as the theoretical position of the welding interface, the theoretical axis of the cantilever beam, and the theoretical contour of the embedded parts, a benchmark model with engineering semantics is formed:

[0074] The theoretical position of the welding interface accurately defines the detection range of the fusion zone (e.g., ±15mm in the thickness direction) to avoid scanning invalid areas. The theoretical axis of the cantilever beam serves as the core reference line for dynamic offset parameter calculations, directly linking to structural force analysis (e.g., the lateral offset vector takes the axis as the origin). The theoretical contour of embedded parts automatically identifies key detection areas (e.g., anchor bar welding sections) to optimize scanning path planning.

[0075] By mapping the 3D defect distribution map to the coordinates of a benchmark mesh model (with an error of ≤0.3mm), parameters such as the depth gradient of the defect relative to the theoretical axis (e.g., the crack growth rate per 10mm of depth along the Z axis) and the lateral offset (e.g., a pore center offset of 2.5mm from the Y axis) can be accurately calculated, providing a quantitative basis for structural mechanics analysis. The grid point coordinates serve as a unified data interface, enabling subsequent deep residual network analysis to correlate ultrasonic signal characteristics (e.g., stress wave amplitude) with theoretical coordinates (e.g., the X-axis force direction). For example, identifying a continuous lack of fusion defect along the length (X-axis) of a cantilever beam can automatically trigger a stress concentration warning. The benchmark model's meshing results directly drive the robotic arm to scan along a pre-set path (e.g., 5mm steps along the X-axis and 2mm layers along the Z-axis), avoiding blind spots caused by manual operation and improving inspection efficiency by 40%. The embedded theoretical geometric features serve as prior knowledge, assisting a joint time-frequency domain threshold segmentation algorithm in eliminating non-defect signals (e.g., pseudo-defect waves generated by reflections from the edges of embedded component contours). For example, when a signal is detected whose Z-axis coordinates exceed the theoretical range of the weld interface thickness (±20mm), the system automatically identifies it as external noise from the parent material, reducing the misjudgment rate by 60%. The system compares the detected defect coordinates with the theoretical coordinates of the embedded part in real time, automatically generating a deviation report (for example, triggering a red alert when the lateral offset vector of a node exceeds 5mm). The system couples the historical defect coordinate set (including four-dimensional information) with structural load data for analysis. For example, using the theoretical axis of a cantilever beam in the baseline model, it accurately calculates the coefficient of the defect's impact on structural stiffness (for example, for every 10% increase in the stress wave attenuation coefficient, the stiffness decreases by 5%).

[0076] In a preferred embodiment of the present invention, a three-dimensional defect distribution map of the welding interface is generated by a time-of-flight inversion algorithm of the ultrasonic echo signal, and spatially aligned with the reference grid model to obtain a defect-reference fusion model, including:

[0077] The probe is attached to the surface of the cantilever beam, emitting high-frequency ultrasonic pulses toward the welding interface and receiving reflected echo signals. Each scan records the echo signal of a position, including the signal amplitude and flight time. Specifically, the high-frequency ultrasonic probe (such as an arc-shaped wedge probe with a center frequency of 5-10MHz) is stably attached to the cantilever beam detection surface through a magnetic fixture or a robotic arm. The detection area is polished in advance to a roughness of Ra ≤ 6.3μm, and a silicone-based coupling agent (thickness 0.2-0.5mm) is applied to ensure that the sound beam is incident vertically on the welding interface. The probe emits high-frequency ultrasonic pulses (pulse width 0.1μs) toward the welding interface and synchronously receives defect reflection echoes and bottom surface echo signals. Each scan records the A-scan signal of a position, including the time domain waveform (amplitude-time curve), flight time (the time point corresponding to the defect echo peak) and the real-time position coordinates of the probe (obtained through the integrated high-precision positioning module, with an accuracy of ±0.1mm); using the "point-by-point stepping and row and column scanning" mode, the step spacing along the cantilever beam length direction (X-axis) is 2mm, and the horizontal (Y-axis) covers 20mm on both sides of the weld. After each layer is scanned, it is increased by 1mm in the thickness direction (Z-axis) until the entire weld interface thickness is covered (such as the designed thickness of 100mm).

[0078] According to the propagation speed of ultrasound in the material, the flight time of the echo signal is converted into the distance between the defect and the probe. Specifically, before testing, the propagation speed of ultrasound in the parent material (such as Q345B steel) is measured using a CSK-IA standard test block. The sound speed at room temperature is approximately 5880m / s. Combined with the on-site temperature (±2°C), the accuracy is corrected to ±10m / s. For the flight time t (unit: μs) of the defect wave in the echo signal, the distance d from the defect to the probe is calculated using the formula d = (v×t) / 2 (v is the sound speed, divided by 2 for round-trip signal propagation).

[0079] According to the actual position and beam angle during scanning, the coordinates of the defect in three-dimensional space are calculated. The defect coordinates obtained from each scan are used as a data point. After multiple scans, point cloud data containing the defect points are formed. Specifically, the three-dimensional coordinates (Xp, Yp, Zp) of the probe center are recorded in real time by a robotic arm encoder or a total station as the scanning position reference. If an oblique incidence probe (such as a 30° / 60° shear wave probe) is used, the coordinates of the defect in three-dimensional space are calculated according to the beam incidence angle θ: the distance along the direction of the sound beam is d, and the distance perpendicular to the sound beam is θ. The thickness direction coordinate of the detection surface is Z = Zp + d × cosθ; the lateral offset coordinate is X = Xp + d × sinθ × cosϕ; Y = Yp + d × sinθ × sinϕ (ϕ is the deflection angle of the probe in the horizontal plane, and the default is 0° along the X-axis direction). Each scan obtains a defect data point (X, Y, Z). If there is no defect at this position, it is marked as a background point. After multiple scans, a three-dimensional point cloud containing thousands to tens of thousands of points is formed, among which the defect points are concentrated in the welding interface area (within the range of ±50mm on the Z axis).

[0080] The point cloud data is mapped to a three-dimensional spatial grid and divided into cubic voxels according to a preset accuracy. The grayscale value of each voxel represents the defect probability of the corresponding area, ultimately forming a three-dimensional defect distribution map. Specifically, the following steps are performed: using the coordinate system of the baseline grid model as a reference, the inspection area is divided into cubic voxels with a preset accuracy (such as 5mm×5mm×5mm). The voxel size must match an integer multiple of the ultrasonic inspection resolution (0.5mm vertically and 1mm horizontally); each defect point (X, Y, Z) is mapped to the corresponding voxel unit, and the number of defect points in each voxel is counted. If the number of defect points in a voxel exceeds a threshold (such as 3), the voxel is marked as a "suspected defect area"; and each voxel is assigned a grayscale value of 0-255 based on the defect point density or signal amplitude mean within the voxel (for example, the higher the defect probability, the higher the grayscale value). For example, the grayscale value of a defect-free voxel is 0, and the grayscale value of a high-density defect voxel is ≥200. Finally, a three-dimensional grayscale map containing the spatial distribution of defects is generated, which can be displayed in three-dimensional visualization through software (such as the projection of defects in the XY plane and the Z-axis depth slice).

[0081] The four corner points of the embedded component and the edge line of the weld interface are extracted from the reference spatial grid model and the 3D defect distribution map. Specifically, the theoretical coordinates of the four corner points of the embedded component (e.g., the corner coordinates of a rectangular embedded component are (X0±a / 2, Y0±b / 2, Z0±c / 2)) and the theoretical edge line of the weld interface (e.g., the XY plane contour line at Z=Z0±t / 2)) are directly retrieved from the reference spatial grid model. Edge detection is then performed on the 3D defect map, and the measured corner points of the embedded component (where the signal amplitude suddenly changes) and the edge line of the weld interface (the boundary of the defect-dense area) are identified using threshold segmentation (e.g., grayscale value ≥ 150). For example, boundary points with sudden grayscale drops are found in the map and fitted into a straight line or curve as the measured edge line.

[0082] The coordinates of the embedded part corner points are matched with the theoretical corner point coordinates in the reference grid, and preliminary translation and rotation angles are calculated to align the overall positions of the two. The distance from each measured defect point to the nearest theoretical point in the reference grid is calculated as the error. The position and posture of the 3D defect distribution map are gradually adjusted until the overall error is less than the set threshold to form a registered defect map. Specifically, the following steps are performed: matching the measured embedded part corner point coordinates with the theoretical corner points of the reference model, calculating preliminary translation (△X = X measured − X theoretical, △Y, △Z) and rotation angles (Euler angles around the X / Y / Z axes to roughly align the measured and theoretical corner points in space); using the iterative closest point (ICP) algorithm, calculating the distance from each measured defect point to the nearest theoretical point in the reference grid as the error, and gradually adjusting the position (translation) and posture (rotation) of the defect map using the least squares method. For example, the error is reduced by 10% in each iteration until the overall root mean square error is less than the set threshold (such as 0.5mm), ensuring that the coordinate deviation between the defect point cloud and the reference grid is within the millimeter range; after the alignment is completed, 10 reference grid points are randomly selected to verify the distance from the measured defect point to the corresponding theoretical point. If the error of more than 90% of the points is less than 1mm, the alignment is considered qualified.

[0083] The registered defect map is embedded into the reference grid model, aligning the coordinates of each defect point with the spatial location of the reference grid. This results in a defect-reference fusion model that incorporates both theoretical geometric features and measured defect data. Specifically, the coordinates of each defect point in the registered defect map are mapped to the coordinate system of the reference grid, for example, mapping the defect point coordinates (1502, 803, 198) to the voxel containing the nearest grid point (1500, 800, 200). Theoretical geometric attributes from the reference model are then added to each defect point, such as whether it lies within the embedded component outline (by determining whether the coordinates are within the theoretical dimension range of the embedded component), its distance from the theoretical axis of the cantilever beam, and whether it is within the theoretical thickness range of the weld interface (±50 mm on the Z axis). The grayscale value data (defect probability) from the 3D defect map is then combined with the coordinates and theoretical features of the reference grid to form a fusion model that includes the measured defect location, spatial distribution density, and theoretical geometric references.

[0084] In an embodiment of the present invention, a time-of-flight inversion algorithm is used to convert the ultrasonic signal's time of flight (with an accuracy of 0.1μs) into a defect distance (with an error of ≤0.3mm). The three-dimensional coordinates are calculated by combining the scanning position (with a coordinate accuracy of ±0.1mm) and the beam angle (with a resolution of 1°), achieving millimeter-level localization of the defect location. Compared to traditional A / B scanning, which only displays two-dimensional cross-sectional defects, this method generates point cloud data containing X / Y / Z coordinates. After voxelization, a three-dimensional map is formed, corresponding grayscale values to defect probabilities. This intuitively displays the three-dimensional distribution of pores (e.g., the depth distribution of a 5mm-diameter spherical pore in the Z direction) and the spatial orientation of cracks (e.g., an unfused crack extending at a 45° angle along the weld interface). During the inspection of a prefabricated bridge component, a linear unfused defect measuring 12mm deep from the surface and 8mm long along the Y axis was successfully located. Traditional methods only determine the presence of a defect but cannot accurately describe its spatial configuration.

[0085] By matching the embedded component corners (theoretical coordinate accuracy ±0.5mm) with the measured corners, the translation (accuracy ±0.2mm) and rotation angle (±0.5°) are calculated, solving the problem of inconsistency between measured data and design drawing benchmarks in traditional inspections. After registration, the average distance error between the defect point and the theoretical point of the benchmark grid is less than 0.8mm, ensuring that the spatial coordinates of defects such as pores and cracks are directly related to the theoretical axis of the cantilever beam and the theoretical position of the welding interface. For example, in the inspection of a balcony slab weld node, the measured crack starting point coordinates (1502, 805, 195) were aligned, and its lateral offset (Y-axis +5mm) and depth gradient (Z-axis -5mm) relative to the theoretical axis could be accurately calculated, providing direct data for evaluating the impact of defects on structural stress.

[0086] In a preferred embodiment of the present invention, based on the spatial coordinate mapping relationship of the defect-reference fusion model, a deep residual network is used to perform feature fusion analysis on the ultrasonic echo signal to obtain the feature vector corresponding to each signal, including:

[0087] The original ultrasonic echo signal is acquired. Each signal corresponds to a detection position and includes a curve showing the amplitude variation over time. Specifically, the probe is stably attached to the cantilever beam detection surface via a magnetic fixture, and position control is achieved with a robotic arm or manual scanning device (positioning accuracy ±0.1mm). Each scan triggers the probe to emit a high-frequency ultrasonic pulse (duration 0.1μs). The receiver synchronously collects the echo signal to form A-scan data, which is a curve showing the amplitude variation over time (sampling rate 200MHz, a single signal contains 1024 time points, and the amplitude voltage value is recorded at each point). The three-dimensional coordinates (X, Y, Z) of the current probe center are recorded in real time using the device's integrated encoder or total station as the detection position corresponding to the signal, ensuring that each echo signal is uniquely bound to a spatial coordinate (for example, the coordinates of a signal are X = 1500mm, Y = 800mm, Z = 200mm).

[0088] Perform frequency domain analysis on the ultrasonic echo signal to generate a time-frequency diagram and extract the energy distribution characteristics of the signal in different frequency bands. The horizontal axis of the time-frequency diagram is time and the vertical axis is frequency. Specifically, the original time domain signal is divided into overlapping time windows (such as a window length of 20μs and an overlap rate of 50%), and the signal in each window is Fourier transformed to obtain the frequency components and amplitudes corresponding to different time points; square the amplitude of each frequency point to obtain the energy value of the frequency band, forming a two-dimensional time-frequency diagram with time (unit: μs) on the horizontal axis and frequency (range: 0-20MHz) on the vertical axis. The brightness or color of each pixel represents the energy intensity of the corresponding time-frequency point (for example, the energy concentration area of 5-10MHz in the high frequency band may indicate a lack of fusion defect); extract the energy statistical characteristics of different frequency bands, such as:

[0089] The energy ratio in the low-frequency band (0-5MHz) reflects the overall acoustic properties of the base material. The energy attenuation rate in the mid- and high-frequency bands (5-15MHz) is an indicator of high-frequency energy loss caused by defects (such as cracks). The peak frequency offset is compared with the reference frequency of the defect-free signal to determine whether there is an interface reflection anomaly.

[0090] The detection position coordinates corresponding to each signal are integrated into the ultrasonic echo signal, time-frequency map, and position coordinates as an input tensor. Specifically, the original amplitude-time curve is normalized (mean subtracted and divided by standard deviation) and padded or truncated to a uniform length of 1024 time points to form a one-dimensional array (shape: 1024,); the two-dimensional time-frequency map is adjusted to a fixed size (e.g., 64×64 pixels), and the pixel values are normalized to the range [0, 1] as a two-dimensional matrix input (shape: 64, 64); the three-dimensional coordinates (X, Y, Z) of the detection position are converted to floating-point numbers and scaled according to the coordinate range of the reference grid model (e.g., if the X-axis range is 1000-2000 mm, the X coordinate is divided by 2000 and normalized to [0.5, 1]) to form a three-dimensional vector (shape: 3,); the time domain array, time-frequency map matrix, and position vector are combined into a multidimensional input tensor. The specific structure is:

[0091] The channel dimension includes 1 time domain channel, 1 time-frequency channel (64, 64), and 1 position channel; the final tensor shape is: (1024, 64, 64, 3), corresponding to the multimodal information fusion of time points, frequency points, and spatial coordinates respectively.

[0092] Based on the input tensor and the trained deep residual network, a comprehensive feature vector containing time-frequency domain signal features and spatial position information is generated, specifically including: receiving multi-dimensional tensors and adapting the three-channel data of time domain, frequency domain and position; using 1D convolution layer (kernel size 5) to extract local features of the time domain signal (such as echo peak time, rising edge slope); processing the time-frequency map through 2D convolution layer (kernel size 3×3) to capture the texture pattern of energy distribution (such as the linear energy attenuation band of cracks and the circular energy concentration area of pores); mapping the three-dimensional coordinates into a high-dimensional feature vector (such as 64 dimensions) through the fully connected layer, and splicing it with the feature map output by the convolution layer; introducing skip connections in the deep network to directly transfer shallow layer features to deep layers to avoid gradient disappearance, for example, adding residual blocks between the 3rd and 6th layers to maintain the integrity of feature propagation; performing average pooling on the convolution feature map and merging it with the position feature vector to form a comprehensive feature vector containing multimodal information (such as 256 dimensions).

[0093] Using labeled ultrasonic signal data (including samples of pores, lack of fusion, cracks, and no defects), we split the training set into a validation set in an 8:2 ratio. We used a cross-entropy loss (for defect classification tasks) and the Adam optimizer (learning rate 0.001). We iterated for 50-100 epochs, monitoring the validation set accuracy (target >95%) and using early stopping to avoid overfitting. The input tensor was forward propagated through the network, passing through convolutional blocks, residual blocks, and pooling layers in sequence, ultimately outputting a fixed-dimensional feature vector (e.g., 256 dimensions) at the fully connected layer. Each dimension in the feature vector corresponds to a different fused feature, for example:

[0094] Dimensions 1-64: waveform details of the time domain signal (such as the interval between multiple reflection echoes);

[0095] Dimensions 65-192: Energy distribution pattern in the time-frequency graph (e.g., the degree of energy attenuation in a specific frequency band);

[0096] Dimensions 193-256: spatial position-related features (such as the distance of the defect from the cantilever beam axis and its relative position in the embedded parts).

[0097] In an embodiment of the present invention, by integrating the original time domain signal (amplitude-time curve), frequency domain energy distribution (time-frequency diagram) and spatial position coordinates (X / Y / Z axes), a three-dimensional input tensor containing the signal's physical characteristics and structural geometric information is constructed, which solves the problem of one-sided information in traditional single-modal analysis (only time domain or frequency domain).

[0098] For example: time domain signal: captures sudden changes in the amplitude of defect reflection waves (such as multiple reflection echoes caused by cracks); frequency domain features: identifies energy attenuation in specific frequency bands (such as the energy loss rate of unfused defects in the 5-8MHz frequency band reaches 30%); spatial coordinates: locates defects in the key stress area of the cantilever beam (such as defects within 5mm from the theoretical axis triggering key warnings); the fusion of the three enables the model to extract the composite features of "cracks at 150mm from the cantilever end and sudden energy drops in the 3-6MHz frequency band", which improves the recognition accuracy by 25% compared with traditional single signal analysis. Residual connections are used to solve the vanishing gradient problem of deep networks, enabling effective training of deep networks with more than 18 layers. This approach can capture deep nonlinear features in ultrasonic signals: Low-level features: extract basic features such as the slope of the rising edge of the signal and the time difference between the peaks through convolutional layers; mid-level features: capture the texture pattern of energy distribution in the time-frequency graph (such as the circular energy concentration area of pores and the linear energy attenuation band of cracks); high-level features: combine spatial coordinates to learn the correlation pattern between defect distribution and structural stress (for example, the influence of unfused defects distributed along the axis of a cantilever beam on stiffness is increased by 40%).

[0099] The detection position coordinates (accuracy ±0.1mm) are used as part of the input tensor, so that the feature vector contains the "structural semantic" information of the defect: Embedded parts association: Automatically identify defects located within the outline of the embedded parts (such as pores at the anchor bar welding, whose spatial coordinates trigger the embedded parts corrosion risk model); Axis offset quantification: Calculate the distance between the defect center and the theoretical axis of the cantilever beam through coordinates (for example, when the lateral offset vector is greater than 10mm, the weight of the corresponding dimension in the feature vector increases); Depth gradient encoding: The Z-axis coordinate is combined with the theoretical thickness of the welding interface to distinguish between surface defects (Z < 5mm) and internal defects (Z > 20mm). The influence coefficient of the former on fatigue life is automatically multiplied by a 1.5 times correction factor.

[0100] Traditional detection relies on engineers to manually extract signal features (such as peak amplitude and number of cycles), which is highly subjective and inefficient. This method, through a data-driven deep learning model, fully automates the entire process from feature extraction to fusion. Time-frequency plots are automatically generated based on the short-time Fourier transform (STFT), avoiding the errors associated with manually selecting window functions. Feature vector dimensions are unified: regardless of signal length, a fixed-dimensional feature vector (e.g., 256 dimensions) is output through the residual network pooling layer, facilitating subsequent threshold segmentation algorithms. Multi-scale feature fusion is achieved by using different convolution kernel sizes (3×3, 5×5) to capture multi-scale features of the signal. For example, the local energy signature of a 5mm pore and the global distribution of a 20mm crack can be simultaneously identified.

[0101] In a preferred embodiment of the present invention, the welding defect type is identified and the defect space coordinate set is marked based on the feature vector and the time-frequency domain joint threshold segmentation algorithm to form a defect coordinate set containing four-dimensional information. The welding defect types include porosity, lack of fusion and cracks, including:

[0102] From the eigenvector, key parameters related to the signal's time-frequency domain characteristics are separated, including time-domain characteristics and frequency-domain characteristics. Time-domain characteristics include peak amplitude, rise time, and duration; frequency-domain characteristics include center frequency, bandwidth, energy center frequency, and characteristic frequency components. Specifically, they include:

[0103] In the time domain waveform of the original ultrasonic echo signal, identify the maximum voltage amplitude of the defect echo, which reflects the strength of the defect reflection energy (for example, due to strong interface reflection, the peak amplitude of a pore is usually more than 30% higher than the base material background signal). Calculate the time required for the signal to rise from 10% to 90% of the peak amplitude to distinguish defect types (for example, due to surface irregularities, the rise time of a crack is shorter than that of a pore, reflecting the rapid reflection characteristics of stress waves). Record the time span from when the defect echo signal first exceeds the noise threshold to when it falls back below the threshold. Due to the large interface area, the duration of an unfused defect is usually 2-3 times longer than that of a pore.

[0104] Frequency domain feature extraction:

[0105] Center frequency: Perform Fourier transform on the signal and calculate the weighted average of the frequency-energy distribution (the frequency band with high energy contributes more to the center frequency). Due to the attenuation of high-frequency energy, the center frequency of the lack of fusion defect is often 5-10% lower than the base value of the base material.

[0106] The bandwidth is defined as the width of the frequency interval where the energy accounts for 90% of the total energy. Due to multi-scale interface reflection, the bandwidth of crack defects is 15-20% wider than that of pores.

[0107] The energy center frequency is determined by normalizing the product of the integral frequency and the corresponding energy to determine the core frequency point of the energy distribution (for example, the energy center of the pore is concentrated at 8-10MHz, while the center of gravity of the crack shifts to 5-7MHz due to the scattering effect).

[0108] Characteristic frequency components are used to identify energy anomalies in specific frequency bands (for example, a sudden drop in energy at 5 MHz may indicate the presence of an unfused interface). Sensitive frequency bands can be screened based on the welding process parameters in the design drawings (such as the base material acoustic impedance matching frequency).

[0109] The time domain features and frequency domain features of each detection signal are integrated into a matrix. Each row of the matrix corresponds to a signal feature vector at a detection location, including the time domain and frequency domain parameters of the signal. Each column of the matrix corresponds to a specific feature parameter, forming a multi-dimensional feature space, specifically including:

[0110] For each inspection location (uniquely identified by the X / Y / Z coordinates recorded by an encoder or total station), a row of data is created, containing all the time-domain and frequency-domain characteristic parameters of the signal at that location (such as peak amplitude, rise time, center frequency, etc., a total of 7-10 features). The matrix column names correspond to specific features (such as "time domain-peak amplitude" and "frequency domain-center frequency") to form a structured table. Each column of data is standardized (for example, normalized to the interval [0, 1]) to eliminate dimensional effects. The feature matrix is horizontally expanded into a two-dimensional table of "inspection location × feature parameter", with each cell storing the feature value of the corresponding location (for example, the peak amplitude of position (1500, 800, 200) is 85mV and the center frequency is 7.2MHz). Visualization tools (such as heat maps) are used to observe the distribution and clustering of different defect types in the feature space (for example, pores are concentrated in the "high peak-narrow frequency band" area, and cracks are distributed in the "fast rise time-low energy center of gravity" area) to assist in verifying the effectiveness of feature separation.

[0111] Based on historical inspection data and defect sample training, the joint judgment conditions of time domain features and frequency domain features are determined for three types of defects: pores, lack of fusion, and cracks. Specifically, the following are the conditions:

[0112] Collect at least 1,000 labeled samples (including signals of pores, lack of fusion, cracks, and no defects) and divide them into training and validation sets at an 8:2 ratio. Perform statistical analysis on the characteristic parameters of each defect type, and calculate the mean, standard deviation, and distribution range (e.g., the mean peak amplitude of pores is 90 mV, with a standard deviation of 10 mV; the mean duration of lack of fusion is 12 μs, with a standard deviation of 3 μs). Conditions are set:

[0113] For pores, set “peak amplitude > 80 mV, bandwidth < 3 MHz, and energy center frequency > 8 MHz” (high reflection in the time domain, high energy in the frequency domain with narrow bandwidth).

[0114] Unfused, the combination has "duration > 10μs, center frequency < 7MHz, and energy attenuation of characteristic frequency component (5MHz) > 20%" (long duration in the time domain, energy loss in the low frequency band in the frequency domain).

[0115] Crack,defined as “rising edge time < 0.8 μs and bandwidth > 5 MHz and periodic oscillation of the characteristic frequency component (6 MHz)” (rapid rise in the time domain and broadband scattering in the frequency domain).

[0116] The conditions are combined through logical "and" and "or" to avoid misjudgment of a single feature (for example, a high peak value alone may be caused by bubbles in the surface coupling agent, and a narrow frequency band must also be met to be judged as a pore).

[0117] According to the characteristic parameters of each detection signal and the joint judgment conditions, the characteristic vectors of each signal in the matrix are judged one by one to obtain the defect signal and the detection position coordinates corresponding to the defect signal, specifically including:

[0118] For each row in the feature matrix (i.e., the signal at each detection location), the time domain and frequency domain features are sequentially checked to see if they meet the joint criteria for the corresponding defect (e.g., first determine whether the condition for porosity is met, then determine whether the condition for lack of fusion is met, and finally determine whether the condition for cracks is met, ensuring mutual exclusivity). For example, a signal with a peak amplitude of 95mV (>80mV), a bandwidth of 2.5MHz (<3MHz), and an energy center frequency of 8.5MHz (>8MHz) meets the porosity condition and is marked as a porosity defect. The coordinates of the detection location are recorded (X=1505, Y=802, Z=198).

[0119] If the signal characteristics meet multiple defect conditions at the same time, it is prioritized as a high-risk type (such as cracks > lack of fusion > pores); if none of them meet the conditions, it is marked as "no defect" or "suspicious signal" (triggering manual review). Combined with the geometric constraints of the benchmark model (such as directly excluding the possibility of defects when the Z-axis coordinate exceeds the theoretical thickness range of the welding interface by ±50mm), it reduces misjudgments caused by boundary reflections.

[0120] According to the spatial coordinate mapping relationship of the defect-benchmark fusion model, the detection position coordinates corresponding to the defect signal are converted into three-dimensional coordinates in the defect-benchmark fusion model. The corresponding welding defect type label is added to each defect coordinate to form a defect coordinate set containing four-dimensional information. The defect-benchmark fusion model contains the theoretical geometric characteristics of the welding node and the measured defect data, that is, the defect coordinate set. The theoretical geometric characteristics include the theoretical axis of the cantilever beam and the theoretical position of the welding interface, specifically including:

[0121] According to the coordinate system of the defect-benchmark fusion model (with the geometric center of the embedded part as the origin, and the X / Y / Z axes corresponding to the cantilever beam length, balcony slab width, and weld interface thickness, respectively), the original coordinates recorded by the testing equipment (such as the probe center coordinates) are converted into three-dimensional coordinates in the benchmark model through translation and rotation (for example, the original coordinates (1505, 802, 198) are converted to the benchmark coordinates (5, 2, -2), in mm, with the origin as the reference point). 10% of the defect coordinates are randomly selected and compared with the theoretical coordinates of the benchmark model using total station measurements to ensure that the deviation is less than 0.5mm (for example, the lateral offset of a crack coordinate from the theoretical axis after conversion is 3mm, recorded as Y axis + 3mm).

[0122] Each defect coordinate is assigned a type label ("pore," "lack of fusion," "crack") and the inspection timestamp is recorded (accurate to the millisecond level, such as 2025-04-1410:05:30.123). Each defect entry contains "X / Y / Z coordinates + timestamp + defect type" (such as (1505, 802, 198, 20250414100530123, pore)) and is associated with theoretical features in the baseline model (such as whether it is within the embedded part outline and the distance from the theoretical axis). The defect coordinate set is stored as a structured file (such as CSV or JSON), containing all 4D information and additional attributes (such as the inspection equipment number, ambient temperature and humidity). It can be imported into BIM systems or 3D visualization software to generate a digital twin model of the welded joint with defect labels.

[0123] In the embodiment of the present invention, the time domain features (peak amplitude, rise time, duration) and frequency domain features (center frequency, bandwidth, energy center frequency) are separated to construct a multi-dimensional feature space containing the signal dynamic waveform and frequency energy distribution. For example:

[0124] Stomata: appear as isolated spikes (high peak value and short duration) in the time domain and as narrow-band high energy concentration in the frequency domain;

[0125] Unfused: Multiple reflections occur in the time domain (with increasing duration), and energy in the 5-8 MHz band in the frequency domain is significantly attenuated.

[0126] Crack: It appears as periodic oscillation in the time domain (with a steep rising edge), and the characteristic frequency components in the frequency domain are offset.

[0127] Compared to traditional single time-domain or frequency-domain analysis, joint features can capture the complex physical properties of defects, increasing the accuracy of identifying three types of defects by over 30%, and the recognition rate of microcracks below 2mm from 60% to 92%. Joint judgment conditions (such as "reflected energy mutation > 20dB and phase shift > 15°" for unfused defects) are trained using historical data to replace manual experience-based judgments, avoiding missed judgments due to subjective differences in operators (such as misjudging the parent material oxide film as unfused), reducing the misjudgment rate by 60%. The detection signal coordinates are converted into three-dimensional coordinates in the defect-reference fusion model (error ≤ 0.8mm), and geometric features such as the theoretical axis of the cantilever beam and the theoretical position of the welding interface are associated. For example:

[0128] The specific X / Y / Z coordinates of cracks (e.g., 1502, 805, 195) and their angle with the theoretical axis (45°) are annotated to visually visualize the spatial distribution of defects within weld joints (e.g., cracks extending along the stress direction). The 3D volume of pores (calculated by the density of adjacent voxels) and the interfacial separation area of unfused defects are quantified, providing direct input parameters for structural mechanics analysis. While traditional methods only provide 2D cross-sectional defect locations, this method provides a three-dimensional representation of the spatial morphology of defects, avoiding safety hazards caused by misidentification (e.g., missing microcracks in the core stress zone). Based on the theoretical axis of the benchmark model, the defect depth gradient (e.g., crack extension of 0.5mm per 10mm along the Z axis) and the lateral offset vector (e.g., pore center offset by 2.5mm from the axis) are automatically calculated, providing direct correlation to structural stress analysis. In one project, a 15% decrease in joint stiffness was confirmed by finite element simulation based on an excess stress wave attenuation coefficient (>10%), providing data support for timely reinforcement.

[0129] The defect coordinate set includes the detection timestamp (accurate to milliseconds), forming a four-dimensional data chain of "time-space-type-parameter", which supports the comparison of historical detection data (such as the expansion trend of defects at the same location over time). For example:

[0130] A dense porosity defect at a certain node showed a lateral shift of 5mm during inspection in 2023. This shift increased to 8mm during retesting in 2025, triggering an update to the structural life prediction model, adjusting the predicted remaining life from 22 years to 18 years. This enables dynamic monitoring of defect development. Defect coordinates and type tags are integrated into the Building Information Model (BIM), linking component design parameters (such as the theoretical outline of embedded parts), construction records (such as welding process parameters), and operation and maintenance data (such as load monitoring values), forming a comprehensive quality traceability system. Application of this system in a certain affordable housing project increased the efficiency of weld node defect rectification by 40% and reduced the cost of subsequent structural monitoring by 25%, achieving closed-loop management from inspection to maintenance.

[0131] A deep residual network automatically extracts 256-dimensional feature vectors, replacing the manual calculation of peak values, frequencies, and other parameters (which traditionally takes 5-10 minutes per signal). This reduces single-node detection time from 30 minutes to 5 minutes. Supporting 16-channel simultaneous acquisition, it is suitable for rapid inspection of prefabricated components in prefabricated buildings. Based on pre-set joint judgment criteria from training data (e.g., "high-frequency energy attenuation rate > 25% and duration > 5μs is considered a crack"), this system automatically classifies defect types, eliminating discrepancies in detection results caused by inconsistent threshold settings in traditional methods and improving the standardization of the inspection process by over 90%.

[0132] Through joint analysis in the time and frequency domains, bottom surface echo noise is suppressed (an 85% recognition rate is maintained when the signal-to-noise ratio is ≥15dB). For example, when testing a surface roughness of Ra≤6.3μm, the algorithm can effectively distinguish between echoes caused by insufficient welding current (local lack of fusion) and those caused by the base material oxide film (functional interface), ensuring the uniformity of the silicon-based coupling agent, thereby avoiding false defect misjudgments. For oblique-incidence scanning (30° / 60° shear wave) in complex areas of fillet welds, combined with the embedded part contour annotation of the benchmark model, false defect waves generated by geometric boundary reflections are automatically filtered out, reducing the missed detection rate from 12% to 3%.

[0133] In a preferred embodiment of the present invention, the dynamic offset parameters of the defect coordinate set relative to the theoretical axis of the cantilever beam are calculated in the defect-reference fusion model. The parameters include the defect depth gradient, the lateral offset vector, and the stress wave attenuation coefficient, including:

[0134] Extract the information of the theoretical axis of the cantilever beam from the defect-reference fusion model, including the direction, position and related geometric parameters of the axis; determine the defect coordinate set, each defect coordinate contains its position information in three-dimensional space, specifically including: retrieve the geometric center coordinates of the embedded part from the reference space grid model as the origin (such as (1500, 800, 200)), and define the theoretical axis of the cantilever beam as a straight line passing through the origin and along the X-axis (the set of all points that meet Y=800mm, Z=200mm); the positive direction of the X-axis points to the cantilever end of the cantilever beam (towards the balcony slab), and the negative direction points to the fixed end (the main structure connection side). The axis position is uniquely determined by the origin coordinates and the X-axis extension range (such as X-axis 1200-1800mm); the geometry The parameters include the spatial direction of the axis (linearly extending along the X-axis), the reference point (origin) and the relative position to the welding interface (the axis is located at the center of the theoretical thickness of the welding interface, Z = 200 mm); all coordinate points marked as "pores", "unfused" and "cracks" are extracted from the defect-reference fusion model. Each coordinate contains three-dimensional position information (X, Y, Z). For example, the coordinates of a crack are (1502, 805, 195), indicating that it is 1502 mm in the length direction of the cantilever beam, 5 mm to the right of the balcony slab width, and 5 mm below the welding interface thickness direction; abnormal coordinate points are eliminated (such as points where the Z axis exceeds the theoretical thickness range of the welding interface by ±50 mm, which are determined to be interference from external parent material) to ensure that only coordinate data within the effective defect area is retained.

[0135] According to the coordinate system and geometric features in the defect-reference fusion model, the depth direction is determined. The depth direction is along the direction perpendicular to the cantilever beam surface and pointing to the inside of the weld. Specifically, according to the definition of the reference model coordinate system, the cantilever beam surface (detection surface) is the XY plane (such as Z = 250mm is the balcony slab base material surface, Z = 150mm is the cantilever beam base material surface), the depth direction is perpendicular to the surface, that is, along the Z axis direction; the depth direction points to the inside of the weld interface, specifically from the surface (Z axis boundary) to the center axis (Z = 200mm ), for example: from the balcony slab surface (Z = 250mm) to the negative direction of the Z axis (cantilever beam parent material), or from the cantilever beam surface (Z = 150mm) to the positive direction of the Z axis (balcony slab parent material), and finally to the core area of the welding interface; verify the depth direction through the theoretical position of the welding interface in the benchmark model (such as the Z axis 150-250mm range) to ensure that the Z axis values of all defect coordinates are within this range, and the depth increase direction is consistent with the positive direction of the Z axis (such as the balcony slab parent material direction is positive and the cantilever beam parent material direction is negative).

[0136] The defect coordinates are grouped according to their positions along the cantilever beam length. For each defect in each group, the coordinate values in the depth direction are calculated, including:

[0137] The cantilever beam is divided into continuous intervals along the length direction (X-axis) at preset intervals (e.g., 5 mm). For example, the X-axis range of 1200-1205 mm is group 1, and 1205-1210 mm is group 2, until the X-axis range of all defect coordinates (e.g., 1200-1800 mm) is covered. For each defect coordinate, it is assigned to the corresponding group according to its X value. For example, X=1502 mm is assigned to the 1500-1505 mm group, and X=1506 mm is assigned to the 1505-1510 mm group. Overlapping boundaries between groups are allowed (e.g., including the left endpoint but not the right endpoint) to avoid repeated classification. For all defects in each group, extract their Z-axis coordinate values (depth position). If there are multiple defects in the same group (for example, the coordinates of three pores are Z = 190, 195, and 200 mm respectively), you can choose to calculate the average value (195 mm) or retain all values for subsequent gradient analysis (such as calculating the maximum and minimum depth changes). If there are no defects in the group (that is, no defects are detected within the length interval), skip the group or mark it as "no defect group" and do not participate in the subsequent gradient calculation.

[0138] For defects in adjacent groups, calculate the ratio of their coordinate difference in the depth direction to the distance difference in the cantilever beam length direction, that is, the defect depth gradient, which specifically includes:

[0139] Arrange all defective groups in order of the X axis (e.g., the X range of group n is 1500-1505 mm, and that of group n+1 is 1505-1510 mm), ensuring that adjacent groups are continuous and have no gaps in the length direction. Extract the representative X value of each group (e.g., the midpoint of the X axis within the group, 1502.5 mm for group n, and 1507.5 mm for group n+1) as the position identifier of the group in the length direction.

[0140] For each pair of adjacent groups, obtain their coordinate values in the depth direction (e.g., the average Z value of the nth group is 195 mm, and the average Z value of the n+1th group is 205 mm), and calculate the depth difference (205-195=10 mm) and the distance difference in the length direction (1507.5-1502.5=5 mm). The gradient value is the depth difference divided by the distance difference (10 mm / 5 mm=2 mm / mm), which means that for every 1 mm extension along the length direction of the cantilever beam, the defect depth increases by 2 mm in the positive direction of the Z axis (balcony slab base material), reflecting the extension trend of the defect in the thickness direction.

[0141] If the depth difference between adjacent groups is negative (e.g., the Z of the rear group is smaller than that of the front group), the gradient is negative, indicating that the defect depth extends in the negative direction of the Z axis (cantilever beam base material). It is necessary to combine the theoretical position of the welding interface to determine whether it points to the core stress area (e.g., near Z = 200 mm); if the distance between adjacent groups is too large (e.g., more than 10 mm) or the number of defects is too small (<2 in a single group), the gradient calculation of the group pair is ignored to avoid errors caused by data sparsity; if the standard deviation of the depth coordinates of a group is greater than 5 mm (indicating high dispersion of defect depth within the group), the median is used instead of the mean to reduce the impact of outliers.

[0142] In this embodiment, the rate of change of defect depth (Z-axis coordinate) is calculated by grouping along the cantilever beam's length (X-axis), visually demonstrating the defect's propagation trend through the thickness. For example, a crack within the X-axis range of 1500-1510 mm has a Z-axis coordinate increasing from 190 mm (near the cantilever beam's base material surface) to 210 mm (extending into the balcony slab's base material). The depth gradient is (210-190) / (1510-1500) = 2 mm / 1 mm, indicating that the crack is extending toward the core stress zone at a rate of 2 mm per millimeter of length, triggering a high-risk warning. (Traditional methods can only detect defects with a single depth and cannot identify propagation trends.) Porosity defects with a depth gradient close to zero (e.g., Z-axis coordinate fluctuations of less than 1 mm at the same X-axis position) are considered localized, isolated defects with low risk. Conversely, a depth gradient greater than 1 mm / 1 mm (e.g., a lack of fusion defect rapidly extending through the thickness) indicates a risk of large-scale separation at the weld interface. The depth gradient is directly related to the stress distribution through the structure's thickness. For example, the core load-bearing area of a cantilever beam is usually located in the middle of the thickness direction (near the theoretical axis of the Z axis). If the defect depth gradient points to this area (such as the gradient is positive, the Z axis coordinate increases toward the core area), its influence coefficient on the structural bending stiffness will automatically increase by 50%, avoiding the bearing capacity estimation deviation caused by misjudgment of the defect location.

[0143] Using the theoretical axis of the cantilever beam (a straight line passing through the origin and along the X-axis, with Y = origin Y and Z = origin Z) as a reference, the vertical distance (lateral offset) from the defect center to the axis and the offset direction (positive or negative value on the Y-axis) are calculated. For example, if the coordinates of the center of a lack of fusion defect are (1600, 810, 200) and the theoretical axis are (1600, 800, 200), the lateral offset vector is Y + 10mm, indicating that the defect is located 10mm to the right of the axis. Incorporating material mechanics theory, for every 5mm increase in offset, the bending stress concentration factor at that location increases by 15%, directly guiding the parameter input of the finite element simulation model. (Traditional methods rely on manual axis measurement, with errors greater than 2mm, leading to stress calculation errors exceeding 20%). During batch testing, if the lateral offset vectors of multiple defects at the same node are all greater than 5mm and deviate to the same side (e.g., the positive Y-axis), the system automatically identifies it as "weld eccentricity," triggering a construction process tracing (e.g., welder positioning deviation), mitigating structural hazards at the source. Risk levels are assigned based on the offset (e.g., green for <5mm, yellow for 5-10mm, and red for >10mm). In one actual project, the lateral offset vector of a crack in a balcony slab weld joint reached 12mm (exceeding the theoretical axis safety threshold of 10mm). Finite element analysis confirmed that the stress amplitude at this location was 30% higher than the design value, predicting fatigue failure risk three years in advance and avoiding the lag inherent in traditional empirical assessments.

[0144] By comparing the attenuation of stress wave amplitudes in defective and non-defective areas, the attenuation coefficient is calculated (e.g., attenuation coefficient = 1 − signal amplitude in the reference area × signal amplitude in the defective area), directly reflecting the impact of defects on the structural vibration transmission efficiency. For example, due to the significant scattering effect, dense porosity defects can have an attenuation coefficient of up to 40% (100 mV in the non-defective area and only 60 mV in the defective area). Finite element simulations verify that the stiffness of this node decreases by 15% compared to the design value, consistent with the results of traditional loading tests (traditional methods require offline testing and take 1-2 weeks, while this method performs real-time online calculations in less than 1 minute). Lack of fusion defects, due to interface separation, result in a sudden change in wave impedance, often with an attenuation coefficient greater than 30% and an accompanying phase shift (>10°). This effectively distinguishes signals due to poor welding process (true defects) from those due to uneven base material quality (non-defect signals), reducing the false positive rate by 70%. The attenuation coefficient serves as a core input parameter, driving the life assessment model to learn the correlation between defect parameters and structural failure time.

[0145] In one project, a joint had an attenuation coefficient of 25%, and the model predicted a remaining life of 22 years (compared to 30 years estimated by traditional accelerated testing). After five years of monitoring, the annual degradation rate of the joint stiffness matched the predicted curve, providing a reliable time window for operation and maintenance decisions (e.g., planning reinforcement plans 15 years in advance). The combination of depth gradient, lateral offset vector, and attenuation coefficient forms a three-dimensional assessment system encompassing "location-morphology-performance." For example: If a crack defect simultaneously meets the conditions of "depth gradient 1.5mm / mm (extending to the core area), lateral offset vector 8mm (stress concentration area), and attenuation coefficient 35% (significant decrease in stiffness)", the system will automatically determine it as "extremely high risk" and trigger the emergency reinforcement process as a priority; if a single pore defect meets the conditions of "depth gradient 0, lateral offset 3mm, and attenuation coefficient 10%", it will be determined as low risk and can be included in routine monitoring. This multi-parameter joint assessment avoids the one-sidedness of a single indicator and increases the accuracy of defect hazard judgment from 60% of traditional methods to 92%; the automatic calculation of dynamic parameters (without manual intervention) is deeply integrated with the geometric characteristics of the benchmark model (such as theoretical axis, welding interface thickness), and the single-node detection time is shortened from 30 minutes to 5 minutes. At the same time, a multi-dimensional report including spatial distribution, extension trend, and mechanical influence is output.

[0146] Dynamic parameters, as the core attributes of the defect coordinate set (four-dimensional information), are seamlessly integrated into the BIM system, forming a closed data loop with design drawings (theoretical coordinates of embedded parts), construction data (welding process parameters), and operation and maintenance monitoring (load history). For example, historical data on the lateral offset vector of a node shows an increase from 5mm to 12mm over three years. Combined with a simultaneous increase in the attenuation coefficient (from 15% to 30%), the system automatically generates a "defect expansion-stiffness degradation" correlation curve, providing real-time input for the structural health monitoring system and enabling digital management of the entire chain, from detection to prediction. Statistical analysis of dynamic parameters (e.g., the mean depth gradient of a batch of components is >0.5mm / mm) can trace welding process issues in the prefabrication process (such as welding gun angle deviation), optimizing the production process from the source and reducing subsequent operation and maintenance costs by over 25%.

[0147] In a preferred embodiment of the present invention, the process of determining the lateral offset vector is as follows:

[0148] Based on the coordinate system in the defect-reference fusion model and the geometric information of the cantilever beam, determine the transverse direction. The transverse direction is the direction perpendicular to the theoretical axis of the cantilever beam and within the plane where the cantilever beam is located. Specifically, it includes:

[0149] The reference origin of the theoretical axis of the cantilever beam is retrieved from the defect-reference fusion model (such as the geometric center coordinates of the embedded part, for example: (1500, 800, 200)). The axis extends along the X-axis (the length direction of the cantilever beam), and all points on the axis satisfy Y = 800 mm and Z = 200 mm (the origin's Y / Z coordinates). The plane where the cantilever beam is located is defined as the XY plane (parallel to the length and width directions of the cantilever beam, and perpendicular to the thickness direction Z axis). The transverse direction is the direction perpendicular to the theoretical axis (X axis) and located in the XY plane, that is, the Y axis direction in the coordinate system (the positive direction of the Y axis is the right side of the balcony slab, and the negative direction is the left side). The transverse offset only reflects the positional deviation of the defect in the width direction (Y axis) of the balcony slab and is not related to the thickness direction (Z axis) (the Z axis is used for depth gradient calculation). Ensure that the transverse direction is strictly confined to the XY plane to avoid confusion with the depth direction.

[0150] For each defect point in the defect coordinate set, project it onto the theoretical axis of the cantilever beam to obtain the corresponding projection point. Specifically, for any defect point coordinate (Xd, Yd, Zd), the coordinates of the point on the theoretical axis corresponding to the X value of Xd are (Xd, Y axis, Z axis), where the Y axis and Z axis take the fixed values of the theoretical axis in the reference model (for example, Y axis = 800mm, Z axis = 200mm, consistent with the origin). Replace the Y and Z coordinates of the defect point with the Y and Z axes of the theoretical axis, keeping the X coordinate unchanged, to obtain the coordinates of the projection point.

[0151] Example: The coordinates of the defect point are (1600, 810, 195), and the point projected onto the theoretical axis is (1600, 800, 200) (the Y axis is adjusted from 810mm to 800mm, and the Z axis is adjusted from 195mm to 200mm. However, according to the definition of the lateral direction, the Z axis adjustment is only used to locate the axis, and the lateral offset only focuses on the Y axis difference).

[0152] Calculate the vector between each defect point and its projection point, i.e., the lateral offset vector of the defect point relative to the theoretical axis of the cantilever beam. This offset includes the magnitude and direction of the offset. Specifically, based on the theoretical axis, the positive Y-axis direction is to the right of the balcony slab, and the negative Y-axis direction is to the left. If the defect point's Y coordinate Yd > the Y axis (e.g., 810mm > 800mm), the offset direction is the positive Y-axis direction (right); if Yd < the Y axis, the offset direction is the negative Y-axis direction (left). Calculate the absolute distance between the defect point and the projection point along the Y axis in millimeters (e.g., 10mm means the defect point is 10mm to the right of the axis).

[0153] The lateral offset vector of each defect point is expressed as (offset direction, offset distance), for example:

[0154] The offset vector of the defect point (1600, 810, 195) is (positive direction of the Y axis, 10 mm);

[0155] The offset vector of the defect point (1550, 790, 210) is (negative direction of the Y axis, 10 mm).

[0156] If the Z coordinate of the defect point exceeds the theoretical thickness range of the welding interface (such as Z < 150mm or Z > 250mm), it is judged as an invalid point and does not participate in the offset vector calculation (to avoid external interference of the base material); for the defect point located on the theoretical axis (Yd = Y axis), the offset vector is marked as (no offset, 0mm).

[0157] In an embodiment of the present invention, the defect point is projected vertically onto the cantilever beam's theoretical axis (a straight line passing through the origin and along the X-axis) as a reference, and the vector offset (such as the deviation in the Y-axis direction) is calculated with an accuracy of ±0.5mm. For example, after the coordinates of a crack defect (1600, 810, 200) are projected onto the axis, the lateral offset vector is Y-axis +10mm, indicating that the defect is located 10mm to the right of the axis. According to the principles of material mechanics, for every 5mm increase in offset, the bending stress concentration factor at that location increases by approximately 15%, directly providing accurate input parameters for finite element simulation (traditional manual measurement errors are greater than 2mm, resulting in stress calculation errors exceeding 20%).

[0158] During batch inspection, if the lateral offset vectors of multiple defects at the same node all deviate toward the positive Y-axis (e.g., the right side of a balcony slab), the system automatically identifies this as "weld eccentricity," indicating welding gun positioning deviation or fixture error, thus mitigating structural hazards at the source of construction. Combining the offset direction and magnitude, a heat map of the defect distribution is generated (e.g., red areas indicate high-risk areas with offsets greater than 10mm). During prefabricated component inspection, the lateral offset vectors were used to locate an unfused defect 12mm from the axis. Load testing confirmed that the stress amplitude at this location increased by 30% compared to the design value, providing an early warning of fatigue failure risk and avoiding the lag inherent in traditional empirical assessments.

[0159] The direction of the transverse offset vector (positive or negative value on the Y-axis) reflects the tendency of defects to shift across the width of the balcony slab. For example, if a large number of defects are concentrated in the positive Y-axis direction (on the right side of the balcony slab), this can be traced back to workpiece clamping misalignment during welding (e.g., the fixture tilted 1° to the right) or incorrect welding torch path planning (offset by 5mm from the theoretical trajectory). This can guide the construction unit to adjust process parameters to prevent recurrence of the same problem. If the transverse offset vectors of defects around cylindrical embedded parts are distributed radially (e.g., with the center of the embedded part as the origin and offset in all directions), this indicates insufficient anchor bar positioning accuracy during welding (e.g., anchor bar verticality deviation >2°), triggering the automated equipment calibration process in the prefabrication process. Defects located near the axis (offset <5mm) have little impact on structural load uniformity, while defects with offsets >10mm (especially those on the tension side of the cantilever beam) significantly alter the section moment of inertia, resulting in a reduction in load-bearing capacity. In one actual project, the transverse offset vectors identified three high-offset defects in critical load-bearing areas, leading to prioritized reinforcement and a 40% increase in joint rectification efficiency.

[0160] The lateral offset vector, as a core attribute of the defect's spatial coordinates, directly contributes to structural stress calculations. For example, when calculating the additional bending moment at a defect (M = F × d, where d is the offset distance and F is the design load), a 10mm offset increases the additional bending moment by 10% compared to the axis, significantly impacting the input parameters of the joint stiffness degradation model. Combined with the depth gradient (z-axis extension trend), a three-dimensional risk assessment system based on "plane offset + thickness extension" is formed. For example, if a crack exhibits both a Y-axis offset of +8mm and a Z-axis depth of +15mm (extending toward the core stress zone), the system automatically assigns twice the impact on the structural bending strength as a typical defect, avoiding the biased nature of single-dimensional assessments. Using a rigid reference (position accuracy of ±0.5mm) to the theoretical axis, false defect offsets caused by positioning errors of the inspection equipment (such as systematic deviations caused by tilted sensor installation) are filtered out. By comparing the distance between the defect coordinates and the axis, outliers exceeding the theoretical width of the weld interface (e.g., ±20mm of the balcony slab's design width) are automatically eliminated, reducing the false positive rate by 60%.

[0161] In a preferred embodiment of the present invention, the process of determining the stress wave attenuation coefficient is as follows:

[0162] Using the principles of ultrasonic testing and related physical models, the propagation process of stress waves in the cantilever beam and weld node is simulated, and the intensity of the stress wave is measured at each defect position in the defect coordinate set and the reference position on the cantilever beam. Specifically, in the defect-free area of the cantilever beam base material (such as the fixed end far away from the welding interface, the theoretical coordinates are verified as a non-welding area by the benchmark model, for example: X = 1000mm, Y = 800mm, Z = 200mm), ensure that the location is defect-free and the material is uniform, which serves as the benchmark reference point for stress wave propagation; extract all valid defect points from the defect coordinate set (excluding abnormal points outside the welding interface, such as pores, lack of fusion, and crack coordinates within the Z-axis range of 150-250mm), for example, the coordinates of a crack are (1502, 805, 195). Use the same high-frequency piezoelectric ceramic probe (center frequency 5-10MHz, arc-shaped wedge matching the curved surface) as used in the signal acquisition stage, attach it to the test surface at the reference position and defect position respectively, apply a uniform amount of silicone-based coupling agent (thickness 0.2-0.5mm), and ensure that the sound beam is incident vertically (record the angle parameters when it is obliquely incident).

[0163] Transmit ultrasonic pulses with the same parameters (e.g., transmitting voltage 150V, pulse width 0.1μs), receive stress wave echo signals, and focus on recording the peak amplitude of the defect echo (reflecting the stress wave intensity). Collect signals three times at each location and take the average value to reduce accidental errors (e.g., the amplitudes at the reference location are 100mV, 98mV, and 102mV three times, with an average of 100mV; the amplitudes at the defect location are 70mV, 72mV, and 68mV three times, with an average of 70mV). Signal calibration is synchronized with the environment:

[0164] Before acquisition, use the CSK-IA standard test block to calibrate the equipment gain to ensure that the bottom echo amplitude at the reference position is stable at 80% of the full screen; simultaneously record the detection environment parameters (temperature 25±2℃, humidity 40±5%RH). Because temperature changes affect the sound velocity, it is necessary to ensure that the detection environment of the reference and defect locations is consistent to avoid intensity measurement deviations caused by environmental differences.

[0165] The stress wave attenuation coefficient is calculated based on the change in stress wave intensity during propagation. This involves extracting the average signal amplitude at the reference location (denoted as Areference, e.g., 100mV) and the average signal amplitude at the defect location (denoted as Adefect, e.g., 70mV). Both are peak voltage values in the time domain waveform, reflecting the energy intensity of the stress wave when it propagates to that location. The difference in signal intensity between the defect location and the reference location is then compared. If Adefect is less than Areference, it indicates that the stress wave has attenuated due to the presence of the defect during propagation (e.g., cracks, pores, scattered or reflected energy). If the two are close (difference less than 5%), it is determined that there is no significant attenuation (possibly due to a small defect or noise signal).

[0166] According to the direction and degree of intensity change, the attenuation coefficient is qualitatively described as follows:

[0167] No attenuation: The difference between the amplitude of the defect position and the reference position is less than 5%, marked as "Level 0";

[0168] Mild attenuation: The amplitude decreases by 5%-20% (e.g., reference 100mV, defect 85mV), marked as "Level 1", which may correspond to small-sized pores;

[0169] Moderate attenuation: The amplitude decreases by 20%-35% (e.g. reference 100mV, defect 65mV), marked as "Level 2", usually indicating lack of fusion or medium crack;

[0170] Severe attenuation: Amplitude reduction > 35% (e.g. reference 100mV, defect below 60mV), marked as "Level 3", indicating large-area defects or through-hole cracks.

[0171] If the signal amplitude at the defect location is abnormally higher than that at the reference location (after eliminating equipment failure), it may be due to enhanced interface reflection (such as the difference in acoustic impedance between the base material and the weld metal). It is necessary to combine time-frequency domain characteristics (such as phase offset) to further determine whether it is a pseudo-defect. For multiple adjacent detection points of the same defect (such as three points in a 5mm×5mm grid), the average of the attenuation coefficients is taken as the final attenuation index of the defect to avoid single-point noise interference.

[0172] In the embodiment of the present invention, by comparing the stress wave intensity at the defect position with that at the reference position, the attenuation coefficient directly reflects the obstruction of the defect on the propagation of the sound wave:

[0173] Porosity / crack differentiation: dense pores have a high attenuation coefficient due to the scattering effect (e.g. > 30%), while cracks have an abnormal attenuation coefficient due to interface reflection (accompanied by phase shift), which can effectively distinguish the degree of damage of different defect types; the attenuation coefficient is positively correlated with the defect volume / area (e.g. 5mm 3 The attenuation of pores is 15%, and that of the 10mm unfused surface is 35%), providing a quantitative basis for the classification of defect hazard (the traditional method only qualitatively judges "the existence of defects").

[0174] Stress wave attenuation is directly related to material continuity, and the attenuation coefficient can represent the change in node stiffness in real time:

[0175] For every 10% increase in the attenuation coefficient, the corresponding node stiffness decreases by approximately 5% (verified by finite element simulation). For example, when the attenuation coefficient of a node exceeds the threshold (25%), the system automatically prompts that the stiffness is insufficient, replacing the time-consuming detection of traditional loading tests (from 2 weeks to 1 minute). Combined with historical attenuation data, it tracks the gradual impact of defect development on structural performance (for example, a 5% annual increase in the attenuation coefficient of a crack indicates intensified interface separation), providing real-time data support for operation and maintenance decisions.

[0176] The attenuation coefficient serves as a bridge between the physical properties of defects and the mechanical properties of structures, improving the accuracy of life prediction:

[0177] By learning from historical data on the correlation between attenuation coefficients and structural failure times (e.g., the lifespan of a node with attenuation greater than 30% is 25% shorter than a normal node), this method overcomes the limitations of traditional empirical formulas and maintains a prediction error within ±10%. Combined with parameters such as load and humidity, the attenuation coefficient can quantify the degradation rate of defects under complex working conditions (e.g., at 80% humidity, the attenuation coefficient growth rate increases by 20%), enabling personalized lifespan prediction. Direct calculation of the attenuation coefficient based on ultrasonic echo signals eliminates the need for additional sensors or offline testing, maintaining a single-node detection time of less than 5 minutes. This method is suitable for batch component inspection in prefabricated buildings. Signal calibration at a reference location (defect-free area) suppresses the influence of environmental noise (effective calculations are maintained even when the signal-to-noise ratio is ≥15dB), reducing the false positive rate by 40% compared to traditional methods.

[0178] In a preferred embodiment of the present invention, the dynamic offset parameters are input into a pre-trained welding quality assessment model to predict the structure life, including:

[0179] The historical dynamic offset parameters of cantilever beam-balcony slab welded joints in various prefabricated buildings were obtained. Specifically, the test data of cantilever beam-balcony slab welded joints was collected from more than 50 prefabricated building projects, including components that have been in service for 10-30 years (covering different regions, ambient humidity / temperature, and load conditions). Three types of dynamic parameters were extracted for each node:

[0180] Defect depth gradient: the rate of change of defect depth along the length of the cantilever beam (e.g., if the depth of a crack increases by 10 mm in the range 1500-1510 mm on the X-axis, the gradient is 1 mm / mm);

[0181] Horizontal offset vector: the plane distance and direction of the defect center from the theoretical axis (e.g. 12mm in the positive direction of the Y axis);

[0182] Stress wave attenuation coefficient: the difference in signal intensity between the defect location and the reference location (e.g., a 30% attenuation indicates significant stress wave energy loss).

[0183] Over 2,000 valid data points were collected, including parameter combinations for different defect types, such as pores, lack of fusion, and cracks. Outliers (e.g., obvious equipment errors with depth gradients greater than 5mm / mm) and duplicate test data (averaging multiple tests at the same node and location) were eliminated. Data was then categorized and stored by project, component type (e.g., precast concrete balcony slabs / steel cantilever beams), and defect type, creating a structured data set (Excel / CSV format).

[0184] Through accelerated fatigue testing, the actual service life data of the welded joints is obtained as labels. Specifically, the actual welded joints are replicated at a 1:1 ratio, using the same base material (Q345B steel) and welding process (such as gas shielded welding, current 200A, voltage 25V) to ensure consistency with historical data samples. Loading plan:

[0185] Fatigue load: simulates balcony slab live load (2.5kN / m²) + wind load combination, using sinusoidal loading (frequency 5Hz, stress ratio R=0.1), with a load amplitude of 70% of the design bearing capacity;

[0186] Environmental coupling: Some specimens were placed in a hot and humid chamber (temperature 40°C, humidity 90% RH) to simulate the corrosion environment in coastal areas, and some were placed in a room temperature dry environment as a control.

[0187] When the specimen has visible cracks (length > 5mm) or the stiffness drops by more than 20% (the deflection of the cantilever end is monitored by the displacement sensor), it is judged as failure and the cumulative number of loading cycles is recorded (converted to the actual service life, such as 10 6 Cycles ≈ 15 years). Three parallel tests were conducted on each specimen, and the average lifespan was used as the label (e.g., a cracked specimen with three lifespans of 22, 24, and 23 years would be labeled 23 years). For retired nodes in historical projects (e.g., components removed after 25 years of use), the remaining lifespan was inferred through on-site load tests to supplement the actual service life data (avoiding extrapolation errors based solely on accelerated testing).

[0188] The historical dynamic migration parameters are scaled to the interval [0, 1], the life data is logarithmically transformed, and the difference between the life data and the true life is measured using the mean square error. Specifically, the depth gradient (usually 0-5mm / mm), lateral migration vector (0-20mm), and attenuation coefficient (0-50%) are uniformly scaled to the interval [0, 1]:

[0189] Depth gradient = (actual value ÷ 5mm / mm) → e.g. 1mm / mm → 0.2, 5mm / mm → 1;

[0190] Lateral offset = (actual distance ÷ 20mm) → e.g. 12mm → 0.6, 20mm → 1 (negative direction is marked with a separate sign, e.g. -10mm on the Y axis → -0.5, normalized after absolute value processing);

[0191] Attenuation coefficient = (actual value ÷ 50%) → e.g. 30% → 0.6, 50% → 1.

[0192] The Min-Max normalization method was used to preserve relative differences between parameters (to avoid Z-score normalization from distorting the distribution of small samples). Lifespan labels (10-50 years) were logarithmically transformed (e.g., ln(20 years) = 3, ln(50 years) = 3.91) to transform the right-skewed distribution into a near-normal distribution, improving the neural network fit. The training and validation sets were divided into a ratio of 8:2 to ensure a balanced sample size across defect types and lifespan ranges (e.g., samples with a lifespan of <20 years and ≥20 years each accounted for 50%).

[0193] Train the fully connected neural network to obtain the trained fully connected neural network, specifically including: network architecture design:

[0194] Input layer: 3 neurons (corresponding to 3 dynamic parameters: depth gradient, lateral offset, and attenuation coefficient);

[0195] Hidden layers: 2 layers (64 neurons each), using ReLU activation function (to capture the nonlinear relationship between parameters), and adding Dropout layers (deactivation rate 0.2) between layers to prevent overfitting;

[0196] Output layer: 1 neuron (outputs the predicted lifespan after logarithmic transformation), using a linear activation function (adapted to regression tasks). Training process:

[0197] The optimizer, Adam (learning rate 0.001), dynamically adjusts the learning rate to balance convergence speed and accuracy;

[0198] Loss function: mean square error (MSE), which calculates the mean squared difference between the predicted lifespan and the true lifespan (e.g., predicted value ln(23) = 3.14, true value ln(23) = 3.14, MSE = 0);

[0199] Iteration strategy:

[0200] The batch size is 32, and the training runs for 100 epochs. The mean absolute error (MSE) is evaluated on the validation set every 5 epochs. Early stopping is triggered if the validation error does not decrease after 10 consecutive epochs to avoid overfitting (convergence is typically achieved after 60-80 epochs). After training, the mean absolute error (MSE) is calculated using a test set (200 data points independent of the training / validation sets) (e.g., the average deviation between predicted and true life expectancy is ≤ ±2 years). A scatter plot of the predicted and true values is then plotted (ideally, the data points are distributed along the y = x line).

[0201] The structure life is predicted based on the trained fully connected neural network and real-time dynamic offset parameters. Specifically, the dynamic parameters obtained by real-time detection (such as the current depth gradient of a node is 0.8mm / mm, the lateral offset is 8mm, and the attenuation coefficient is 25%) are scaled according to the normalization method of the training phase (0.8÷5=0.16, 8÷20=0.4, 25%÷50%=0.5) to form an input vector [0.16, 0.4, 0.5]. The pre-processed parameters are input into the trained neural network, and the output value is calculated through forward propagation (such as the logarithmic life of 3.25), which is restored to the actual life prediction value (e) through exponential transformation. 3.25 The prediction results are accompanied by a confidence score (for example, a variance of activation values in the neural network output layer of less than 0.05 is marked as "high confidence"). These predictions are integrated with the BIM system to mark the remaining lifespan of nodes in the 3D model (e.g., red warning for less than 15 years, yellow warning for 15-25 years, and green for more than 25 years). Maintenance recommendations are also automatically generated (e.g., if the lifespan is less than 10 years, reinforcement is recommended within three months).

[0202] In an embodiment of the present invention, the historical dynamic offset parameters of welding nodes in different projects (such as depth gradient, lateral offset vector, attenuation coefficient) are used in combination with the real life labels obtained by accelerated fatigue testing (such as the actual failure time of a node is 25 years) to construct a "defect parameter-life" mapping relationship, avoiding the rough estimation of traditional methods that rely on material manuals or empirical formulas (such as the error of traditional methods in estimating life often exceeds ±20%). The fully connected neural network automatically learns the complex nonlinear relationship between dynamic parameters and life (such as for every 10% increase in stress wave attenuation coefficient, the life is shortened by 15%; when the depth gradient is greater than 1mm / mm, the life impact weight increases by 30%). In an actual project, the model controlled the life prediction error of nodes containing dense pores within ±10%, which is twice as accurate as the traditional accelerated test extrapolation method (error ±30%).

[0203] The model integrates the defect's spatial location (lateral offset vector), extension trend (depth gradient), and mechanical properties (attenuation coefficient) to avoid the bias of a single parameter. For example, a defect with a lateral offset greater than 10mm alone might be classified as "high risk." However, combined with a depth gradient less than 0.5mm / mm (the defect has not extended into the core area) and an attenuation coefficient less than 20% (limited stiffness reduction), the model will adjust the lifespan prediction (from 20 years to 25 years), avoiding over-reinforcement. Trained with historical data under various environmental (temperature, humidity) and load conditions, the model can tailor its prediction logic to the specific node's service environment (for example, when humidity exceeds 80%, the weighting of the attenuation coefficient on lifespan is increased by 20%), addressing the traditional method's inability to account for environmental coupling effects. Single-node lifespan prediction takes less than 1 second (based on forward propagation of a trained neural network), enabling automated quality assessment of prefabricated components in prefabricated buildings (e.g., a production line can test 200 nodes per hour and simultaneously generate a lifespan report). This represents a 1,000-fold improvement in efficiency compared to traditional testing methods (which take 72 hours to test a single node).

[0204] like Figure 2 As shown, an embodiment of the present invention further provides an intelligent detection system for prefabricated building quality, comprising:

[0205] An acquisition module is used to collect ultrasonic echo signals from the welding nodes between the cantilever beam end and the balcony slab;

[0206] The fusion module is used to construct a reference spatial grid model based on the theoretical installation coordinates of the embedded parts, generate a three-dimensional defect distribution map of the welding interface through the time-of-flight inversion algorithm of the ultrasonic echo signal, and perform spatial registration with the reference grid model to obtain a defect-reference fusion model;

[0207] A processing module is used to perform feature fusion analysis on ultrasonic echo signals using a deep residual network based on the spatial coordinate mapping relationship of the defect-reference fusion model to obtain a feature vector corresponding to each signal. The module then identifies the welding defect type and labels the defect spatial coordinate set based on the feature vector and a joint threshold segmentation algorithm in the time and frequency domains to form a defect coordinate set containing four-dimensional information. The welding defect types include porosity, lack of fusion, and cracks.

[0208] A calculation module, configured to calculate dynamic offset parameters of a defect coordinate set relative to a theoretical axis of the cantilever beam in a defect-reference fusion model, the parameters including a defect depth gradient, a lateral offset vector, and a stress wave attenuation coefficient;

[0209] A prediction module is used to input dynamic offset parameters into a pre-trained weld quality assessment model to predict the structure life.

[0210] The above is a preferred embodiment of the present invention. It should be pointed out that for ordinary technicians in this technical field, several improvements and modifications can be made without departing from the principles of the present invention. These improvements and modifications should also be regarded as within the scope of protection of the present invention.

Claims

1. An intelligent detection method for prefabricated building quality, characterized in that: The method comprises: Collect ultrasonic echo signals from the welding nodes between the cantilever beam end and the balcony slab; A reference spatial grid model is constructed based on the theoretical installation coordinates of the embedded parts. A three-dimensional defect distribution map of the welding interface is generated using the time-of-flight inversion algorithm of the ultrasonic echo signal. This map is then spatially aligned with the reference grid model to obtain a defect-reference fusion model. Based on the spatial coordinate mapping relationship of the defect-reference fusion model, a deep residual network is used to perform feature fusion analysis on the ultrasonic echo signals to obtain the eigenvector corresponding to each signal. Based on the eigenvector and a joint threshold segmentation algorithm in the time and frequency domains, the welding defect types are identified and the defect spatial coordinate sets are marked to form a defect coordinate set containing four-dimensional information. The welding defect types include porosity, lack of fusion, and cracks. Calculating dynamic offset parameters of the defect coordinate set relative to the theoretical axis of the cantilever beam in the defect-reference fusion model, the parameters including defect depth gradient, lateral offset vector and stress wave attenuation coefficient; The dynamic offset parameters are input into a pre-trained weld quality assessment model to predict the structural life.

2. The intelligent detection method for prefabricated building quality according to claim 1, characterized in that: The reference space grid model is constructed based on the theoretical installation coordinates of the embedded parts, including: Extract the theoretical coordinates of embedded parts from the architectural design drawings, including the coordinates of each embedded part in three-dimensional space; determine the spatial range of the grid to be constructed based on the theoretical coordinates of the embedded parts; According to the resolution of ultrasonic testing and the spatial range of the grid, the size of the grid unit is set. Taking the theoretical coordinates of the embedded parts as a reference, a three-dimensional rectangular coordinate system is established in the detection area, where the origin is the geometric center of the embedded parts, the X-axis is along the length of the cantilever beam, the Y-axis is along the width of the balcony plate, and the Z-axis is along the thickness of the welding interface. Starting from the origin, the grid lines are evenly divided along the X-axis, Y-axis and Z-axis to form a regular cubic grid; The theoretical geometric features of the welded joints are embedded in the cubic mesh to form a reference space mesh model containing the coordinates of all mesh points and the theoretical geometric features, including the theoretical position of the weld interface, the theoretical axis of the cantilever beam, and the theoretical outline of the embedded parts.

3. The intelligent detection method for prefabricated building quality according to claim 2, characterized in that: The three-dimensional defect distribution map of the weld interface is generated by the time-of-flight inversion algorithm of the ultrasonic echo signal, and spatially aligned with the reference grid model to obtain a defect-reference fusion model, including: The probe is attached to the surface of the cantilever beam, emitting high-frequency ultrasonic pulses to the welding interface and receiving reflected echo signals. Each scan records the echo signal of a position, including signal amplitude and flight time. According to the propagation speed of ultrasonic waves in the material, the flight time of the echo signal is converted into the distance between the defect and the probe; The coordinates of the defect in three-dimensional space are calculated based on the actual position and beam angle during scanning. The defect coordinates obtained from each scan are used as a data point. After multiple scans, point cloud data containing the defect points is formed; Map the point cloud data to a three-dimensional space grid and divide the cube voxels according to the preset accuracy. The grayscale value of each voxel represents the defect probability of the corresponding area, and finally form a three-dimensional defect distribution map; Extract the four corner points of the embedded parts and the edge line of the welding interface from the reference space grid model and the three-dimensional defect distribution map; Match the coordinates of the embedded part corner points with the theoretical corner point coordinates in the reference grid, calculate the preliminary translation and rotation angle to align the overall positions of the two, calculate the distance from each measured defect point to the nearest theoretical point in the reference grid as the error, and gradually adjust the position and posture of the 3D defect distribution map until the overall error is less than the set threshold to form a registered defect map; The registered defect map is embedded into the reference grid model so that the coordinates of each defect point correspond to the spatial position of the reference grid, and finally a defect-reference fusion model containing theoretical geometric features and measured defect data is obtained.

4. The intelligent detection method for prefabricated building quality according to claim 3, characterized in that: Based on the spatial coordinate mapping relationship of the defect-reference fusion model, a deep residual network is used to perform feature fusion analysis on the ultrasonic echo signal to obtain the feature vector corresponding to each signal, including: Obtain the original ultrasonic echo signal, each signal corresponds to a detection position, and includes a curve showing the amplitude changing over time; Perform frequency domain analysis on ultrasonic echo signals to generate a time-frequency diagram and extract the energy distribution characteristics of the signal in different frequency bands. The horizontal axis of the time-frequency diagram is time and the vertical axis is frequency. The detection position coordinates corresponding to each signal are integrated into the input tensor, including the ultrasonic echo signal, time-frequency map and position coordinates; Based on the input tensor and the trained deep residual network, a comprehensive feature vector containing time-frequency domain signal features and spatial position information is generated.

5. The intelligent detection method for prefabricated building quality according to claim 4, characterized in that: Based on the feature vector and the time-frequency domain joint threshold segmentation algorithm, the welding defect type is identified and the defect space coordinate set is marked to form a defect coordinate set containing four-dimensional information. The welding defect types include pores, lack of fusion and cracks, including: From the feature vector, key parameters related to the time-frequency domain characteristics of the signal are separated, including time-domain characteristics and frequency-domain characteristics. Time-domain characteristics include peak amplitude, rise time, and duration; frequency-domain characteristics include center frequency, bandwidth, energy center frequency, and characteristic frequency components. The time domain features and frequency domain features of each detection signal are integrated into a matrix. Each row of the matrix corresponds to a signal feature vector at a detection location, including the time domain and frequency domain parameters of the signal. Each column of the matrix corresponds to a specific feature parameter, forming a multi-dimensional feature space. Based on historical inspection data and defect sample training, the joint judgment conditions of time domain features and frequency domain features are determined for three types of defects: pores, lack of fusion, and cracks. According to the characteristic parameters of each detection signal and the joint judgment conditions, the characteristic vectors of each signal in the matrix are judged one by one to obtain the defect signal and the detection position coordinates corresponding to the defect signal; According to the spatial coordinate mapping relationship of the defect-reference fusion model, the detection position coordinates corresponding to the defect signal are converted into three-dimensional coordinates in the defect-reference fusion model. The corresponding welding defect type label is added to each defect coordinate to form a defect coordinate set containing four-dimensional information.

6. The intelligent detection method for prefabricated building quality according to claim 5, characterized in that: In the defect-reference fusion model, the dynamic offset parameters of the defect coordinate set relative to the theoretical axis of the cantilever beam are calculated. The parameters include the defect depth gradient, the lateral offset vector, and the stress wave attenuation coefficient, including: Extract the theoretical axis information of the cantilever beam from the defect-reference fusion model, including the axis direction, position, and related geometric parameters; determine the defect coordinate set, each of which contains its position information in three-dimensional space; According to the coordinate system and geometric features in the defect-reference fusion model, the depth direction is determined, which is along the direction perpendicular to the cantilever beam surface and pointing to the inside of the weld; The defect coordinate set is grouped according to the position in the length direction of the cantilever beam, and for each defect in the group, its coordinate value in the depth direction is calculated; For defects in adjacent groups, the ratio of their coordinate difference in the depth direction to the distance difference in the cantilever beam length direction is calculated, that is, the defect depth gradient.

7. The intelligent detection method for prefabricated building quality according to claim 6, characterized in that: The process of determining the lateral offset vector is: According to the coordinate system in the defect-reference fusion model and the geometric information of the cantilever beam, the transverse direction is determined. The transverse direction is the direction perpendicular to the theoretical axis of the cantilever beam and in the plane where the cantilever beam is located. For each defect point in the defect coordinate set, project it onto the theoretical axis of the cantilever beam to obtain the corresponding projection point; The vector between each defect point and its projection point is calculated, that is, the lateral offset vector of the defect point relative to the theoretical axis of the cantilever beam, which includes the magnitude and direction information of the offset.

8. The intelligent detection method for prefabricated building quality according to claim 7, characterized in that: The process of determining the stress wave attenuation coefficient is: Using the principles of ultrasonic testing and related physical models, the propagation process of stress waves in the cantilever beam and weld node is simulated, and the intensity of the stress wave is measured at each defect position in the defect coordinate set and at the reference position on the cantilever beam. The stress wave attenuation coefficient is calculated based on the change in the intensity of the stress wave during its propagation.

9. The intelligent detection method for prefabricated building quality according to claim 8, characterized in that: Dynamic offset parameters are fed into a pre-trained weld quality assessment model to predict structural life, including: Obtain the historical dynamic offset parameters of cantilever beam-balcony slab welding nodes in different prefabricated buildings; Through accelerated fatigue testing, the actual service life data of welding nodes is obtained as labels; The historical dynamic offset parameters are scaled to the interval [0, 1], the life data are logarithmically transformed, and the mean square error is used to measure the difference between the life data and the true life; Training the fully connected neural network to obtain a trained fully connected neural network; The structural life is predicted based on the trained fully connected neural network and real-time dynamic offset parameters.

10. An intelligent detection system for prefabricated building quality, characterized in that: The system executes the method according to any one of claims 1 to 9, comprising: An acquisition module is used to collect ultrasonic echo signals from the welding nodes between the cantilever beam end and the balcony slab; The fusion module is used to construct a reference spatial grid model based on the theoretical installation coordinates of the embedded parts, generate a three-dimensional defect distribution map of the welding interface through the time-of-flight inversion algorithm of the ultrasonic echo signal, and perform spatial registration with the reference grid model to obtain a defect-reference fusion model; A processing module is used to perform feature fusion analysis on ultrasonic echo signals using a deep residual network based on the spatial coordinate mapping relationship of the defect-reference fusion model to obtain a feature vector corresponding to each signal. The module then identifies the welding defect type and labels the defect spatial coordinate set based on the feature vector and a joint threshold segmentation algorithm in the time and frequency domains to form a defect coordinate set containing four-dimensional information. The welding defect types include porosity, lack of fusion, and cracks. A calculation module, configured to calculate dynamic offset parameters of a defect coordinate set relative to a theoretical axis of the cantilever beam in a defect-reference fusion model, the parameters including a defect depth gradient, a lateral offset vector, and a stress wave attenuation coefficient; A prediction module is used to input dynamic offset parameters into a pre-trained weld quality assessment model to predict the structure life.

Images