Complex metal product internal defect analysis method based on ray detection

Abstract

The invention discloses a complex metal product internal defect analysis method and system based on radiographic inspection, a spiral track scanning path is generated through a workpiece three-dimensional model, a six-axis cooperative mechanical arm is utilized to drive a miniature X-ray source to move along the path, and + / -5 [mu] m positioning compensation is realized in combination with a piezoelectric ceramic driving mechanism. And the detector array module is matched to carry out multi-mode ray scanning and real-time data acquisition. Collected data are transmitted to a real-time data processing module through an LVDS differential interface, an FDK algorithm is executed by a JetsonTX2 embedded GPU to complete tomographic image reconstruction, a 3D-ResNet neural network model is loaded through a TensorRT inference engine to achieve intelligent defect recognition, and finally a detection report is generated and uploaded to an MES system. The system comprises a ray generation module, a multi-degree-of-freedom mechanical arm bearing module and the like, a high-voltage generator circuit comprises a ZVS resonance circuit, and a motion control module adopts an FPGA + ARM dual-core architecture.

Description

Technical Field

[0001] The present invention provides an analysis method, and in particular relates to an analysis method for internal defects of complex metal products based on radiographic detection. Background Art

[0002] In the manufacturing of complex metal products, accurate detection of internal defects is crucial for product quality control. Traditional X-ray inspection devices typically consist of a X-ray source, a detector, a mechanical transmission mechanism, and an image analysis system. The X-ray source transmits radiation through the workpiece, while the detector receives the attenuated radiation and converts it into an electrical signal. After image reconstruction, defect identification is performed manually or using simple algorithms. The mechanical transmission mechanism typically uses linear or simple curved motion to drive the X-ray source and detector to scan the workpiece.

[0003] However, existing devices have significant shortcomings. First, the single scanning path makes it difficult to fully inspect complex curved metal surfaces, resulting in missed inspections in some areas. Second, the fixed relative position of the X-ray source and detector prevents adjustment of the inspection geometry based on workpiece size and inspection requirements, impacting inspection accuracy and efficiency. Third, data processing capabilities are limited, and traditional image reconstruction algorithms are time-consuming and rely on manual extraction of defect features, failing to meet the real-time requirements of online inspection. Fourth, the system suffers from poor stability and lacks an effective environmental compensation mechanism. In complex industrial environments such as temperature fluctuations and electromagnetic interference, inspection results fluctuate significantly. Summary of the Invention

[0004] In order to solve the above problems, the present application provides a method for analyzing internal defects of complex metal products based on X-ray detection, which solves the difficult problems of poor adaptability of detection paths, unadjustable parameters, low processing efficiency and insufficient stability.

[0005] In order to solve the above technical problems, the present invention provides the following technical solution: a method for analyzing internal defects of complex metal products based on radiographic detection, comprising the following steps:

[0006] S1. Scan Path Planning: Generates a spiral scanning path based on the workpiece 3D model. The path point cloud data is transmitted to the high-precision motion control module via Gigabit Industrial Ethernet.

[0007] S2. Multimodal ray scanning:

[0008] The cone beam is emitted by the micro X-ray source of the ray generating module, and the high voltage generator circuit adopts a full-bridge inverter topology to maintain a stable output of 160kV / 2mA;

[0009] The six-axis collaborative robotic arm drives the radiation source along the planned path, and the three-dimensional fine-tuning platform at the end achieves ±5μm positioning compensation through a piezoelectric ceramic drive mechanism;

[0010] S3. Real-time data collection:

[0011] The rays that penetrate the workpiece are captured by the detector array module. The amorphous silicon flat panel detector outputs digital signals through the photoelectric conversion circuit. The AD9279 analog front-end chip samples the signals at a 12-bit resolution.

[0012] The collected data is transmitted to the real-time data processing module through the LVDS differential interface;

[0013] S4. Dynamic Motion Compensation: The absolute encoders of the robotic arm joints provide real-time feedback on position data. The motion controller, using the AD2S1210 resolver decoder chip, calculates position errors and drives the servo circuit for closed-loop correction.

[0014] S5. Tomographic image reconstruction: The image acquisition card receives data through the CameraLink interface, and the Jetson TX2 embedded GPU executes the FDK algorithm to reconstruct the 3D volume data.

[0015] S6. Intelligent Defect Identification: The TensorRT inference engine of the defect analysis unit loads a pre-trained 3D-ResNet neural network model to output porosity, slag inclusion depth, and crack direction parameters.

[0016] S7. Output of test results: Generate a report containing the three-dimensional coordinates of the defect and quantitative analysis, and upload it to the MES system through the main control communication module.

[0017] Preferably, the step S1 specifically includes:

[0018] Discretize the workpiece CAD model into a voxel grid and calculate the optimal relative position of the ray source and detector;

[0019] Dynamically adjust the scanning speed based on the adaptive step size algorithm:

[0020] For areas with a curvature radius greater than 50 mm, a uniform scanning speed of 15 mm / s is used;

[0021] In areas with a curvature radius of ≤50mm, the speed is reduced to 3mm / s and vibration compensation of the fine-tuning platform is activated.

[0022] Preferably, in step S3:

[0023] The photoelectric conversion circuit includes a time-integrated sampling unit that completes eight sampling accumulations within a 10ms exposure period.

[0024] A dual-threshold comparison circuit is used to eliminate dark current noise, and the comparator chip is ADCMP600.

[0025] Preferably, step S5 comprises:

[0026] Geometric distortion correction: Generate the system point spread function using the tungsten alloy calibration block of the online calibration module;

[0027] Scatter Correction: The scattered component is calculated by Monte Carlo algorithm and subtracted from the raw data.

[0028] Preferably, the training method of the neural network model in step S6 is:

[0029] Input layer: 128×128×128 volume data, 3D convolution kernel is used between layers;

[0030] Feature extraction: 4-level residual blocks connected to spatial pyramid pooling layer;

[0031] Output layer: Outputs the defect probability map through the Sigmoid activation function.

[0032] Preferably: based on the method of claims 1-5, the system includes a ray generating module, a multi-degree-of-freedom robotic arm carrying module, a high-precision motion control module, a detector array module, a real-time data processing module and a main control communication module;

[0033] in:

[0034] The ray generating module is rigidly connected to the end effector of the robotic arm, and the high voltage generator circuit includes a zero voltage switching ZVS resonant circuit;

[0035] The multi-axis motion controller of the motion control module adopts FPGA+ARM dual-core architecture and drives the robotic arm through the CAN bus;

[0036] The detector array module is installed at a 30° inclination angle to the radiation source, and the center distance can be adjusted within a range of 200-500mm.

[0037] The real-time data processing module includes a real-time image pre-processing circuit based on Xilinx Artix-7 FPGA.

[0038] Preferably: the high voltage generator circuit comprises:

[0039] PI closed-loop feedback circuit uses ADP197 power controller to achieve voltage fluctuation ≤ 0.5%;

[0040] High-frequency transformer, the secondary winding adopts a segmented shielding structure to reduce leakage inductance.

[0041] Preferably, an online temperature compensation module is also included:

[0042] The platinum resistance temperature sensor is mounted on the surface of the detector;

[0043] The PT100 signal conditioning circuit outputs a 4-20mA signal to the main control module, triggering the start and stop of the cooling system.

[0044] One or more technical solutions provided in the embodiments of this application have at least the following technical effects or advantages:

[0045] This device is an online detection system for internal defects of complex metal products based on the principle of cone-beam ray imaging. The core detection mechanism consists of a multi-degree-of-freedom robotic arm integrated with a micro X-ray source, a high-precision flat-panel detector and a real-time processing unit. Its working principle is as follows: the motion control module drives the robotic arm to drive the ray source to move along a pre-planned spiral trajectory, and the high-voltage generator circuit is synchronously activated to generate 160kV stable cone-beam rays to penetrate the workpiece. The attenuated rays are captured by the amorphous silicon detector array and converted into digital signals. The embedded GPU reconstructs the three-dimensional volume data in real time based on the FDK algorithm. At the same time, the pre-trained 3D convolutional neural network is used to intelligently identify and quantitatively analyze the defect characteristics. Combined with the piezoelectric ceramic dynamic compensation mechanism of the end fine-tuning platform and the online temperature calibration module, it finally realizes sub-millimeter-level online non-destructive detection of defects such as pores, cracks and slag inclusions inside complex curved metal parts.

[0046] Other advantages, objects and features of the present invention will be described in part in the following description and, in part, will be apparent to those skilled in the art based on an examination of the following or may be learned from the practice of the invention. BRIEF DESCRIPTION OF THE DRAWINGS

[0047] Figure 1 This is a timing diagram of the overall detection process of a method for analyzing internal defects of complex metal products based on radiographic detection according to the present invention;

[0048] Figure 2 This is a system hardware architecture module diagram of a method for analyzing internal defects of complex metal products based on radiographic detection according to the present invention;

[0049] Figure 3 This is a timing diagram of data processing and motion control of a method for analyzing internal defects of complex metal products based on radiographic detection according to the present invention;

[0050] Figure 4 This is a flowchart of the temperature compensation module of a method for analyzing internal defects of complex metal products based on radiographic inspection according to the present invention;

[0051] Figure 5 This is a timing diagram of the neural network model training and inference of a method for analyzing internal defects of complex metal products based on radiographic inspection according to the present invention. DETAILED DESCRIPTION

[0052] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0053] It should be noted that the terms “vertical”, “horizontal”, “up”, “down”, “left”, “right” and similar expressions used herein are for illustrative purposes only and do not represent the only implementation method.

[0054] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by those skilled in the art to which this invention pertains; the terms used herein in the specification of the present invention are for the purpose of describing specific embodiments only and are not intended to limit the present invention; the term "and / or" used herein includes any and all combinations of one or more of the associated listed items.

[0055] like Figure 1 and Figure 2 As shown in the figure, a method for analyzing internal defects in complex metal products based on X-ray inspection is implemented through seven steps: spiral trajectory scanning path planning, multimodal X-ray scanning, real-time data acquisition, dynamic motion compensation, tomographic image reconstruction, intelligent defect identification, and test result output. The spiral trajectory is generated using a 3D workpiece model, and the path point cloud data is transmitted via Gigabit Ethernet. A miniature X-ray source emits cone-beam radiation, and a full-bridge inverter topology maintains a stable output of 160kV / 2mA. A six-axis collaborative robotic arm drives the radiation source, and a piezoelectric ceramic drive mechanism achieves ±5μm positioning compensation. The detector array module captures the penetrating radiation, and the AD9279 chip samples the data at 12-bit resolution, with the data transmitted via an LVDS differential interface. Position error is calculated and closed-loop corrected using the AD2S1210 chip. The Jetson TX2 embedded GPU executes the FDK algorithm to reconstruct 3D volume data. The TensorRT inference engine loads the 3D-ResNet model to output defect parameters. Finally, a report is generated and uploaded to the MES system. During scan path planning, the scanning speed is dynamically adjusted based on the curvature radius. Areas with a curvature radius greater than 50mm are scanned at a constant speed of 15mm / s. Areas with a curvature radius ≤50mm are scanned at a reduced speed of 3mm / s, and vibration compensation for the fine-tuning platform is activated. The photoelectric conversion circuit for data acquisition completes eight sampling accumulations within a 10ms exposure period, and a dual-threshold comparator circuit is used to eliminate dark current noise. Image reconstruction includes geometric distortion correction and scattering correction. Neural network model training utilizes a specific input layer, feature extraction, and output layer structure.

[0056] In this embodiment, the end of the multi-degree-of-freedom robotic arm is rigidly connected to the three-dimensional fine-tuning platform through a flange. The three-dimensional fine-tuning platform includes an X / Y / Z piezoelectric ceramic drive mechanism with a travel range of ±100μm and a positioning accuracy of ±5μm; the miniature X-ray source is fixed to the upper surface of the three-dimensional fine-tuning platform through an insulating bracket, and the line connecting the center of its emission window and the center of the receiving surface of the detector array module forms a 30° detection inclination angle, and the angle α between the line and the normal of the workpiece detection area can be dynamically changed in the range of 0° to 45° through the posture adjustment of the robotic arm; the detector array module is fixed to the base of the robotic arm through a cantilever bracket, and the center distance between its receiving surface and the ray source emission window is continuously adjustable from 200 to 500mm through an electric slider.

[0057] The high-voltage generator circuit is integrated at the rear end of the ray source and includes a zero-voltage switching resonant circuit and a segmented shielded high-frequency transformer. The secondary winding of the transformer adopts a four-layer segmented structure, with inter-layer distributed capacitance ≤2pF and leakage inductance ≤10μH. The output end of the high-voltage generator circuit is connected to the ray source through a high-voltage cable. The high-voltage cable is sheathed with an electromagnetic shielding layer, and its shielding effectiveness is ≥60dB@10MHz.

[0058] The detector array module integrates an amorphous silicon flat-panel detector and a photoelectric conversion circuit. The photoelectric conversion circuit includes an 8-channel time-integrated sampling unit. Each sampling unit completes 8 sampling accumulations within a 10ms exposure period, effectively improving the signal-to-noise ratio. The output end of the photoelectric conversion circuit is connected to the AD9279 analog front-end chip via a differential signal line. The differential signal line adopts an impedance matching design with a characteristic impedance of 100Ω±10%.

[0059] The real-time data processing module is connected to the detector array module via a CameraLink interface, and includes a Xilinx Artix-7 FPGA and a DDR3 cache array. The FPGA internally implements data format conversion, bad pixel correction, and data compression algorithms, with a compression ratio of ≥4:1. The real-time data processing module is connected to the Jetson TX2 embedded GPU via a high-speed PCIe bus, with a transmission bandwidth of ≥8GB / s.

[0060] The multi-axis motion controller adopts an FPGA+ARM dual-core architecture and is connected to the robotic arm joint servo driver via a CANopen bus with a communication rate of 1Mbps. The robotic arm joint has an internal integrated absolute encoder with a resolution of ≥22 bits. The encoder feeds back posture data to the motion controller in real time via an SPI interface. The motion controller calculates position errors based on the AD2S1210 resolver decoder chip, and the generated compensation instructions drive the servo circuit through PWM signals to achieve closed-loop correction, with a position tracking error of ≤±0.05°.

[0061] The online temperature compensation module includes a PT100 temperature sensor mounted on the surface of the detector. The sensor output signal is converted into a 4-20mA current signal by a signal conditioning circuit and transmitted to the main control module via a shielded cable. The main control module dynamically adjusts the detector gain parameter according to the temperature change curve, and the temperature coefficient compensation accuracy is ≤±0.01% / ℃.

[0062] Based on the above implementation, the coordinated motion of the 3D fine-tuning platform and the robotic arm, combined with adaptive scanning speed control, enables the system to achieve a spatial resolution of 0.1mm when inspecting complex curved surfaces, with a defect detection rate of ≥99.5% @ 0.2mm equivalent defects. The ZVS resonant topology of the high-voltage circuit, combined with the cable shielding, ensures that the system's radiated interference complies with the GB / T17626.6-2017 standard, ensuring reliable detector signal acquisition. The heterogeneous computing architecture, combining FPGA pre-processing and GPU reconstruction, achieves single-frame image reconstruction time of ≤50ms, meeting the requirements of real-time inspection at 10fps. The temperature compensation module extends the detector's operating temperature range to -20°C to 60°C, with a temperature drift of ≤±0.05% / °C, significantly improving the system's stability in industrial applications. The variable tilt angle and adjustable center-to-center distance inspection layout enable the system to adapt to the inspection needs of workpieces ranging from φ50mm to φ500mm, expanding the device's application scenarios.

[0063] In solving the problem of internal defect detection of complex metal products, the robotic arm provides six-degree-of-freedom spatial positioning capability, achieving full coverage scanning of complex surfaces; the three-dimensional fine-tuning platform compensates for micro-positioning errors through piezoelectric ceramic drive to ensure the relative position accuracy of the radiation source and the detector; the adaptive scanning speed control dynamically adjusts the acquisition parameters according to the curvature of the workpiece to balance detection efficiency and accuracy; the ZVS topology and segmented transformer of the high-voltage circuit improve energy conversion efficiency and reduce electromagnetic interference; the time integration sampling unit and the dual-threshold comparison circuit work together to effectively suppress noise and improve weak signal detection capabilities; the multimodal data processing architecture realizes rapid conversion from raw data to defect parameters; the temperature compensation module eliminates the impact of ambient temperature changes on the detection results, ensuring the long-term stability of the system.

[0064] like Figure 3 and Figure 4As shown in the figure, the complex metal product internal defect analysis system based on the above-mentioned method consists of a radiation generation module, a multi-degree-of-freedom robotic arm support module, a high-precision motion control module, a detector array module, a real-time data processing module, and a main control and communication module. The radiation generation module is rigidly connected to the end of the robotic arm. The high-voltage generator circuit includes a zero-voltage switching (ZVS) resonant circuit and a PI closed-loop feedback circuit. An ADP197 power controller is used to achieve a voltage fluctuation of ≤0.5%. The secondary winding of the high-frequency transformer adopts a segmented shielding structure. The multi-axis motion controller of the motion control module uses an FPGA + ARM dual-core architecture and drives the robotic arm via the CAN bus. The detector array module is mounted at a 30° angle to the radiation source, with an adjustable center distance between 200 and 500 mm. The real-time data processing module includes real-time image pre-processing circuitry based on a Xilinx Artix-7 FPGA. The system also includes an online temperature compensation module. A platinum resistance temperature sensor is mounted on the detector surface. A PT100 signal conditioning circuit outputs a 4-20 mA signal to the main control module to trigger the cooling system start and stop.

[0065] In this embodiment, the complete implementation of this device also relies on the following existing technologies and structures: the ray source window uses a 500μm thick beryllium metal film as the ray exit window, and its transmittance to 100keV X-rays is ≥95%; the receiving surface of the detector array module is covered with a 0.2mm thick cesium iodide scintillator, which is coupled to the amorphous silicon photodiode through an optical fiber array, and the X-ray-visible light conversion efficiency is ≥55%; the transmission system of each joint of the robotic arm uses a harmonic reducer with a reduction ratio of 1:100 and a backlash of ≤30 arc seconds; the power MOSFET of the ZVS resonant circuit in the high-voltage generator circuit uses IXFH23N60P2, with a switching frequency of 100kHz and a conversion efficiency of ≥92%; the bad pixel correction algorithm implemented in the FPGA uses a median filter combined with a linear interpolation method , the correction accuracy of fixed bad points is ≥99.8%; the cooling system of the embedded GPU adopts a liquid cooling circulation device, which transfers heat to the external radiator through ethylene glycol aqueous solution, and can maintain the GPU core temperature ≤85°C at an ambient temperature of 65°C; the MES system interface module supports the OPCUA protocol, and realizes two-way data interaction with the factory information system through industrial Ethernet, and the data transmission delay is ≤100ms; the device casing is made of aviation aluminum alloy 6061-T6, the surface is hard anodized, the film thickness is ≥25μm, and the protection level reaches IP65; the cantilever bracket of the detector array module has an integrated stress sensor, which triggers a safety alarm when the load exceeds 110% of the rated value; the motion control software adopts a ROS-based architecture, through MoveIt! The motion planning library generates optimal scanning trajectories with a trajectory planning error of ≤±0.1mm. The 3D-ResNet neural network model is trained using MixUp data augmentation technology, linearly interpolating volume data of different defect types with a parameter of α = 0.2, effectively improving the recognition accuracy of small sample defects. The high-voltage cable utilizes a three-layer insulation structure: an inner layer of polytetrafluoroethylene, a middle layer of silicone rubber, and an outer layer of chloroprene rubber. It has a withstand voltage of ≥200kV and a minimum bend radius of 80mm. These complementary technologies and material selections further enhance the device's engineering implementation, ensuring the system's reliability and detection accuracy in industrial environments.

[0066] Specifically, during the implementation of this solution, the installation of the ray generation module and the six-axis collaborative robotic arm adopts a modular quick-release interface design, which achieves rapid and precise assembly through flanges and positioning pins, ensuring that the installation error of the ray source is within ±0.02mm. The FPGA+ARM dual-core architecture multi-axis motion controller in the high-precision motion control module is connected to the robotic arm joint servo driver via a dedicated terminal block via the CAN bus. Before connection, a network analyzer is required to perform impedance matching debugging on the CAN bus to ensure the integrity of the communication signal. Before the system is started, a tungsten alloy calibration block is required to perform geometric calibration of the ray detection system. The calibration block is placed in the center of the detection area, the ray is emitted through the ray generation module, the detector array module collects projection data, and the calibration software is used to calculate the system point spread function to complete the calibration of the geometric distortion correction parameters. The Xilinx Artix-7 FPGA in the real-time data processing module needs to be pre-burned with a customized configuration file through the JTAG interface to implement the data preprocessing algorithm function. It is physically connected to the Jetson TX2 embedded GPU through the PCIe slot, and the driver installation and data transmission protocol configuration are completed under the Linux operating system. When inspecting complex metal products, the three-dimensional model of the workpiece is first imported into the scanning path planning software, and the spiral trajectory scanning path is generated using an adaptive step-size algorithm. The path point cloud data is then transmitted to the high-precision motion control module via Gigabit Industrial Ethernet. The motion control module uses the AD2S1210 resolver decoding chip to calculate the position error in real time based on the path data and the posture data fed back by the absolute encoder of the robotic arm joint, and drives the servo circuit to perform closed-loop correction on the robotic arm movement. During the inspection process, the high-voltage generator circuit adopts a full-bridge inverter topology, maintaining a stable output of 160kV / 2mA through a PI closed-loop feedback circuit and an ADP197 power controller. The cone-beam radiation emitted by the micro X-ray source penetrates the workpiece and is captured by a detector array module mounted at a 30° angle to the source. The amorphous silicon flat-panel detector completes eight sampling accumulations within a 10ms exposure period through an internal photoelectric conversion circuit. A dual-threshold comparator circuit eliminates dark current noise. The converted digital signal is sampled at 12-bit resolution by the AD9279 analog front-end chip and transmitted to the real-time data processing module via an LVDS differential interface for preprocessing. The preprocessed data is then transmitted to the Jetson TX2 embedded GPU via the CameraLink interface, where the FDK algorithm is executed for tomographic image reconstruction. The reconstructed 3D volume data is input into the TensorRT inference engine of the defect analysis unit, which loads a pretrained 3D-ResNet neural network model for intelligent defect identification. A report containing the 3D coordinates and quantitative analysis of the defects is generated and uploaded to the MES system via the main control communication module.When the online temperature compensation module is working, the platinum resistance temperature sensor monitors the detector surface temperature in real time. Its output signal is converted into a 4-20mA standard signal by the PT100 signal conditioning circuit and transmitted to the main control module. The main control module controls the start and stop of the cooling system according to the preset temperature threshold and dynamically adjusts the detector gain parameters.

[0067] Although the present invention has been disclosed above in terms of preferred embodiments, it is not intended to limit the present invention. Anyone familiar with this technology can make various changes and modifications without departing from the spirit and scope of the present invention. Therefore, the scope of protection of the present invention should be based on the definition of the claims.

Claims

1. A method for analyzing internal defects of complex metal products based on radiographic detection, characterized in that: The following steps are involved: S1. Scan Path Planning: Generates a spiral scanning path based on the workpiece 3D model. The path point cloud data is transmitted to the high-precision motion control module via Gigabit Industrial Ethernet. S2. Multimodal ray scanning: The cone beam is emitted by the micro X-ray source of the ray generating module, and the high voltage generator circuit adopts a full-bridge inverter topology to maintain a stable output of 160kV / 2mA; The six-axis collaborative robotic arm drives the radiation source along the planned path, and the three-dimensional fine-tuning platform at the end achieves ±5μm positioning compensation through a piezoelectric ceramic drive mechanism; S3. Real-time data collection: The rays that penetrate the workpiece are captured by the detector array module. The amorphous silicon flat panel detector outputs digital signals through the photoelectric conversion circuit. The AD9279 analog front-end chip samples the signals at a 12-bit resolution. The collected data is transmitted to the real-time data processing module through the LVDS differential interface; S4. Dynamic Motion Compensation: The absolute encoders of the robotic arm joints provide real-time feedback on position data. The motion controller, using the AD2S1210 resolver decoder chip, calculates position errors and drives the servo circuit for closed-loop correction. S5. Tomographic image reconstruction: The image acquisition card receives data through the CameraLink interface, and the Jetson TX2 embedded GPU executes the FDK algorithm to reconstruct the 3D volume data. S6. Intelligent Defect Identification: The TensorRT inference engine of the defect analysis unit loads a pre-trained 3D-ResNet neural network model to output porosity, slag inclusion depth, and crack direction parameters. S7. Output of test results: Generate a report containing the three-dimensional coordinates of the defect and quantitative analysis, and upload it to the MES system through the main control communication module.

2. The method for analyzing internal defects of complex metal products based on radiographic inspection according to claim 1, characterized in that: The step S1 specifically includes: Discretize the workpiece CAD model into a voxel grid and calculate the optimal relative position of the ray source and detector; Dynamically adjust the scanning speed based on the adaptive step size algorithm: For areas with a curvature radius greater than 50 mm, a uniform scanning speed of 15 mm / s is used; In areas with a curvature radius of ≤50mm, the speed is reduced to 3mm / s and vibration compensation of the fine-tuning platform is activated.

3. The method for analyzing internal defects of complex metal products based on radiographic inspection according to claim 1, characterized in that: In the step S3: The photoelectric conversion circuit includes a time-integrated sampling unit that completes eight sampling accumulations within a 10ms exposure period. A dual-threshold comparison circuit is used to eliminate dark current noise, and the comparator chip is ADCMP600.

4. The method for analyzing internal defects of complex metal products based on radiographic inspection according to claim 1, characterized in that: The step S5 comprises: Geometric distortion correction: Generate the system point spread function using the tungsten alloy calibration block of the online calibration module; Scatter Correction: The scattered component is calculated by Monte Carlo algorithm and subtracted from the raw data.

5. The method for analyzing internal defects of complex metal products based on radiographic inspection according to claim 1, characterized in that: The training method of the neural network model in step S6 is: Input layer: 128×128×128 volume data, 3D convolution kernel is used between layers; Feature extraction: 4-level residual blocks connected to spatial pyramid pooling layer; Output layer: Outputs the defect probability map through the Sigmoid activation function.

6. A system for analyzing internal defects of complex metal products based on radiographic inspection, characterized by: Based on the method of claims 1-5, the system includes a ray generating module, a multi-degree-of-freedom robotic arm carrying module, a high-precision motion control module, a detector array module, a real-time data processing module and a main control communication module; in: The ray generating module is rigidly connected to the end effector of the robotic arm, and the high voltage generator circuit includes a zero voltage switching ZVS resonant circuit; The multi-axis motion controller of the motion control module adopts FPGA+ARM dual-core architecture and drives the robotic arm through the CAN bus; The detector array module is installed at a 30° inclination angle to the radiation source, and the center distance can be adjusted within a range of 200-500mm. The real-time data processing module includes a real-time image pre-processing circuit based on Xilinx Artix-7 FPGA.

7. The complex metal product internal defect analysis system based on radiographic inspection according to claim 6, characterized in that: The high voltage generator circuit comprises: PI closed-loop feedback circuit uses ADP197 power controller to achieve voltage fluctuation ≤ 0.5%; High-frequency transformer, the secondary winding adopts a segmented shielding structure to reduce leakage inductance.

8. The complex metal product internal defect analysis system based on radiographic inspection according to claim 6, characterized in that: Also includes online temperature compensation module: The platinum resistance temperature sensor is mounted on the surface of the detector; The PT100 signal conditioning circuit outputs a 4-20mA signal to the main control module, triggering the start and stop of the cooling system.

Images