Pipeline defect detection method based on fusion of magnetic eddy current and machine vision

Abstract

The application relates to the field of nondestructive testing, and discloses a pipeline defect detection method based on fusion of magnetic eddy current and machine vision, which comprises the following steps: acquiring an image through a depth camera to extract defect features and solve target three-dimensional coordinates, planning a third-order polynomial trajectory according to the three-dimensional coordinates and motor deceleration limit, controlling a robot chassis to smoothly decelerate to a parking point, controlling a mechanical arm to drive an eddy current probe to approach a defect, combining an end compliant mechanism and a displacement disturbance constraint formula to generate a static contact pre-pressure, making the probe adhere to a pipe wall to physically lock a lift-off value, adaptively configuring double-frequency electromagnetic excitation parameters according to visual defect features, collecting a composite signal and inputting the one-dimensional convolutional neural network, and outputting a quantitative evaluation result of the defect physical depth. The application solves the problem of signal lift-off value fluctuation caused by chassis vibration in mobile robot detection, corrects the nonlinear error under single-frequency excitation, and improves the defect quantitative detection precision in a complex engineering environment.

Description

Technical Field

[0001] This invention relates to the field of nondestructive testing technology, specifically to a pipeline defect detection method based on the fusion of magnetic eddy current and machine vision. Background Technology

[0002] This invention relates to the fields of non-destructive testing and robot control automation, specifically to a pipeline defect detection method based on the fusion of magnetic eddy current and machine vision. In petrochemical and oil and gas storage and transportation scenarios, outdoor inspections of metal pipelines typically rely on mobile robots equipped with non-destructive testing instruments. After the robot chassis moves to the suspected defect area, the robotic arm uses its probe to attach to the pipeline surface, acquiring physical state data of the pipeline metal to assess its structural integrity.

[0003] In real-world outdoor service environments, mobile robot chassis are subject to ground or wind loads, resulting in chassis vibration and normal displacement disturbances. Existing automated inspection systems often use robotic arms to rigidly press probes against the pipe surface, lacking buffering and compensation mechanisms for these disturbances. When the chassis vibrates, this rigid structure can cause high-frequency detachment between the probe contact surface and the pipe wall, leading to fluctuations in the equivalent lift-off value of electromagnetic detection. This physical contact instability results in a decrease in the signal-to-noise ratio of the acquired composite signal and baseline drift, making it impossible to extract effective defect features.

[0004] Furthermore, existing mobile eddy current testing equipment mostly employs a fixed-frequency single-frequency electromagnetic excitation mode. Limited by the skin effect of the electromagnetic field, a single excitation frequency cannot simultaneously address both surface defect resolution and deep defect penetration. When faced with complex surface corrosion pits of varying depths and equivalent apertures, traditional equipment relies solely on the phase angle or amplitude changes of a single-frequency signal for evaluation. This makes it difficult to establish a mapping relationship between the electromagnetic response and the actual three-dimensional geometric dimensions, resulting in nonlinear measurement biases in the final quantitative output, which fails to meet the demands for accurate defect size evaluation in engineering settings. Summary of the Invention

[0005] The technical problem solved by this invention is that when existing mobile robots perform non-destructive testing of pipeline surfaces, the vibration generated by the chassis movement and the rigid contact of the robotic arm can easily cause fluctuations in the probe lift-off value, which in turn leads to a decrease in the signal-to-noise ratio of the detection signal and a large error in the quantitative evaluation of defects.

[0006] To address the above problems, the present invention provides the following technical solution: The first aspect of this invention provides a pipeline defect detection method based on the fusion of magnetic eddy current and machine vision, comprising the following steps: Visual forward perception and target space calculation steps: acquire images of the pipe surface using a depth camera, extract defect features, and calculate the global three-dimensional coordinates of the target defect; Dynamic gait interception and smooth approximation control steps: The braking distance between the target defect and the robot chassis is calculated based on the global three-dimensional coordinates. A smooth deceleration trajectory is planned, and the robot chassis is controlled to travel along the smooth deceleration trajectory to the target parking point. Physical locking steps for robotic arm attitude fixation and lift-off values: The robotic arm is controlled to drive the eddy current probe loaded at the end to approach the target defect. Static contact pre-pressure is generated through the end-effector conforming mechanism to control the eddy current probe to adhere to the pipe surface. Adaptive excitation and quantitative defect evaluation steps: Configure the electromagnetic excitation parameters of the eddy current probe according to the defect characteristics, collect the composite magnetic eddy current signal, and generate a quantitative evaluation result for the target defect based on the composite magnetic eddy current signal.

[0007] Preferably, the specific process of the visual forward perception and target space solution steps includes: A lightweight convolutional network is used to extract defect features from images and output the defect classification confidence score. Based on the defect characteristics, a multi-level coordinate system mapping is performed to calculate the global three-dimensional coordinates of the target defect in the global coordinate system.

[0008] Preferably, in the dynamic gait interception and smooth approximation control steps, the process of planning a smooth deceleration trajectory includes: The straight-line distance from the center of the robot chassis to the target parking point is calculated and used as the braking distance. Trigger a system state machine interrupt to obtain the robot chassis’s current initial running speed and the maximum allowable deceleration of the underlying drive motors; The minimum threshold for the desired braking time is calculated based on the initial operating speed and maximum deceleration, and the braking time is configured to be a time value greater than or equal to the minimum threshold. Based on the braking distance, initial operating speed, and braking time, a reference displacement trajectory equation containing third-order polynomial coefficients is generated as a smooth deceleration trajectory.

[0009] Preferably, the expression for the reference displacement trajectory equation is: ,in, For time The changing braking displacement, , , Let be the coefficients of a third-order polynomial, and satisfy: , , ,in, The initial velocity, Braking distance, This refers to the braking time.

[0010] Preferably, the process of the controller robot chassis traveling along a smooth deceleration trajectory to the target parking point includes: the robot chassis downgrading the dynamic cruising gait to a low-speed crawling gait according to the smooth deceleration trajectory, and switching to a stationary standing gait when it reaches the target parking point; While standing at the station, perform secondary precise positioning of the target defect.

[0011] Preferably, the end-effector conforming mechanism includes an elastic compression spring. In the physical locking steps of the robotic arm's attitude fixation and lift-off values, the process of generating static contact pre-pressure through the end-effector conforming mechanism includes: Obtain the maximum expected displacement disturbance amplitude of the robot chassis in the normal direction; Control the downward stroke of the robotic arm and generate a formula that satisfies the constraints. Static contact preload; in, For static contact preload, Let be the stiffness coefficient of the elastic compression spring. This represents the maximum displacement disturbance amplitude.

[0012] Preferably, the end-compliance mechanism further includes a viscous damper; The damping coefficient of the viscous damper satisfies: ; in, The damping coefficient of the viscous damper. Let be the stiffness coefficient of the elastic compression spring. The equivalent physical mass of the end-effector conforming mechanism.

[0013] Preferably, the specific process of the adaptive stimulus and quantitative defect evaluation step includes: Defect features are used as visual semantic labels. Based on the skin effect equation, a preset frequency mapping table is retrieved, and dual-frequency electromagnetic excitation parameters are sent to the eddy current probe as electromagnetic excitation parameters. The collected composite magnetic leakage signal is input into a one-dimensional convolutional neural network for network regression, and the predicted physical depth and lateral equivalent pore size of the corrosion pit are output as quantitative evaluation results. After completing the quantitative evaluation of defects, the system status is reset.

[0014] A second aspect of the present invention provides an electronic device, comprising: a processor and a memory; the memory being used to store a computer program; The processor is used to execute computer programs to implement the aforementioned pipeline defect detection method based on the fusion of magnetic eddy current and machine vision.

[0015] A third aspect of the present invention provides a computer-readable storage medium storing a computer program, which, when executed by a processor, implements the aforementioned pipeline defect detection method based on the fusion of magnetic eddy current and machine vision.

[0016] This invention provides a pipeline defect detection method based on the fusion of magnetic eddy current and machine vision. It has the following beneficial effects: 1. This invention uses visual forward-looking perception to calculate the three-dimensional coordinates of the target, combines this with the maximum deceleration of the chassis drive motor to calculate the braking time, and then constructs a third-order polynomial reference displacement trajectory equation to control the chassis gait degradation. This control method enables the robot to achieve smooth deceleration and parking when approaching the defective target, reducing the mechanical impact inertia generated during sudden stops, and laying a stable platform foundation for the subsequent high-precision posture determination of the robotic arm.

[0017] 2. This invention introduces a compliant mechanism comprising an elastic element and a viscous damper at the end of the robotic arm, and sets physical constraints on the contact preload and damping coefficient based on the maximum displacement disturbance amplitude in the chassis normal direction. This structural feature transforms traditional rigid detection into controlled flexible contact, ensuring that the probe contact surface remains in close contact with the pipe wall even when the chassis vibrates. This physically suppresses lift-off fluctuations caused by high-frequency probe jitter, thus guaranteeing the signal-to-noise ratio of the magnetic eddy current signal.

[0018] 3. This invention utilizes defect semantic features extracted by a visual network to drive the eddy current probe to adaptively configure dual-frequency electromagnetic excitation parameters, and inputs the acquired composite magnetic eddy current signal into a one-dimensional convolutional neural network for spatial size regression. This method combines the physical characteristics of different frequency band signals on surface resolution and deep penetration, overcoming the nonlinear measurement bias caused by conventional fixed single-frequency excitation when facing complex corrosion morphologies, and improving the quantitative evaluation accuracy of defect physical depth and equivalent pore size. Attached Figure Description

[0019] Figure 1 This is a topology diagram of a pipeline defect detection system based on the fusion of magnetic eddy current and machine vision in an embodiment of the present invention. Figure 2 This is the main flowchart of the pipeline defect detection method based on the fusion of magnetic eddy current and machine vision in an embodiment of the present invention; Figure 3 This is a flowchart of visual forward perception and target space calculation in an embodiment of the present invention; Figure 4 This is a schematic diagram of the dynamic gait interception and smooth approximation control principle in an embodiment of the present invention; Figure 5 This is a schematic diagram illustrating the physical locking principle of the robotic arm's posture fixation and lift-off values ​​in an embodiment of the present invention. Figure 6This is a flowchart of adaptive excitation and quantitative defect evaluation in an embodiment of the present invention; Figure 7 This is a comprehensive data chart comparing the approximation braking displacement and detection effect of a robot chassis according to an embodiment of the present invention; in: Figure 7 (a) is a graph of the smooth approximation of the braking displacement; Figure 7 (b) is a comparison chart of physical lock-in state and signal quality; Figure 7 (c) Comparison chart of quantitative evaluation accuracy of surface corrosion pit defects (50 sample set). Detailed Implementation

[0020] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0021] Example: See attached document Figure 1 , Figure 1 This is a topology diagram of a pipeline defect detection system based on the fusion of magnetic eddy current and machine vision according to an embodiment of the present invention.

[0022] This invention provides a pipeline defect detection system based on the fusion of magnetic eddy current and machine vision, which may include: Computing gateway, robot chassis, depth camera, robotic arm, and eddy current probe.

[0023] The computing gateway, as the core processing and scheduling unit of the system, communicates and connects with other hardware modules.

[0024] The robot chassis is connected to the computing gateway. The robot chassis receives underlying motion control commands to execute specified walking states.

[0025] An inertial measurement unit and a motion odometer can be installed inside the robot chassis.

[0026] The inertial measurement unit is used to acquire the yaw angle, pitch angle, roll angle and corresponding inertial motion parameters of the robot chassis, while the motion odometer is used to acquire the position increment information of the robot chassis during the movement.

[0027] The inertial measurement unit and the motion odometer are respectively connected to the computing gateway to provide the computing gateway with real-time pose data of the robot chassis.

[0028] In this embodiment, the robot chassis can be a legged robot chassis, such as a quadruped robot chassis.

[0029] The quadruped robot chassis features dynamic cruising gait, low-speed static crawling gait, and stationary standing gait.

[0030] It is understandable that, given the ability to achieve the aforementioned cruise, deceleration approach, and parking functions, the present invention may also employ other forms of mobile robot chassis.

[0031] The depth camera is mounted on the front of the robot's chassis.

[0032] The depth camera continuously transmits environmental image data containing depth information to the computing gateway via a data bus.

[0033] The robotic arm is mounted on the upper part of the robot chassis.

[0034] A force detection unit can be set in the end effector area of ​​the robotic arm. The force detection unit is used to acquire the contact force or torque feedback data related to the contact force at the end of the robotic arm along the detection axis of the eddy current probe.

[0035] An end-effector compliance mechanism is further installed between the end effector region of the robotic arm and the eddy current probe.

[0036] The end-effector conforming mechanism has at least a passive extensional degree of freedom along the detection axis of the eddy current probe, which is used to absorb the normal micro-displacement disturbance generated by the robot chassis and robotic arm after the eddy current probe comes into contact with the pipe surface.

[0037] The robotic arm has an extended and retracted posture, and receives spatial position and torque control commands from the computing gateway via a field control bus.

[0038] The eddy current probe is placed at the end effector area of ​​the robotic arm.

[0039] The eddy current probe may include an electromagnetic excitation unit and an electromagnetic response detection unit. The electromagnetic excitation unit is used to generate an alternating magnetic field according to the electromagnetic excitation parameters sent by the computing gateway, so as to induce eddy currents inside the conductive pipe. The electromagnetic response detection unit is used to acquire electromagnetic response signals affected by pipe materials and defects. The eddy current probe can also be connected to a computing gateway via signal conditioning and analog-to-digital conversion interfaces.

[0040] The composite magnetic eddy current signal described in this embodiment refers to the electromagnetic response signal formed by the combined effect of the pipe material state and defects after the electromagnetic excitation unit applies an alternating magnetic field to the conductive pipe and induces eddy currents inside the pipe. The electromagnetic response signal includes at least signal components that can characterize the amplitude and phase changes of the detection response.

[0041] The eddy current probe receives the excitation parameters set by the computing gateway, performs electromagnetic excitation and signal acquisition, and transmits the acquired composite magnetic eddy current signal back to the computing gateway.

[0042] See attached document Figure 2 , Figure 2 This is a flowchart of a pipeline defect detection method based on the fusion of magnetic eddy current and machine vision according to an embodiment of the present invention.

[0043] This invention provides a pipeline defect detection method based on the fusion of magnetic eddy current and machine vision, comprising the following steps: In this embodiment, the target spatial information output by machine vision is used to drive the robot chassis and robotic arm to complete the spatial alignment of the eddy current probe. The defect category semantic information output by machine vision is further used to determine the electromagnetic excitation parameters used for magnetic eddy current detection. The electromagnetic response information obtained by magnetic eddy current detection is used to further obtain the dimensional quantification features of the target defect along the pipe wall thickness direction and the transverse direction, thereby forming a hierarchical fusion detection link between visual spatial information, visual semantic information and magnetic eddy current detection information.

[0044] S10, perform visual forward perception and target space calculation. During the robot chassis in dynamic cruising gait, the depth camera acquires image data of the pipe surface, the computing gateway extracts features from the image data, and when it is determined that there is a target defect within the preset detection range, it calculates the three-dimensional coordinates of the target defect in the global world coordinate system, and calculates the normal vector of the local surface of the pipe where the target defect is located based on the local depth points around the target defect. S20: Execute dynamic gait interception and smooth approximation control. The computing gateway generates a hardware state machine interrupt signal to stop the robot chassis's routine path planning task. The robot chassis performs trajectory replanning based on the calculated 3D coordinates. The target parking point is determined based on the 3D coordinates of the target defect and the reachable workspace of the robotic arm. The robot chassis gradually switches from dynamic cruising gait to low-speed static crawling gait, and switches to stationary standing gait after reaching the target parking point. It stops smoothly at the target parking point. After stopping, the image and depth data of the target defect area are reacquired, and the 3D position and local surface normal vector of the target defect are precisely located a second time. S30, execute the robot arm posture fixation and physical locking of the lift-off value. After the robot chassis stops, the robot arm extends outward to drive the eddy current probe to press down on the target defect surface. After the eddy current probe establishes contact with the pipe surface, the end-effector conforming mechanism set between the robot arm end execution area and the eddy current probe undergoes conforming displacement along the normal direction of the local surface of the pipe to absorb the normal displacement disturbance transmitted by the robot chassis and keep the equivalent lift-off value between the detection coil inside the eddy current probe and the pipe surface within the set lift-off value tolerance range. S40, perform adaptive excitation and quantitative defect evaluation. The computing gateway determines the electromagnetic excitation frequency band that matches the corresponding defect category based on the defect category features extracted from the image data, and determines at least two discrete excitation frequencies and corresponding excitation amplitudes within the electromagnetic excitation frequency band to form multiple sets of electromagnetic excitation parameters. The eddy current probe injects an alternating magnetic field into the pipeline in sequence according to the multiple sets of electromagnetic excitation parameters and collects the coupled composite magnetic eddy current signal. The computing gateway performs fusion calculation on the composite magnetic eddy current signals obtained under different excitation frequencies, outputs the size quantification features of the target defect, and after completion, controls the robotic arm to restore the folding posture and controls the robot chassis to restore the dynamic cruising gait.

[0045] The steps in the method provided in the embodiments of the present invention will be described in detail below.

[0046] See attached document Figure 3 , Figure 3 This is a flowchart of visual forward perception and target space calculation according to an embodiment of the present invention.

[0047] In the above method, the process of performing visual forward perception and target space calculation described in step S10 may specifically include the following sub-steps: S101, perform image acquisition and target feature extraction.

[0048] During the dynamic cruise gait of the robot chassis, a depth camera mounted at the front of the robot chassis is configured with a fixed forward tilt angle to cover the surface area of ​​the pipe within a set distance in front.

[0049] The depth camera acquires color image data and matching depth point cloud data of the pipe surface and transmits them to the computing gateway.

[0050] A target detection network is deployed within the computing gateway.

[0051] Specifically, the object detection network is a model built on a lightweight convolutional architecture. Its network topology may include a backbone network for extracting multi-scale image features, a neck network for fusing features, and a head for outputting prediction results.

[0052] To adapt to the computing power limitations of edge computing devices, a fine-grained convolution (GSConv) module can be introduced into the neck network to preserve the semantic features of tiny defects on the pipe surface while reducing the computational load.

[0053] (During forward inference, the computing gateway first performs preprocessing operations such as scaling and normalization on the received color image data to form a preset dimension (e.g., size 1). The tensor is used as the input data for the object detection network.

[0054] After convolutional computation, the object detection network outputs a set containing defect categories, confidence scores, and predicted bounding box coordinates.

[0055] The computation gateway compares the confidence score output by the network with a preset confidence threshold.

[0056] The confidence threshold can be determined empirically by those skilled in the art based on the tolerance for false negative and false positive rates in actual engineering (e.g., the value range is between 0.5 and 0.7).

[0057] When a prediction result has a confidence score greater than a preset confidence threshold, it is determined that a target defect exists in the image field of view, and the corresponding two-dimensional bounding box and semantic label representing the physical properties of the defect are extracted.

[0058] Semantic labels contain basic classification data on defect types, including categories such as surface microcracks or corrosion pits with visible corrosion features and a certain erosion depth along the pipe wall thickness direction.

[0059] The computation gateway extracts the geometric center point of the two-dimensional bounding box and sets it as the positioning reference point of the target defect in the pixel coordinate system.

[0060] Understandably, the aforementioned object detection network needs to be trained before deployment.

[0061] The training process typically involves: constructing an image sample set containing various real-world pipe damage states; The defective regions in the sample set are manually bounded by rectangles and assigned actual defect category labels; The sample set is input into the initial network, and the difference between the predicted value and the true label is calculated using a loss function (e.g., using the full intersection-union (CIoU) loss function to calculate the bounding box regression error, and using the cross-entropy loss function to calculate the classification error). The network weight parameters are iteratively updated through the backpropagation algorithm until the loss function converges.

[0062] When multiple candidate defects that meet the confidence requirements exist in the same image, the computing gateway determines the current target defect based on at least one of the following: the confidence of the candidate defect, the distance information represented by the depth data corresponding to the candidate defect, and the preset defect detection priority.

[0063] Unselected candidate defects do not trigger the current target approach process and are identified and scheduled by the computing gateway in subsequent cruise images.

[0064] S102 performs multi-level coordinate system mapping and global 3D coordinate calculation.

[0065] The computing gateway obtains the pixel coordinate data of the positioning reference point and extracts the positioning reference point and multiple effective depth points in its neighborhood from the depth point cloud data synchronously output by the depth camera. Based on the median, mean or local point cloud fitting results of multiple effective depth points, the true physical depth value corresponding to the target defect is determined to reduce the positioning error caused by noise, missing points and reflections from the pipe surface of a single depth pixel.

[0066] The computing gateway performs forward mapping calculations of three-dimensional coordinates based on the camera imaging geometry model.

[0067] According to the principle of pinhole imaging, an image point on a two-dimensional image corresponds to a ray in three-dimensional space that connects the optical center of the camera to the target object. By combining the depth value of the optical axis direction corresponding to the image point provided by the depth camera, the position parameters representing the target point in physical space can be calculated.

[0068] Let the homogeneous pixel coordinates of the target defect in the pixel coordinate system be... ,in The x-coordinate of the image pixels. The vertical coordinate of the image pixels.

[0069] Let the depth value output by the depth camera at this positioning reference point be... The depth value This indicates the depth of the target point along the optical axis of the depth camera.

[0070] The computation gateway retrieves the pre-calibrated depth camera intrinsic parameter matrix. Solve for the 3D coordinates of the target defect in the depth camera's body coordinate system: ; And further construct its homogeneous coordinates ; When the robot chassis is in dynamic cruising mode, it will generate multi-degree-of-freedom mechanical pose changes, which will cause the spatial normal vector of the depth camera fixed to the robot chassis to be in a state of real-time change.

[0071] To obtain the target contact posture of the subsequent robotic arm end effector area, the computing gateway further selects effective depth points within a preset neighborhood around the target defect positioning reference point, and back-projects the effective depth points onto the depth camera body coordinate system to form a local three-dimensional point cloud set near the target defect.

[0072] The computing gateway performs local surface fitting on the local 3D point cloud set and determines the unit normal vector of the local surface of the pipe where the target defect is located based on the normal direction of the fitted surface. .

[0073] Local surface fitting can employ at least one of least squares plane fitting, principal component analysis, or random sampling consensus algorithm. For cylindrical pipe surfaces, the local surface normal vector at the corresponding location can also be calculated based on the cylindrical surface parameters obtained from the local point cloud fitting.

[0074] The computational gateway uses the three-dimensional position of the target defect and the local surface normal vector together as the data basis for subsequent end-effector position and attitude control.

[0075] To obtain the absolute position of the target defect in the external real physical environment, the computing gateway performs multi-level coordinate system homogeneous transformation operations.

[0076] Let the homogeneous transformation matrix from the camera body coordinate system to the robot chassis body coordinate system be: .

[0077] The matrix for The homogeneous transformation matrix, whose parameters are determined by the mechanical mounting position and orientation of the depth camera on the robot chassis, can be obtained in advance through the camera-robot chassis extrinsic parameter calibration program and called as a known constant matrix when the mechanical mounting relationship between the depth camera and the robot chassis remains unchanged.

[0078] The three-dimensional coordinates of the target defect in the global world coordinate system mentioned below refer to the global world coordinate system established during system initialization. The obtained spatial location.

[0079] The homogeneous transformation matrix from the robot chassis coordinate system to the global world coordinate system is set as follows: .

[0080] The computing gateway reads the robot's body pose angle data and chassis motion odometer data output by the inertial measurement unit inside the robot chassis in real time.

[0081] The color image data and depth data output by the depth camera, the attitude data output by the inertial measurement unit, and the position data output by the motion odometer all carry corresponding acquisition timestamps.

[0082] The computing gateway performs time alignment on the data output by the inertial measurement unit and the motion odometer based on the acquisition timestamp corresponding to the target image; When the data sampling times of different sensors are inconsistent, the computing gateway interpolates the robot chassis pose data at adjacent times based on the target image acquisition time to obtain the robot chassis pose corresponding to the target image acquisition time.

[0083] The computation gateway updates the homogeneous transformation matrix from the robot chassis body coordinate system to the global world coordinate system based on the time-aligned robot chassis pose, in order to reduce the spatial positioning error of the target defect caused by the dynamic motion of the robot chassis.

[0084] Attitude and position data includes the instantaneous yaw, pitch, and roll angles of the fuselage.

[0085] The computing gateway continuously corrects the transformation matrix using real-time updated pose angle data and chassis odometer data. This is to compensate for the external spatial positioning error caused by the dynamic movement of the chassis.

[0086] In this embodiment, a global world coordinate system is established during the system initialization phase. The initial position of the robot chassis when it starts performing the inspection task can be set as the origin of the global world coordinate system, and the direction of each coordinate axis of the global world coordinate system can be determined according to the initial orientation of the robot chassis.

[0087] As described below.

[0088] The computing gateway calculates the absolute three-dimensional coordinates of the target defect in the global world coordinate system based on the above spatial parameters. .

[0089] The absolute coordinate inverse equation is configured as follows: ; For the unit normal vector of the local surface of the pipe where the target defect is located, the computation gateway only uses the rotation component in the homogeneous transformation matrix to perform coordinate system transformation.

[0090] set up The homogeneous transformation matrix The corresponding rotation submatrix, The homogeneous transformation matrix The corresponding rotation submatrix represents the unit normal vector of the local surface of the pipe where the target defect is located in the global world coordinate system. Represented as: The computing gateway reconstructs the two-dimensional pixel information existing in the image sensor into three-dimensional spatial feature information through the above-mentioned homogeneous coordinate transformation operation, and outputs the three-dimensional coordinates of the target defect. As the data foundation for the robot chassis to plan the target parking point, and to include the three-dimensional coordinates of the target defect. and the unit normal vector of the local surface of the pipe where the target defect is located. Together, they serve as the data foundation for subsequent target pose planning at the end of the robotic arm.

[0091] See attached document Figure 4 , Figure 4This is a schematic diagram of dynamic gait interception and smooth approximation control according to an embodiment of the present invention. In the above method, the execution of dynamic gait interception and smooth approximation control process described in step S20 may specifically include the following sub-steps: S201, execution state machine interrupt and normal path pause.

[0092] Once the computing gateway calculates the absolute three-dimensional coordinates of the target defect through the visual network, the system needs to switch from the conventional macroscopic environmental inspection task to the microscopic local inspection task.

[0093] The computing gateway generates a hardware state machine interrupt signal with a preset priority and sends the signal to the robot chassis via the control bus.

[0094] After receiving the interruption signal, the robot chassis suspends the execution of the original global navigation path planning nodes, hands over the control of the chassis to the local approximation control logic, and enters the target approximation control state.

[0095] S202, perform smooth deceleration trajectory replanning.

[0096] The robot chassis has a fixed initial running speed during dynamic cruise gait.

[0097] If direct braking is performed after receiving the target coordinates, the inertial force generated by the sudden speed change may cause the robot chassis to become unstable and overturn.

[0098] Therefore, the computational gateway uses a polynomial interpolation algorithm to generate a smooth deceleration curve, applies continuity constraints to the displacement and velocity during the braking process, and limits the braking time and trajectory parameters to ensure that the reference acceleration corresponding to the braking process is within the acceleration range allowed by the robot chassis drive mechanism.

[0099] The calculation gateway obtains the current initial velocity of the robot chassis. and the current physical location of the fuselage center .

[0100] Based on the pre-set optimal operating distance range of the robotic arm, the computing gateway calculates the coordinates of a target parking point located within the working radius of the robotic arm, using the three-dimensional coordinates of the target defect as a reference. .

[0101] In practical implementation, the coordinates of the target parking point The settings need to ensure that the robot chassis maintains a preset projection offset distance relative to the target defect within the current drivable support plane of the robot chassis, so as to ensure that after the robot chassis stops, the target defect falls into the reachable workspace envelope of the robotic arm and avoids interference with the robot body.

[0102] Computing gateway computer body center physical location Coordinates of the target parking point Straight-line braking distance The distance It can be calculated and The distance between the projected coordinates on the horizontal plane is obtained.

[0103] Set the desired braking time as .

[0104] The expected braking time It is limited by the maximum available deceleration of the robot chassis drive motor. .

[0105] To avoid saturation overload of the bottom motor, braking time The value of must satisfy the inequality constraints: ; The computing gateway can calculate the current speed of the robot chassis. Distance to the target parking point and the maximum allowable deceleration of the robot chassis. Adaptive determination of braking time .

[0106] The above inequality is used to determine the initial lower limit of the braking time.

[0107] The computing gateway determines the braking time. After obtaining the reference displacement trajectory, the corresponding reference acceleration is calculated based on the reference displacement trajectory, and it is determined whether the absolute value of the reference acceleration over the entire braking time interval exceeds the maximum permissible deceleration. .

[0108] When the maximum permissible deceleration is exceeded, the braking time is increased. Alternatively, the target parking point can be readjusted until the desired result is achieved: ; In an exemplary operating condition, when the above parameters meet the corresponding braking constraints, the braking time is... The time can be between 0.5 seconds and 1.5 seconds; Under other operating conditions, the braking time is adjusted accordingly based on the current motion state, and is not limited to the range of the examples above.

[0109] When the calculated remaining distance from the current position of the robot chassis to the target parking point is insufficient to complete normal braking within the allowable deceleration range, the computing gateway does not perform instantaneous forced braking. Instead, it re-determines the target parking point that meets the braking distance requirements and is within the reachable workspace of the robotic arm, or controls the robot chassis to cross the current target and re-plan the approach trajectory to avoid robot chassis instability due to sudden changes in braking torque.

[0110] Set the motion time variable as Establish a framework for time Third-order polynomial displacement function ; In the formula, , , and The constant coefficients of the polynomial to be solved are denoted as .

[0111] To ensure a smooth trajectory, the dynamic boundary conditions are substituted into the calculation of the gateway: At the initial moment, displacement The derivative of the initial velocity, i.e., the displacement. At the moment of braking termination, displacement Termination speed is the speed at which braking is completed. .

[0112] The computational gateway obtains the analytical expression for the reference displacement trajectory during the smooth deceleration process by solving the above system of equations: ; The analytical expression for the reference displacement trajectory specifies the robot's movement from 0 to... The continuous displacement state within the time interval allows for smooth changes in the reference displacement and reference velocity of the robot chassis, and this is achieved by controlling the braking time. Constraints are applied to keep the corresponding reference acceleration within the allowable range of the chassis drive mechanism.

[0113] S203 performs underlying gait switching and smooth docking.

[0114] The robot chassis's underlying motion control unit is based on the reference displacement trajectory generated above. It can calculate the desired angle and output torque of each leg joint in real time.

[0115] At the principle level, smooth deceleration requires the robot to change the support period of its feet on the ground.

[0116] The robot chassis adjusts the phase difference between the gait frequency and the support phase based on the virtual model control mechanism, gradually downgrading the original high-frequency dynamic gait such as diagonal fast step.

[0117] As the current speed gradually decreases, the robot chassis first switches from dynamic cruising gait to low-speed static crawling gait, and uses the static crawling gait to complete low-speed position correction near the target parking point; after the robot chassis reaches the target parking point and the current speed decreases below the preset parking threshold, the robot chassis further switches to a stationary standing gait.

[0118] Through the aforementioned dynamic control process, the robot chassis gradually consumes kinetic energy while stopping at the target parking point coordinates in accordance with the preset parking position requirements. Place.

[0119] For conventional legged robot control logic, such as the calculation of the Jacobian matrix at the foot end and the single-leg spring damping model involved in virtual model control, those skilled in the art can write programs and optimize parameters based on the existing motion control framework. This is well-known technology in the field and will not be elaborated here.

[0120] S204, perform secondary fine positioning of the target after parking.

[0121] After the robot chassis enters the stationary standing gait, the computing gateway controls the depth camera to reacquire color images and depth data containing the target defect area, and re-identifies the target defect in the corresponding local search area based on the three-dimensional coordinates of the target defect obtained during the robot chassis's dynamic cruise phase.

[0122] The computing gateway recalculates the 3D position of the target defect and the normal vector of the local surface of the pipe where the target defect is located based on the image and depth data re-acquired after parking. It also updates the 3D coordinates of the target defect in the global world coordinate system using the current pose of the robot chassis after parking. and unit normal vector The updated version and The target precision positioning result is used for end-effector control of robotic arms.

[0123] The subsequent target approximation and attitude determination of the robotic arm prioritizes the target precision positioning results obtained after parking, thereby reducing the impact of odometer cumulative error, foot slippage error and body posture changes on the contact positioning accuracy of the robotic arm caused by dynamic deceleration, gait switching and parking of the robot chassis.

[0124] See attached document Figure 5 , Figure 5 This is a schematic diagram of the physical locking principle of the robot arm's posture fixation and lift-off values ​​according to an embodiment of the present invention.

[0125] In the above method, the physical locking process for the robot arm's pose determination and lift-off value described in step S30 may specifically include the following sub-steps: S301 performs target approximation and kinematic pose determination for the robotic arm.

[0126] After the robot chassis completes docking, the computing gateway retrieves the three-dimensional coordinates of the target defect and the unit normal vector of the local surface of the pipe where the target defect is located. Combined with the current docking pose of the robot chassis and the pre-calibrated extrinsic parameters of the robotic arm base relative to the robot chassis, the target position and local surface normal direction of the target defect in the coordinate system of the robotic arm base are obtained through inverse coordinate transformation.

[0127] The computing gateway constructs the target spatial pose of the robotic arm end based on the target location and the normal direction of the local surface, so that the detection axis of the eddy current probe is basically parallel to the normal direction of the local surface of the pipe where the target defect is located, and the detection end of the eddy current probe faces the pipe surface.

[0128] Based on the target spatial pose of the robotic arm end effector, the computation gateway uses the robotic arm inverse kinematics algorithm to solve the target joint angles corresponding to each joint, and generates robotic arm motion control commands based on the target joint angles.

[0129] The robotic arm receives angle commands from the computing gateway, breaks free from its initial folded posture, drives all joints to extend in tandem, and guides the eddy current probe at its end to approach the detection position corresponding to the target defect along a direction that is collinear with the unit normal vector of the local surface of the pipe where the target defect is located and points towards the pipe surface, until the eddy current probe contacts the pipe surface and applies a preset downward pressure.

[0130] The preset downward pressure can be achieved by monitoring the feedback value of the force detection unit at the end of the robotic arm, such as by obtaining the contact force component along the detection axis of the eddy current probe through a six-dimensional force / torque sensor, combined with the torque closed-loop control of each joint motor, to prevent excessive downward pressure from damaging the probe or insufficient downward pressure from causing poor contact.

[0131] For the specific DH parameter modeling and Jacobian matrix inverse solution algorithm of the robotic arm, those skilled in the art can configure it according to conventional robotics theory. It is a well-known technology in this field and will not be elaborated here.

[0132] S302, executes end-effector deformation and lift-off value locking.

[0133] The lift-off value characterizes the physical distance between the effective detection surface of the detection coil inside the eddy current probe and the metal surface of the pipe being tested, along the local surface normal direction.

[0134] In this embodiment, the eddy current probe has a wear-resistant contact surface that contacts the pipe surface, and the detection coil is fixedly disposed inside the wear-resistant contact surface.

[0135] When the wear-resistant contact surface maintains stable contact with the pipe surface, the equivalent lift-off value between the detection coil and the pipe metal surface is determined by the fixed mechanical packaging size of the probe itself.

[0136] In actual working environments, even when the robot chassis is parked, fluctuations in the holding torque of its leg motors and environmental factors such as external wind loads can still cause disturbances in the normal displacement of the robot body.

[0137] If this minute disturbance is directly and rigidly transmitted to the eddy current probe, it will cause the lift-off value to change dynamically, thereby affecting the characteristics of the subsequent electromagnetic signal.

[0138] To reduce or even eliminate this normal displacement disturbance, a passive end-effector conformation mechanism is installed between the end effector region of the robotic arm and the eddy current probe.

[0139] The end-effector specifically includes a linear sliding shaft arranged along the probe normal, an elastic compression spring sleeved on the linear sliding shaft, and a viscous damper arranged in parallel.

[0140] To adapt to the curvature of the pipeline and the small angular positioning error at the end of the robotic arm, a passive attitude adjustment connection structure with a preset small angular swing degree of freedom can be set between the end-effector and the eddy current probe, so that the wear-resistant contact surface of the eddy current probe can be slightly adaptively adjusted according to the local surface posture of the pipeline after contacting the pipeline.

[0141] In another embodiment, the wear-resistant contact surface of the eddy current probe can be configured as an arc-shaped contact surface adapted to the curvature of the outer surface of the pipe to be tested.

[0142] The eddy current probe is slidably connected to the bottom end of the linear sliding shaft.

[0143] When the robotic arm performs a downward pressing action to bring the eddy current probe into contact with the pipe surface, the elastic compression spring generates an initial compression, providing a stable static contact pre-pressure to the eddy current probe.

[0144] When the robot chassis experiences a normal displacement disturbance, the end effector undergoes passive mechanical elastic deformation to absorb the disturbance energy.

[0145] Establish the single-degree-of-freedom normal dynamic equilibrium equations for this physical process: ; In the formula, This represents the dynamic relative displacement of the end-effector when subjected to disturbance. The velocity represents the relative displacement. Acceleration representing relative displacement, This represents the equivalent physical mass of the eddy current probe and its follower components. This represents the damping coefficient of a viscous damper. This represents the stiffness coefficient of an elastic compression spring. This indicates the static contact preload generated by the downward pressure of the robotic arm. This represents the transient contact force between the eddy current probe and the pipe surface.

[0146] During the system physical parameter design phase, to prevent the elastic mechanism from undergoing continuous oscillations under disturbances, the system needs to be configured as a critically damped or overdamped state, i.e., the damping coefficient needs to be adjusted accordingly. Stiffness coefficient With equivalent physical mass The relationship between them is satisfied .

[0147] In addition, to reduce the disturbance amplitude at the expected maximum normal displacement The possibility of the lower eddy current probe detaching will As one of the constraints for selecting static contact preload, and combined with the dynamic response of the end-effector to influence transient contact force. Perform a check to ensure that the transient contact force under normal operating conditions remains greater than zero.

[0148] The effective compliance stroke of the end-effector along the probe detection axis is greater than or equal to the maximum normal displacement disturbance amplitude expected by the system, and a preset safety margin is reserved to prevent the linear sliding shaft from reaching the mechanical limit position under the maximum disturbance state, so that the displacement disturbance of the robot chassis is rigidly transmitted to the eddy current probe again.

[0149] By designing the constraints on the damping ratio and static preload as described above, it is ensured that the wear-resistant contact surface of the eddy current probe remains in contact with the pipe surface and will not physically detach.

[0150] By relying on the predetermined fixed mechanical encapsulation distance between the wear-resistant contact surface of the eddy current probe and the internal detection coil, as long as the wear-resistant contact surface continues to maintain contact with the pipe surface, the equivalent lift-off value between the internal detection coil and the pipe metal surface can be maintained in a preset state.

[0151] Through the compliant absorption process of the above-mentioned purely mechanical structure, the equivalent lift-off value between the effective detection surface of the internal detection coil of the eddy current probe and the metal surface of the pipe is kept within the set lift-off value tolerance range, providing stable physical boundary conditions for subsequent quantitative analysis of electromagnetic signals.

[0152] See attached document Figure 6 , Figure 6 This is a flowchart of adaptive excitation and quantitative defect evaluation according to an embodiment of the present invention.

[0153] In the above method, the adaptive excitation and quantitative defect evaluation process described in step S40 may specifically include the following sub-steps: S401 performs semantic-driven electromagnetic parameter mapping and signal acquisition.

[0154] The computation gateway establishes parameter mapping relationships based on the defect category features extracted in the preceding steps.

[0155] The parameter mapping relationship can be pre-established through offline calibration tests before the system is put into operation. Specifically, for the material and wall thickness parameters of the pipeline to be tested, standard calibration specimens containing different defect types, different defect depths, and different transverse dimensions are used to collect the corresponding magnetic eddy current responses at multiple candidate excitation frequencies and excitation amplitudes. Based on at least one evaluation index among defect response amplitude, phase difference, signal-to-noise ratio, or defect parameter prediction error, the preferred excitation frequency band and excitation amplitude range corresponding to different defect types are determined, and the resulting mapping relationship between pipeline material parameters, pipeline wall thickness, defect type, and electromagnetic excitation parameters is pre-stored in the computing gateway.

[0156] Specifically, the penetration ability of electromagnetic fields within metal pipes is limited by the skin effect, and the computational gateway has a pre-defined skin depth equation: .

[0157] In the formula, The depth of penetration of the alternating magnetic field. The excitation frequency is one of the electromagnetic excitation parameters. The magnetic permeability of the pipe material. The electrical conductivity of the pipe material.

[0158] Before performing an inspection task, at least one of the following basic parameters of the pipeline to be inspected—material type, magnetic permeability, electrical conductivity, and pipe wall thickness—can be pre-stored in the computing gateway.

[0159] For pipelines with consistent material parameters in the same inspection task, the computing gateway directly calls the pre-stored corresponding material parameters; For inspection targets that include pipe sections of different materials or specifications, the corresponding material parameters can be retrieved based on the identification information of the currently inspected pipe section.

[0160] When the semantic label output in step S10 is a surface microcrack category, the computing gateway determines the corresponding high-frequency excitation band by looking up a table based on the above equation, and selects multiple discrete excitation frequencies within the high-frequency excitation band to concentrate electromagnetic energy on the shallow surface of the pipe to improve the sensitivity of surface defect detection. When the semantic label is a corrosion pit category, the gateway mapping is calculated to determine the corresponding relatively low-frequency excitation band, and multiple discrete excitation frequencies are selected within the relatively low-frequency excitation band to increase the penetration depth δ, thereby obtaining the electromagnetic response related to the corrosion depth.

[0161] The computational gateway determines the excitation amplitude corresponding to each discrete excitation frequency based on the pre-established parameter mapping relationship, and forms at least two sets of electromagnetic excitation parameters with different excitation frequencies.

[0162] The eddy current probe sequentially performs electromagnetic excitation and signal acquisition according to each set of electromagnetic excitation parameters, thereby obtaining the multi-frequency composite magnetic eddy current response of the target defect under different electromagnetic penetration depths.

[0163] Let the number of discrete excitation frequencies used in a single quantitative detection be . Each excitation frequency is expressed as: The excitation amplitude corresponding to each excitation frequency is expressed as follows: .

[0164] Data was collected at each excitation frequency. Composite magnetic eddy current signal with discrete sampling points.

[0165] The standard electromagnetic signal acquisition and analog-to-digital conversion procedures are executed simultaneously during this stage.

[0166] For routine operations involving alternating electromagnetic field excitation and underlying hardware data acquisition, those skilled in the art can use general-purpose data acquisition cards and programmable signal generation circuits for configuration, which are well-known technologies in the field and will not be elaborated here.

[0167] S402, Perform quantitative assessment of defects and system status reset.

[0168] After acquiring the composite eddy current signal, the computational gateway first performs preprocessing operations such as bandpass filtering and phase rotation on the raw signals obtained at each excitation frequency, extracts the real and imaginary voltage components corresponding to the time sequence of the signal, and then... The real and imaginary signals obtained at each excitation frequency are concatenated according to the channel dimension to construct a dimension of The multi-frequency two-dimensional feature tensor is used as the model input, where The number of discrete sampling points corresponding to each excitation frequency. The number of discrete excitation frequencies used for a single quantitative detection.

[0169] The computing gateway inputs the preprocessed multi-frequency two-dimensional feature tensor into the shared feature extraction network of the quantitative prediction neural network to extract high-level features of the magnetic eddy current signal.

[0170] This quantitative prediction neural network employs a regression network architecture that includes hidden layers.

[0171] The computing gateway also normalizes and encodes the excitation frequency and amplitude parameters used in this detection to form an electromagnetic excitation parameter vector. The electromagnetic excitation parameter vector is then fused with the high-level features extracted from the multi-frequency two-dimensional feature tensor by the quantitative prediction neural network and input into the fully connected layer at the end of the network, so that the quantitative prediction neural network can distinguish the magnetic eddy current response under different electromagnetic excitation conditions.

[0172] In practice, the quantitative prediction neural network can adopt a one-dimensional convolutional neural network (1D-CNN) model.

[0173] Its necessary internal hierarchical structure includes: One-dimensional convolutional and max-pooling layers are used to extract local temporal variations and response differences corresponding to different excitation frequencies in the aforementioned multi-frequency two-dimensional feature tensors, as well as fully connected layers at the end of the network used for feature tiling and numerical mapping.

[0174] Before entering the final regression output layer, the high-level features of magnetic eddy currents extracted by the convolutional network can be concatenated with the normalized electromagnetic excitation parameter vector, and the corresponding regression output head can be called according to the defect category information output by the machine vision.

[0175] The fully connected layer directly outputs the final size quantization features of the target defect.

[0176] For the corrosion pit category, the size quantification feature can include a quantified value characterizing the physical depth of the corrosion pit. And quantitative values ​​characterizing the equivalent lateral dimensions of corrosion pits ; For each crack category, the size quantization feature can include a quantized value characterizing the physical depth of the crack. And the quantized value characterizing the transverse feature size of the crack .

[0177] The quantitative prediction neural network can be configured with a shared feature extraction network and regression output heads corresponding to different defect categories. The computation gateway selects the corresponding regression output head according to the semantic label of the defect category obtained in step S10, thereby obtaining the size quantification features corresponding to the current defect category.

[0178] The computing gateway can also determine whether the currently acquired signal meets the preset quantitative conditions based on the amplitude range, signal-to-noise ratio, or model prediction confidence index of the composite magnetic eddy current signal.

[0179] When the signal quality is lower than a preset threshold, the computing gateway controls the robotic arm to readjust the contact pre-pressure or re-execute at least one electromagnetic signal acquisition. If the signal quality conditions are still not met after a preset number of repeated acquisitions, the location, defect type, and abnormal detection status of the corresponding target defect are recorded, and the quantitative calculation of the current target ends.

[0180] When the quantitative calculation is successful, the calculation gateway writes the world coordinates, defect category, and size quantification results of the target defect into the list of detected targets. When quantitative calculation fails, the calculation gateway writes the world coordinates of the target defect, the defect category, and the detection anomaly identifier into the abnormal target list.

[0181] Regardless of whether the quantitative calculation of the current target defect is successful or terminated due to signal quality abnormalities, the computing gateway will issue a system reset signal after the current detection process ends.

[0182] See attached document Figure 7 , Figure 7 This is a comprehensive data chart comparing the robot chassis approach braking displacement and detection effect according to an embodiment of the present invention.

[0183] During the inspection of a carbon steel crude oil pipeline operating outdoors in a chemical plant area, the relative magnetic permeability of the pipeline material was set. Approximately 200, conductivity Approximately S / m, wall thickness is 8.0mm.

[0184] The robot chassis employs a quadrupedal locomotion configuration and is moving along a maintenance passage parallel to the pipeline axis using a dynamic cruising gait. Its current initial operating speed is... m / s.

[0185] The computing gateway receives the image transmitted back from the depth camera, extracts it through a lightweight convolutional network, determines that there is a surface corrosion defect in front, and outputs a classification confidence score of 0.88.

[0186] The computing gateway synchronously calculates the global three-dimensional absolute coordinates of the target defect and measures the straight-line braking distance from the center of the robot chassis to the target parking point. m.

[0187] At this point, a system state machine interrupt is triggered, and the system enters the smooth deceleration trajectory replanning phase.

[0188] The maximum allowable deceleration of the drive motor at the bottom of the quadruped robot's chassis. .

[0189] Based on the braking time constraint formula described above, the desired braking time is... satisfy: ; To ensure a smooth braking process and to reserve redundant cycles for gait switching control by the underlying virtual model, the computation gateway will calculate the braking time. The adaptive configuration is set to 2.0s.

[0190] Substituting the aforementioned analytical expression for the reference displacement trajectory, we solve for the coefficients of the third-order polynomial: ; ; ; This yields the specific reference displacement trajectory equation, corresponding to the appendix. Figure 7 Braking displacement curve in the middle: ; Taking its derivative, we can see that, At time s, the braking termination speed converges to 0 m / s.

[0191] Based on the planned curve, the robot chassis smoothly downgrades from a diagonal fast walk to a low-speed crawl within 2.0 seconds, and finally stops steadily at the target parking point with a stationary standing gait.

[0192] After docking, the robotic arm extends and drives the eddy current probe loaded at the end to press down.

[0193] Based on the fluctuation characteristics of actual wind load and the static holding torque of the chassis motor, the maximum expected displacement disturbance amplitude of the robot chassis in the normal direction under this working condition was determined. m.

[0194] Equivalent physical mass of the end-effector conforming mechanism kg. To satisfy the damping ratio formula When the system enters an overdamped or critically damped state, select the stiffness coefficient of the elastic compression spring. N / m, and configure the damping coefficient of the viscous damper N·s / m.

[0195] Combining constraint formulas The gateway calculates the downward stroke of the robotic arm to generate static contact preload. N.

[0196] At this point, the minimum extreme value of the transient contact force N remains greater than 0.

[0197] This parameter configuration ensures, at the physical level, that when the chassis experiences a 15mm normal vibration, the wear-resistant contact surface of the eddy current probe still adheres to the metal surface of the pipe by relying on a transient contact force of more than 12.5N.

[0198] During the electromagnetic detection phase, since the visual semantic label is corrosion pit, the computing gateway retrieves a preset low-frequency mapping table based on the skin effect equation and sends the data to the eddy current probe. kHz and The dual-frequency electromagnetic excitation parameters are measured at kHz. The acquired composite magnetic eddy current signal is input into a one-dimensional convolutional neural network, which then regresses to output the quantized value of the predicted physical depth of the corrosion pit. mm, quantized value of transverse equivalent aperture 5.8mm.

[0199] After the inspection is completed, the system is reset and resumes the patrol task at 0.8 m / s.

[0200] To objectively evaluate the data reliability and detection accuracy of this system, comparative experiments were conducted using standard artificial defect calibration blocks with known precision at the core technical nodes of the above-mentioned specific application embodiments. The relevant statistical histograms and scatter plots are presented in the appendix. Figure 7 middle.

[0201] In the verification of the physical locking effect of the lift-off value, the test environment was artificially introduced with periodic chassis normal oscillations with an average value of 10mm.

[0202] The control group used a traditional structure with an eddy current probe rigidly connected to the end of a robotic arm, while the experimental group used the detection system with an end-effector conforming mechanism of this invention. Experimental data showed that the traditional rigid detection system, due to the high-frequency detachment of the probe from the pipe wall, exhibited an equivalent lift-off value fluctuation of up to 12.4 mm, a signal-to-noise ratio drop to 6.2 dB, and baseline drift resulted in the inability to extract 64% of the signal features from deep corrosion pits. The system of the present invention measures that the transient contact force between the probe and the surface of the test block is maintained in the range of 8.0N to 26.5N throughout the entire cycle, the physical fluctuation range of the lift-off value is less than 0.2mm, and the corresponding magnetic eddy current signal signal-to-noise ratio is stable at more than 34.5dB.

[0203] In the accuracy verification test of defect quantitative evaluation, for the prediction of corrosion pits with a physical depth range of 1.0 mm to 6.0 mm, the root mean square error (RMSE) of the prediction results of 50 samples was calculated using the actual geometric size of the standard test block as the benchmark.

[0204] Statistical data shows that the root mean square error of depth prediction using the traditional fixed single-frequency excitation combined with phase angle analysis method is 0.85 mm. The visual semantic-driven adaptive multi-frequency excitation combined with network regression method of the present invention reduces the root mean square error of depth prediction to 0.18 mm and the measurement nonlinear error is significantly converged, thus providing a reliable technical implementation method for automated nondestructive testing of robots in complex engineering environments.

[0205] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A pipeline defect detection method based on the fusion of magnetic eddy current and machine vision, characterized in that, Includes the following steps: Visual forward perception and target space calculation steps: Images of the pipe surface are acquired using a depth camera, defect features are extracted, and the global three-dimensional coordinates of the target defect are calculated. Dynamic gait interception and smooth approximation control steps: The braking distance between the target defect and the robot chassis is calculated based on the global three-dimensional coordinates, a smooth deceleration trajectory is planned, and the robot chassis is controlled to travel along the smooth deceleration trajectory to the target parking point. Physical locking steps for robotic arm attitude fixation and lift-off values: The robotic arm is controlled to drive the eddy current probe loaded at the end to approach the target defect. Static contact pre-pressure is generated through the end-effector conforming mechanism to control the eddy current probe to adhere to the pipe surface. Adaptive stimulus and quantitative defect evaluation steps: Configure the electromagnetic excitation parameters of the eddy current probe according to the defect characteristics, collect composite magnetic eddy current signals, and generate quantitative evaluation results for the target defect based on the composite magnetic eddy current signals.

2. The pipeline defect detection method based on the fusion of magnetic eddy current and machine vision according to claim 1, characterized in that, The specific process of the visual forward perception and target space calculation steps includes: The defect features in the image are extracted using a lightweight convolutional network, and the defect classification confidence score is output. Based on the defect characteristics, a multi-level coordinate system mapping is performed to calculate the global three-dimensional coordinates of the target defect in the global coordinate system.

3. The pipeline defect detection method based on the fusion of magnetic eddy current and machine vision according to claim 1, characterized in that, The process of planning a smooth deceleration trajectory in the dynamic gait interception and smooth approximation control steps includes: The straight-line distance from the center of the robot chassis to the target parking point is calculated and used as the braking distance. Trigger a system state machine interrupt to obtain the current initial running speed of the robot chassis and the maximum allowable deceleration of the underlying drive motor; Calculate the minimum threshold of the desired braking time based on the initial operating speed and the maximum deceleration, and configure the braking time to be a time value greater than or equal to the minimum threshold; Based on the braking distance, the initial operating speed, and the braking time, a reference displacement trajectory equation containing third-order polynomial coefficients is generated as the smooth deceleration trajectory.

4. The pipeline defect detection method based on the fusion of magnetic eddy current and machine vision according to claim 3, characterized in that, The expression for the reference displacement trajectory equation is: ; in, For time The changing braking displacement, Let be the coefficients of a third-order polynomial, and satisfy . , ; in, The initial running speed, The braking distance is... The braking time is mentioned.

5. The pipeline defect detection method based on the fusion of magnetic eddy current and machine vision according to claim 1, characterized in that, The process of controlling the robot chassis to travel along the smooth deceleration trajectory to the target parking point includes: The robot chassis downgrades its dynamic cruising gait to a low-speed crawling gait based on the smooth deceleration trajectory, and switches to a stationary standing gait when it reaches the target parking point. Under the stationary standing gait, the target defect is precisely located a second time.

6. The pipeline defect detection method based on the fusion of magnetic eddy current and machine vision according to claim 1, characterized in that, The end-effector compliance mechanism includes an elastic compression spring. In the physical locking step of the robotic arm's pose fixation and lift-off value, the process of generating static contact pre-pressure through the end-effector compliance mechanism includes: Obtain the maximum expected displacement disturbance amplitude of the robot chassis in the normal direction; Controlling the downward stroke of the robotic arm to generate a constraint formula The static contact preload; in, For the static contact preload, Let be the stiffness coefficient of the elastic compression spring. The maximum displacement disturbance amplitude is given.

7. The pipeline defect detection method based on the fusion of magnetic eddy current and machine vision according to claim 1, characterized in that, The end-effector also includes a viscous damper; The damping coefficient of the viscous damper satisfies in, The damping coefficient of the viscous damper is... This is the stiffness coefficient of the elastic compression spring. The equivalent physical mass of the end-effector conforming mechanism.

8. The pipeline defect detection method based on the fusion of magnetic eddy current and machine vision according to claim 1, characterized in that, The specific process of the adaptive excitation and quantitative defect evaluation step includes: The defect features are used as visual semantic labels. Based on the skin effect equation, a preset frequency mapping table is retrieved, and dual-frequency electromagnetic excitation parameters are sent to the eddy current probe as electromagnetic excitation parameters. The collected composite magnetic eddy current signal is input into a one-dimensional convolutional neural network for network regression, and the predicted physical depth and lateral equivalent pore size of the corrosion pit are output as the quantitative evaluation result. After completing the quantitative evaluation of defects, the system status is reset.

9. An electronic device, characterized in that, include Processor and memory; The memory is used to store computer programs; The processor is used to execute the computer program to implement the pipeline defect detection method based on the fusion of magnetic eddy current and machine vision as described in any one of claims 1 to 8.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program that, when executed by a processor, implements the pipeline defect detection method based on the fusion of magnetic eddy current and machine vision as described in any one of claims 1 to 8.

Images