A quadruped robot inspection and risk early warning system based on multi-source perception

Abstract

The present application relates to the technical field of quadruped robots, and discloses a quadruped robot inspection and risk early warning system based on multi-source sensing, which comprises a feature quantization module, a frequency domain detection module, a fault monitoring module and a decoupling control module. The feature quantization module calculates the material uncertainty based on environmental data and timely outputs a ground exploration trigger flag; the frequency domain detection module responds to the flag and issues a perturbation torque instruction to the leg, and combines a vibration response signal to solve contact parameters to determine whether to re-plan a path; the fault monitoring module issues a frequency sweep torque instruction when the leg is empty, generates a degradation index and outputs a hardware protection instruction; and the decoupling control module uses the above instructions as a feedforward signal, filters out the body same-frequency interference component, and outputs smooth macroscopic pose data. The present application overcomes the limitations of single visual sensing, realizes closed-loop control of contact verification, dynamic detection and stability decoupling, and improves the operation safety and stability of the robot in complex environments.

Description

Technical Field

[0001] This invention relates to the field of quadruped robot technology, specifically to a quadruped robot inspection and risk warning system based on multi-source perception. Background Technology

[0002] With the development of intelligent equipment technology, quadruped robots are widely used in inspection tasks in complex and unknown environments. Existing robots mainly rely on visual sensors for environmental mapping and obstacle avoidance planning. However, visual perception struggles to penetrate the surface and directly obtain the physical stress properties of the ground material. When facing soft, collapsed, or mirror-like water-reflective geology, the visual system is prone to misjudgment. When the robot transfers its weight entirely to the target support leg based on purely visual planning, if the contact surface slips or sinks, the passive force detection after contact often exhibits significant response lag, causing the robot to completely lose balance before performing any avoidance maneuvers.

[0003] Meanwhile, under prolonged high-intensity operation, quadruped robots experience increased microscopic backlash and fatigue wear in their internal joint actuators, reducers, and transmission links. Existing hardware monitoring methods typically rely on periodic shutdowns of external equipment for inspection, or can only monitor macroscopic over-limit conditions of current and temperature. They cannot perform frequency domain flaw detection on the pure transmission characteristics of internal mechanical links during dynamic gait during continuous task execution, making it difficult to trigger underlying proactive defense actions before catastrophic damage to transmission components. To address these detection and sensing issues, if a high-frequency physical excitation signal is actively injected into the legs during movement to detect environmental impedance or hardware transfer function, this excitation stress wave will propagate upwards along the rigid structure to the torso, causing high-frequency interference pulses to superimpose on the state acquisition data of the onboard inertial measurement unit. This results in state estimation drift, disrupting the closed-loop stability of the system's top-level macroscopic gait calculation and attitude stabilization control. Summary of the Invention

[0004] To address the shortcomings of existing technologies, this invention provides a quadruped robot inspection and risk warning system based on multi-source perception. This system solves the problems of quadruped robots being prone to instability due to visual misjudgment caused by complex materials and passive perception lag, being unable to monitor the mechanical degradation of internal transmission links online during continuous tasks, and causing structural data crosstalk to the top-level macroscopic attitude stabilization control when actively implementing physical vibration detection.

[0005] To achieve the above objectives, the present invention is implemented through the following technical solution: a quadruped robot inspection and risk warning system based on multi-source perception. The system mainly includes a feature quantization module, a frequency domain detection module, a fault monitoring module, and a decoupling control module. Through multimodal data interaction and feedforward control of hardware status, the quadruped robot can actively identify terrain risks and protect its own mechanical health.

[0006] Specifically, the feature quantization module performs a priori assessment of the environment at the robot's predicted landing point by fusing color environmental image data and 3D depth point cloud data. This module calculates the local pixel variance of the image to extract texture divergence features, extracts the proportion of pixels exceeding the brightness limit to obtain reflectivity feature parameters, and calculates the ratio of invalid data points within the local grid of the point cloud to obtain depth missing density features.

[0007] The system performs linear superposition of the above three types of features according to preset weights to quantify the material uncertainty of the target area. When the uncertainty is greater than the set confidence level safety threshold, it indicates that there is a risk of slippage or collapse, and the system immediately outputs a ground exploration trigger flag.

[0008] The frequency domain detection module responds to the ground-penetrating trigger flag and executes vision-guided active tactile detection. This module first matches the corresponding excitation frequency parameters from a built-in mapping data table based on the dominant visual feature type causing the material uncertainty to exceed limits, and then generates a sinusoidal perturbation torque command at a specific frequency. To avoid the detection action affecting the stability of the aircraft, the system monitors the normal contact reaction force at the foot tip and the displacement of the aircraft's center of mass to accurately capture the virtual load time window when the target leg has just touched the ground and is not yet fully bearing weight, and issues a superimposed torque command within this window.

[0009] Subsequently, a Fast Fourier Transform (FFT) is simultaneously performed on the time series corresponding to the test torque signal and the broadband vibration velocity response time series to extract the resonance peak amplitude and phase angle deviation of the frequency response function within the test frequency band. Based on this, the equivalent contact stiffness and estimated dynamic friction coefficient are calculated and output. When the calculated contact parameters are lower than the physical safety threshold, the system actively outputs a path replanning command to avoid high-risk areas.

[0010] The fault monitoring module utilizes the alternating gait characteristics of quadruped robots. When the target leg is in a pure, airborne state with its limbs off the ground and the resultant force of the three-dimensional forces within the zero-noise threshold, it sends a broadband sweep torque command with a linearly increasing frequency to the leg joint. The frequency band completely covers the natural frequency range of the transmission mechanical components. By calculating the ratio of the sweep excitation to the vibration response through frequency domain transformation, the real-time structural transfer function is obtained. The system then generates a degradation index characterizing mechanical wear or loosening by calculating the integral of the sum of squares of the frequency band deviations between this function and the health baseline function.

[0011] Based on the preset list of graded degradation thresholds, the system executes a step-by-step hardware protection strategy: reducing the maximum running speed of the joint in the case of mild degradation; limiting the maximum output torque and sending a request to the gait planning program to modify the calculation logic of the force polygon in the case of moderate degradation; and directly shutting off the stator current and triggering the electromagnetic brake in the case of severe degradation to prevent catastrophic mechanical failure.

[0012] To address the high-frequency mechanical vibrations introduced during frequency domain detection and fault monitoring, the decoupling control module employs an adaptive notch filter based on the least mean square algorithm for state observation correction. This module directly extracts the issued superimposed torque command or swept torque command as a reference feedforward signal vector, and combines it with an adaptive weight vector to calculate and estimate the interference components in real time. This allows for the precise filtering of in-frequency vibration noise from the original mixed state data sequence, outputting smooth macroscopic pose data.

[0013] Finally, the solver, which includes the whole-body motion control algorithm model, uses the above smooth macroscopic pose data as the body's center of mass state variable. When it receives the path replanning instruction, it transforms the repulsion region parameters attached to the instruction into spatial inequality constraints, substitutes them into the quadratic programming optimization solution, and calculates the final motor execution instruction that maintains the body's global dynamic balance.

[0014] This invention achieves cross-modal deep fusion of visual prior assessment and active frequency domain detection at the foot, breaking through the limitations of traditional passive force perception. It can accurately identify high-risk terrain such as low friction or soft collapse before landing and during the virtual load stage. At the same time, the active detection mechanism is extended to the robot's swing phase, utilizing the pure airborne state to eliminate interference from ground contact nonlinearity, and realizing online frequency domain impedance analysis of mechanical components.

[0015] Furthermore, the constructed active excitation and feedforward filtering decoupling mechanism eliminates the contamination of high-frequency test signals on the robot's macroscopic pose estimation and attitude closed-loop control, ensuring the robot's motion stability during high-density state detection and fault diagnosis.

[0016] This invention provides a quadruped robot inspection and risk warning system based on multi-source perception. It has the following beneficial effects:

[0017] 1. This invention calculates material uncertainty and outputs ground-penetrating trigger flags through a feature quantization module. Combined with a frequency domain detection module, it issues a perturbation torque command within the virtual load time window of the target leg. It uses the synchronously acquired vibration response signal to calculate the equivalent contact stiffness of the contact surface and estimate the friction coefficient. Before the robot's body weight is substantially transferred to the ground-touching leg, it actively injects high-frequency physical excitation to extract the real mechanical parameters of the contact surface in advance. When the parameters are determined to be lower than the physical threshold, it outputs a path replanning command. This overcomes the appearance error of pure visual perception and the lag of passive force perception, and effectively avoids the risk of robot instability caused by ground collapse or slippage in complex areas.

[0018] 2. This invention uses a fault monitoring module to issue a broadband sweep torque command within a reuse time window when the target leg is in a pure air state. It calculates and obtains the real-time structural transfer function, and generates a degradation index characterizing the micro-wear state through differential integral calculation with the health benchmark function. It fully reuses the swing phase of the robot's gait, and realizes pure vibration testing of the internal mechanical transmission link in the absence of external contact resistance coupling. Based on the generated degradation index and the hierarchical degradation threshold list, it automatically triggers hardware protection commands such as speed limiting, torque reduction, or power-off braking. This achieves dynamic flaw detection and low-level active defense of the mechanical link without interrupting normal movement tasks.

[0019] 3. This invention extracts the perturbation torque command or frequency sweep torque command issued from the bottom layer by decoupling the control module as a reference feedforward signal. With the help of the built-in adaptive notch filter, it estimates in real time and filters out the corresponding co-frequency interference components from the original mixed state data sequence, and outputs smooth macroscopic pose data to the whole body motion control algorithm model. It isolates the structural crosstalk caused by the high-frequency active physical excitation of the bottom layer transmitted upward along the rigid link to the airborne inertial measurement unit from the source of signal processing. This ensures that when performing high-frequency geological exploration or transmission flaw detection, the macroscopic gait calculation and balance stabilization control loop of the top layer of the system can still obtain pure state feedback, thus ensuring the control accuracy of the whole body motion posture and the stability of dynamic balance control. Attached Figure Description

[0020] Figure 1 This is a schematic diagram of the system modules of the present invention;

[0021] Figure 2 This is a schematic diagram of the material uncertainty quantification method based on visual multimodal fusion of the present invention;

[0022] Figure 3 This is a schematic diagram of the geological exploration and dynamic risk avoidance mechanism during the ground contact period of the supporting leg according to the present invention;

[0023] Figure 4 This is a schematic diagram of the gait phase time division multiplexing logic and the frequency sweep signal generation process of the present invention;

[0024] Figure 5 This is a schematic diagram of the dynamic bidirectional decoupling and system stability control process based on feedforward phase shift of the present invention;

[0025] Figure 6 This is a schematic diagram illustrating the comparative verification of the pitch attitude stability of a quadruped robot under high-frequency exploration according to the present invention. Detailed Implementation

[0026] The technical solutions in 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.

[0027] This invention provides a quadruped robot inspection and risk warning system based on multi-source perception. The system is applied to the hardware entity of the quadruped robot, which includes an onboard main control platform, a multi-source perception sensor array, and a chassis structure containing multiple leg actuators.

[0028] The onboard main control platform is located inside the torso of the quadruped robot's hardware entity and includes an embedded main control board. The embedded main control board integrates a central processing unit and a hardware co-processing unit. It receives and processes multimodal sensor data, performs environmental feature extraction, generates active frequency domain detection commands, calculates ground impedance, and executes algorithms for whole-body motion control of the robot.

[0029] The chassis structure includes four symmetrically distributed leg actuators. Each leg actuator is physically connected to the torso via a hinge mechanism. Each leg actuator has three active degrees of freedom, equipped with a first joint servo motor, a second joint servo motor, and a third joint servo motor, used to control the leg's roll, pitch, and knee extension / retraction movements. Each joint servo motor integrates an absolute encoder, a temperature sensor, and a motor driver unit. The motor driver unit receives low-level torque control commands from the embedded main control board and transmits rotor position and temperature data back to the embedded main control board.

[0030] The multi-source sensing sensor array includes an environmental sensing component and a body state sensing component. The environmental sensing component includes depth vision cameras rigidly fixed to the front and sides of the torso. The depth vision cameras are used to acquire color environmental image data and corresponding 3D depth point cloud data in the direction the machine is moving forward.

[0031] The body state perception component includes an airborne inertial measurement unit, foot torque sensors, and broadband vibration sensors. The airborne inertial measurement unit is fixed to the physical center of the torso and is used to collect real-time data on the three-axis acceleration, three-axis angular velocity, and body attitude of the chassis in the global coordinate system.

[0032] Each leg actuator has a ground-contacting foot structure at its end. A foot torque sensor is installed between the end link of the third joint servo motor and the ground-contacting foot structure to measure the three-dimensional contact reaction force data when the foot contacts the ground surface. A broadband vibration sensor is embedded in a rigid substrate inside the ground-contacting foot structure to collect broadband vibration response signals along the normal direction of the contact surface.

[0033] The airborne main control platform establishes physical connections and communication links with the multi-source sensing sensor array and the servo motors of each joint via an airborne bus network. The depth vision camera connects to the embedded main control board via a first high-speed communication interface to transmit video streams and point cloud streams. The airborne inertial measurement unit and broadband vibration sensor are connected to the acquisition channel of the embedded main control board via a second serial communication bus. The foot torque sensor and the motor driver units of each joint servo motor are connected in series on the same controller area network communication bus, forming a closed-loop data transmission and reception network uniformly scheduled by the embedded main control board.

[0034] See attached document Figure 1 This invention provides a quadruped robot inspection and risk warning system based on multi-source perception. This system operates as a control logic unit embedded within an embedded main control board. The system includes a feature quantization module, a frequency domain detection module, a fault monitoring module, and a decoupling control module.

[0035] The feature quantization module acquires color environmental image data and 3D depth point cloud data from a depth vision camera via a data interface. This module calculates local pixel differences in the color environmental image data to extract texture variance features, extracts the proportion of high-exposure pixels as reflectivity features, and calculates the local point cloud missing rate as depth missing features in the 3D depth point cloud data. Based on preset weighting coefficients, this module linearly superimposes these three feature values ​​to generate a material uncertainty matrix covering the machine's forward direction.

[0036] Combining the global coordinates of the predicted landing point for the next cycle calculated by the robot's gait planning program, the feature quantization module extracts the value of the corresponding coordinate position from the material uncertainty matrix. When the extracted value is greater than the set confidence threshold, the module outputs a ground-penetrating trigger flag for that landing point to the frequency domain detection module and identifies that area as a high-priority detection zone.

[0037] The frequency domain detection module receives the ground-penetration trigger flag. This module internally stores a preset mapping table between visual feature types and mechanical vibration frequencies. For the dominant feature type in the material uncertainty matrix that causes the values ​​to exceed limits, this module extracts the corresponding detection frequency value from the mapping table. This module monitors the normal contact force data of the foot-end torque sensor corresponding to the target leg in real time. When the normal contact force is detected to jump from zero to the contact threshold, and within a time window during which the robot's chassis center of mass has not shifted towards the currently contacting leg, this module generates a superimposed torque command.

[0038] The superimposed torque command is configured to add a sinusoidal perturbation torque at the extracted detection frequency to the original macroscopic motion control torque. The frequency domain detection module sends the superimposed torque command to the joint servo motor of the corresponding leg, and simultaneously receives the vibration response signal collected along the normal direction by the broadband vibration sensor. This module performs a fast Fourier transform on the sent perturbation torque sequence and the received vibration response signal sequence, calculates the ratio of the two in the frequency domain, and obtains the real-time mechanical frequency response function of the contact area. This module extracts the amplitude and phase data of the frequency response function and calculates the estimated friction coefficient and contact stiffness of the contact area. When the estimated friction coefficient or contact stiffness is less than the set physical safety threshold, the module outputs a path replanning command to interrupt the current action.

[0039] The fault monitoring module receives gait phase state machine data from the robot's base layer. When it determines that a certain leg is in an airborne state and the corresponding foot torque sensor value is zero, the module sends a broadband sweep torque command with a preset frequency band to the joint servo motor of that specific leg. Simultaneously, the module acquires the structural vibration response signal collected by the broadband vibration sensor under this sweep excitation.

[0040] The fault monitoring module calculates the real-time structural transfer function of the corresponding leg mechanical link using the input sweep torque and the output response signal. This module retrieves the healthy transfer function of the leg, measured under factory baseline conditions, from the main control platform's non-volatile memory. The module calculates the integral of the square of the difference between the real-time structural transfer function and the healthy transfer function within a preset test frequency band, generating a degradation index. This module compares the degradation index with a preset list of graded degradation thresholds and outputs corresponding hardware protection commands. These hardware protection commands specifically include commands to reduce the current joint's maximum operating speed, limit the maximum output torque, or trigger a power-off brake to stop the machine.

[0041] The decoupling control module receives the high-frequency perturbation torque sequence generated by the frequency domain detection module and the fault monitoring module as a reference feedforward signal. This module simultaneously acquires raw state data containing high-frequency physical crosstalk collected by the airborne inertial measurement unit and foot torque sensors. This module is equipped with an adaptive notch filter based on the least mean square algorithm. Using the reference feedforward signal and real-time updated adaptive weighting coefficients, this module subtracts the same-frequency interference vibration component from the raw state data, outputting filtered, smoothed macroscopic pose data. This module inputs the smoothed macroscopic pose data, path replanning instructions, and hardware protection instructions into the robot's whole-body motion control algorithm model, calculating and generating the final motor execution instructions to maintain the robot's dynamic balance, and distributing them to the servo motors of each joint via the airborne bus network.

[0042] See attached document Figure 2 The method steps of the present invention are executed by the logic unit in the aforementioned embedded main control board. The method includes environmental data alignment, local visual feature extraction, uncertainty matrix generation, and ground probe marker triggering steps.

[0043] The embedded main control board acquires color image frames output from the depth vision camera and synchronized raw depth image frames. By reading the pre-calibrated camera intrinsic parameter matrix and extrinsic translation and rotation matrix, it spatially maps and aligns the 2D pixel coordinates in the color image frames with the 3D point cloud coordinates in the raw depth image frames. The aligned data is then converted into a global elevation map matrix with the robot's torso geometric center as the origin, which is used for subsequent feature calculation of local regions.

[0044] For the aligned color image frames, perform local texture feature extraction. Convert the color image frames into grayscale matrices and set their size to [value missing]. A pixel sliding window is used. The grayscale matrix is ​​traversed, and the local variance of the pixel grayscale values ​​within each sliding window is calculated. This local variance value is defined as the texture divergence feature. ,in This is the coordinate index of the center point of the current sliding window in the global elevation map matrix.

[0045] For the aligned color image frames, local reflection feature extraction is performed. The color space of the color image frames is converted from the RGB model to the HSV model, and the V channel data representing brightness is extracted. A preset brightness threshold is set. In the same way as above Within the sliding window, the V channel value is counted if it exceeds the preset brightness threshold. The number of pixels. Calculate this number as a percentage of the total number of pixels within the window, and define it as the reflective feature parameter. This parameter is used to quantify the physical appearance of specular reflection or water stain reflection in a surface area.

[0046] For the synchronized raw depth image frames, depth missing rate feature extraction is performed. Under certain material conditions of ambient light absorption or specular reflection, the active ranging beam of the depth vision camera cannot return, resulting in zero or invalid values ​​in the point cloud data. Within the local grid region corresponding to the global elevation map, the total number of data points with zero or invalid depth values ​​is counted, and the ratio of this number to the theoretically expected total number of data points in that grid is calculated, defined as the depth missing density feature. .

[0047] Based on the extracted three types of feature parameters, a material uncertainty evaluation model is constructed. For any coordinate unit in the global elevation map matrix... Its material uncertainty value Calculated based on the following linear superposition formula: ;

[0048] In the formula, This is the texture weight coefficient. This is the reflectivity weighting coefficient. This represents the weighting coefficient for deep missing values. , , All are dimensionless constants, whose values ​​are pre-determined and stored in memory during the system initialization phase based on the signal-to-noise ratio characteristics of the onboard sensors. A complete material uncertainty matrix is ​​generated by traversing the global coordinate system.

[0049] Receive the robot's full-body motion control algorithm in real time to predict the landing point coordinates for the next motion cycle. ,in This represents the x-coordinate of the estimated landing point. This represents the ordinate of the estimated landing point. The quantized value corresponding to this coordinate is located and extracted from the material uncertainty matrix. . Quantify this value With respect to the system's set confidence level safety threshold Perform numerical comparisons. When the conditions are met... At that time, it was determined that the physical surface area corresponding to the estimated landing point had material properties that could not be measured purely by visual geometry. The logic processing unit of the main control board then triggered the ground probe flag for the leg actuator located at the estimated landing point. Boolean values ​​are determined by The bit is set to 1, and this state is sent to the downlink drive control layer. This operation is completed before the leg contacts the ground, establishing a pre-judgment process to avoid the problem of passive sensing lag.

[0050] See attached document Figure 3 The method steps of the present invention are used to physically extract contact parameters of the contact surface and perform path blocking control during the transition period when a specific leg of the robot contacts the ground but has not yet borne the main load of the body.

[0051] When the ground probe trigger flag is set, At this time, the logic control unit retrieves the texture divergence feature value, reflectivity feature value, and depth missing density feature value corresponding to the coordinates of the current target landing point. The system executes comparison logic to determine the type of feature component that dominates the uncertainty calculation (i.e., the one with the largest weighted product). A feature frequency mapping data table is pre-configured in the system memory, establishing a correspondence between different visual feature types and corresponding mechanical test frequency bands. Based on the determined dominant feature component type, the system retrieves and extracts the corresponding excitation frequency parameter from the feature frequency mapping data table. .

[0052] The system monitors the foot-end torque sensor located at the end of the leg about to touch the ground in real time to obtain its Z-axis normal contact reaction force. Set the contact detection threshold. The contact determination threshold Set to a value greater than the sensor's zero-load floor noise level. When detected... When the conditions are met, the initial contact between the foot and the physical ground surface is determined. Simultaneously, the kinematic trajectory of the chassis center of mass in the three-dimensional global coordinate system is monitored within the current control cycle. A virtual load time window is defined as the continuous time period from the moment the normal contact reaction force exceeds the contact determination threshold until the moment the body center of mass undergoes a set displacement change towards the support polygon corresponding to the contacting leg. .

[0053] Within the time window for determining whether a virtual load is in effect In this state, the underlying motor driver operates under the original kinematic command torque. Based on this, a micro-amplitude high-frequency test torque signal is superimposed to generate the actual output torque. The control law formula is defined as follows:

[0054] ;

[0055] In the formula, The set perturbation amplitude reference value is such that the tangential torque component generated by this reference value is less than the critical value of the static friction force at the foot end; These are the excitation frequency parameters extracted in the preceding steps; This represents the relative time variable within the virtual load time window. The actual output torque is generated by the underlying motor driver. The force is transmitted to the foot contact surface through the rigid link of the leg, forming a physical high-frequency domain excitation force injected into the ground.

[0056] Within the same virtual load time window of injecting high-frequency excitation force Inside, a broadband vibration sensor inside the foot collects in real time the broadband vibration velocity response time series along the normal direction at the interface between the foot and the contact surface. The logic processing unit extracts the corresponding test torque time series. That is, the aforementioned superimposed components The logic processing unit simultaneously performs Fast Fourier Transform calculations on the test torque time series and the broadband vibration velocity response time series to obtain the frequency domain complex sequences. and .

[0057] Based on the frequency domain transformation results, calculate the frequency response function (equivalent mechanical impedance) of the current physical contact area. The calculation formula is:

[0058] ;

[0059] Calculate and obtain the frequency response function Subsequently, the logic processing unit extracts the resonance peak amplitude and the corresponding phase angle deviation data of the frequency response function within the test frequency band. Based on the contact mechanics mapping algorithm pre-stored in the system, the equivalent contact stiffness of the ground surface is calculated and output using the extracted resonance peak amplitude. The estimated coefficient of kinetic friction of the contact surface is calculated and output using phase angle deviation data. .

[0060] After obtaining geological parameters, the system executes threshold comparison and risk avoidance logic. The calculated equivalent contact stiffness is then used... With the preset safety stiffness threshold Numerical comparisons will be performed to estimate the coefficient of kinetic friction. With the preset safety friction threshold Perform numerical comparisons. When there is... or If any of the conditions are met, the system determines that there is a substantial physical risk of ground subsidence, collapse, instability, or slippage at the current support point.

[0061] In response to the physical risk assessment results, the system will operate within the current virtual load time window. Internally, an interrupt command to terminate the center of mass transfer is directly sent to the robot's whole-body motion control algorithm, blocking further transfer of body weight to that leg. Simultaneously, the system sends repulsion zone parameters for the current foot position coordinates to the gait planning module. Based on the received repulsion zone parameters, the gait planning module recalculates the inverse kinematics, generating a sequence of crossing or circumventing trajectories that raise the currently grounded leg joint and change the foot position coordinates. The updated trajectory sequence directly overwrites the original instruction stack and is sent to the underlying motor driver to control the joint servo motors, completing a closed-loop motion to avoid physical risks.

[0062] See attached document Figure 4 The method steps of the present invention are used to apply physical excitation commands to the internal mechanical links using a specific motion phase during the movement of a robot.

[0063] The logic control unit within the embedded main control board receives phase state machine data from each leg actuator in real time from the gait planning program. Simultaneously, the logic control unit reads the force values ​​from the corresponding foot torque sensors of each leg in real time via the onboard communication bus.

[0064] The system is configured with logic to determine the airborne state, which includes two conditions: first, phase state machine data indicates that the target leg is currently in the swing phase of leaving the ground; second, the resultant force of the three-dimensional forces output by the foot torque sensor corresponding to the target leg is less than the sensor's zero-point noise threshold for a continuous predetermined sampling period. When both conditions are met simultaneously, the system determines that the target leg has entered a pure airborne state without physical coupling constraints from the ground.

[0065] After determining that the target leg is in a pure airborne state, the system extracts the complete swing time period from when the leg lifts off the ground until it prepares to touch down for the next movement. To avoid interference from the macroscopic inertial force impact generated during the initial acceleration off the ground and the final deceleration upon landing, the logic control unit extracts a continuous time interval in the middle of the complete swing time period, defining it as a reused test time window. .

[0066] During the reuse test time window Internally, the logic control unit's internal generator calculates and generates a wideband swept-frequency torque command sequence. The broadband sweep torque command sequence is configured as a continuous sinusoidal signal whose frequency increases linearly with time. The system presets the start frequency of this sweep signal. With termination frequency The frequency band determined by the start and end frequencies covers the inherent frequency range of the base set at the factory for the target leg transmission mechanical component.

[0067] Wideband sweep frequency torque command sequence The specific generating formula is defined as follows:

[0068] ;

[0069] In the formula, This is a preset sweep frequency excitation amplitude constant, set to a value that does not change the current macroscopic swing kinematic trajectory of the swinging leg. For the relative time variable within the virtual load time window, It is a sine function.

[0070] After generating the broadband swept-frequency torque command sequence, the system adds it as an additional component to the original swing phase control torque command vector of the target leg joint servo motor. The final execution command, including the additional component, is sent to the corresponding underlying motor driver. The joint servo motor generates continuously varying torque fluctuations based on this command. Because it is in an airborne state and not affected by external resistance, the mechanical vibration wave generated by this torque fluctuation is transmitted only along a unidirectional internal mechanical physical link: motor rotor, reducer backlash, rigid connecting rod, and foot end, providing an excitation source input for the pure extraction of subsequent structural response characteristics.

[0071] The method steps of the present invention are used to quantify the degree of physical degradation of the internal mechanical structure of a robot by frequency domain analysis after acquiring the time domain response signal.

[0072] With the target's legs in the air, a broadband sweep frequency torque command sequence is injected. The same reuse test time window Internally, the system synchronously activates the high-speed sampling channel of a broadband vibration sensor located inside the foot-end structure at the end of the leg. The broadband vibration sensor collects in real-time the structural vibration velocity response time sequence transmitted along the internal mechanical linkage to the end due to the frequency sweep excitation of the joint motor, denoted as... .

[0073] During the reuse test time window After completion, the logic processing unit within the embedded main control board extracts the buffered input signal sequence. With output response signal sequence The logic processing unit simultaneously performs Fast Fourier Transform calculations on the two sets of time series, converting the time-domain data into a frequency-domain complex sequence, and obtaining the frequency-domain sequence of the swept-frequency input. With the frequency domain sequence of vibration response .in, This represents the angular frequency variable, and its value range corresponds to the frequency band formed by the starting and ending frequencies set in the preceding steps. ,in, The starting frequency, This is the termination frequency.

[0074] The logic processing unit uses the frequency domain complex number sequence obtained by transformation to calculate the real-time structural transfer function of the target leg at this moment. The formula for calculating the real-time structure transfer function is defined as follows:

[0075] ;

[0076] Since the extraction operation takes place in a pure, airborne state where the legs are not constrained by ground contact resistance, the calculation results exclude the interference of external environmental loads in the frequency domain. The position and amplitude of its resonance peak depend solely on the structural stiffness, internal damping, and clearance state of the mechanical transmission link formed by the current motor rotor, reducer, rigid connecting rod, and foot.

[0077] After completing real-time status extraction, the system accesses the onboard non-volatile memory module via the internal memory bus. The logic processing unit calls the matching health baseline function based on the hardware physical number of the current target leg. The health baseline function is the structural transfer function data measured and stored using the same set of swept-frequency excitation parameters during the robot's factory calibration phase, under the same clean, airborne conditions.

[0078] To quantify the wear and loosening of the current mechanical structure, the logic processing unit performs a real-time structural transfer function. With health benchmark function Perform difference integration. Calculate the frequency band. Integrating the sum of squared deviations of the amplitudes of the two factors generates the micro-degradation index. The formula for calculating the microdegradation index is defined as follows:

[0079] ;

[0080] In the formula, This indicates the operation of taking the amplitude of a complex function. When the reducer inside the target leg experiences increased microscopic backlash, surface fatigue spalling of the bearing balls, or initial loosening of the connecting rod fasteners, the inherent resonant frequency of the transmission link will shift, resulting in a significant increase in the absolute value of the difference within the predetermined integration frequency band, thereby causing the calculated microscopic degradation index to change. The numerical value jumps. This quantified value serves as the objective trigger input for subsequent hardware fault warnings and protection actions.

[0081] The method steps of this invention are used to automatically execute graded physical protection actions for the underlying hardware of a robot based on quantified mechanical degradation data.

[0082] The embedded main control board's non-volatile memory module is pre-configured with a degradation threshold list consisting of three discrete values. The degradation threshold list includes a first-level warning threshold. Level II warning threshold and Level 3 shutdown threshold Furthermore, the three thresholds satisfy a strict numerical increasing relationship: The logic processing unit receives the microscopic degradation index of a specific leg calculated in the preceding steps. It also performs real-time numerical comparisons with the degradation threshold list.

[0083] When the numerical comparison result meets the condition At this time, the logic processing unit changes the current hardware state machine of the leg to a normal wear state. In response to this state, the system issues a speed limit command to the underlying motor driver corresponding to the leg via the controller area network bus. This speed limit command is configured to modify the maximum allowable speed parameter in the driver's internal register, reducing it to a first preset ratio of the rated maximum speed to reduce the kinetic energy impact of the mechanical link during movement.

[0084] When the numerical comparison result meets the condition When this occurs, the logic processing unit changes the current leg's hardware state machine to a secondary risk state, indicating a significant increase in physical clearance in the transmission link or potential peeling of internal bearings. In response to this state, the system issues a torque limiting command to the motor driver corresponding to that leg. This torque limiting command is configured to modify the driver's internal maximum permissible output torque parameter, reducing it to a second preset proportion of the rated maximum torque. Simultaneously, the logic processing unit sends an asymmetric load request to the robot's gait planning program, instructing the program to modify the force polygon calculation logic of the support phase, forcibly reducing the body weight load shared by the faulty leg during ground contact.

[0085] When the numerical comparison result meets the condition When this occurs, the logic processing unit changes the current hardware state machine of the leg to a critical fault state, indicating an absolute risk of immediate physical breakage or mechanical jamming in the mechanical transmission link. In response to this state, the system triggers the highest-priority hardware protection interrupt procedure. The logic processing unit bypasses the conventional trajectory control link via the onboard bus and directly sends enable / disconnect commands to the drive layer of all joint servo motors in the leg, cutting off the stator power supply current to the motors. Simultaneously, the logic processing unit outputs a power-off closing control level to the electromagnetic brake mechanism integrated at the joint shaft end, physically locking the failed joint in its current position via mechanical friction plates, preventing the leg from freely falling due to gravity.

[0086] After executing any of the aforementioned low-level protection actions, the logic processing unit generates an intrinsic fault interrupt vector containing the physical number of the faulty leg, the specific value of the degradation index at the trigger time, and the protection status code that has been executed. This vector is pushed to the top-level scheduling task stack of the embedded main control board, and the top-level scheduling system sends the hardware maintenance request to the remote monitoring console through the external wireless communication interface, terminating the participation of that specific leg in subsequent routine inspection tasks.

[0087] See attached document Figure 5 The method steps of the present invention are used to isolate the structural crosstalk caused by the high-frequency physical excitation signal to the macroscopic kinematic state estimation and balance control link of the top layer when the robot performs high-frequency physical excitation at the bottom layer.

[0088] During the operating cycle when the frequency domain detection module or fault monitoring module sends high-frequency perturbation torque commands or frequency sweep torque commands to the underlying joint servo motors, the mechanical stress wave generated by the high-frequency torque is transmitted to the torso along the rigid linkage structure of the quadruped robot. The onboard inertial measurement unit, fixed to the torso, synchronously acquires the motion state of the robot in the global coordinate system. The onboard inertial measurement unit is set to... The original mixed-state data sequence output at time t is The original mixed-state data sequence physically superimposes the low-frequency motion components generated by the macroscopic gait of the organism and the high-frequency vibration interference components generated by the high-frequency torque transmission.

[0089] To separate high-frequency vibration disturbance components from the original mixed-state data sequence, the decoupling control module within the embedded main control board constructs an adaptive notch filter based on the least mean square algorithm. The decoupling control module directly extracts the known high-frequency perturbation torque command or swept-frequency torque command sequence sent to the lower layer from the internal data bus, defining it as a reference feedforward signal vector. .

[0090] Adaptive notch filters utilize reference feedforward signal vectors With adaptive weight vector updated in real time Calculate the estimated interference components. The filter converts the original mixed-state data sequence... Subtract the estimated interference components to output the filtered, smoothed macroscopic pose data. The formula for calculating smoothed macroscopic pose data is defined as follows:

[0091] ;

[0092] In the formula, This represents the transpose of the adaptive weight vector.

[0093] To achieve dynamic tracking of the filter weights, the system updates the adaptive weight vector periodically based on the smooth macroscopic pose data from the filter output, following the minimum mean square error criterion. The weight update law is defined as follows:

[0094] ;

[0095] In the formula, This is the adaptive weight vector for the next sampling period; This is a preset step size factor constant used to control the convergence speed of weight iteration. Its value is calibrated based on the sampling frequency of the airborne inertial measurement unit.

[0096] Smooth macroscopic pose data are obtained through the above calculations. Subsequently, the decoupled control module inputs the purified low-frequency kinematic state data into the whole-body motion control algorithm model configured on the embedded main control board. The whole-body motion control algorithm model includes a model predictive control layer based on multi-rigid-body dynamics. This model predictive control layer uses the input smooth macroscopic pose data as the current body center of mass state variable, and combines it with the preset body motion trajectory target to solve a constrained quadratic programming optimization problem.

[0097] During the solution process, if the decoupling control module synchronously receives the path replanning command triggered by the frequency domain detection module in the aforementioned embodiment, the whole-body motion control algorithm model extracts the repulsion zone coordinate parameters contained in the command and substitutes them as additional spatial inequality constraints into the quadratic programming optimization model. Through optimization, the whole-body motion control algorithm model calculates and outputs the basic support reaction force and desired kinematic torque required by each leg joint in the next control cycle. This desired kinematic torque command is then distributed to the corresponding joint servo motors via the airborne communication bus to maintain the macroscopic attitude balance of the chassis throughout the entire process of physical detection and anomaly avoidance.

[0098] Specific application examples:

[0099] A quadruped robot equipped with the system of this invention is set up to perform an autonomous inspection task in an abandoned factory. The factory floor has large areas of glare and soft silt deposits caused by ruptured underground pipes.

[0100] When the robot approaches the edge of a puddle, its onboard depth vision camera captures images of the surrounding environment and depth point clouds. The feature quantization module extracts features using the acquired color environmental image data and 3D depth point cloud data. Considering the strong specular reflection in the puddle area and the failure of the active ranging beam, the system calculates high reflectivity feature parameters. and depth missing density features According to the formula:

[0101] ;

[0102] The system calculates the estimated landing point at the edge of the puddle. Material uncertainty Due to strong reflections and missing point clouds, the calculated values ​​are... Exceeded the system's preset security confidence threshold The feature quantization module outputs the ground-penetration trigger flag in advance based on this information. .

[0103] When the robot swings its right front leg to the edge of the puddle and makes initial contact with the ground, the foot torque sensor detects the normal contact reaction force. If the contact threshold is exceeded, and the load has not yet shifted towards the right foreleg, the system determines that it has entered the virtual load time window. The frequency domain detection module injects high-frequency perturbation torque commands into the joint motor of the right front leg. Wideband vibration sensors synchronously acquire vibration response speed The system computes frequency domain complex sequences using Fast Fourier Transform. and And using the formula:

[0104] ;

[0105] Solve for the local frequency response function Based on the resonance amplitude and phase data extracted from the frequency response function, the equivalent contact stiffness of the silty ground at this location was calculated. Far below the safety stiffness threshold The frequency domain detection module immediately sent a path replanning command to the underlying layer. The whole-body motion control algorithm model extracted the repulsion zone parameters and modified the gait trajectory. The robot quickly raised its right foreleg and reselected a solid support point to cross the puddle, successfully avoiding the risk of falling due to the soft ground subsidence.

[0106] During the subsequent movement, the decoupling control module extracts the perturbation torque signal sent from the underlying layer as a feedforward, and uses an adaptive notch filter to filter out the high-frequency vibration interference components that are transmitted from the legs to the body trunk in real time. This ensures that the smooth macroscopic pose data output by the airborne inertial measurement unit remains stable, and the whole-body balance control is not negatively affected by the physical flaw detection action.

[0107] In a simulated complex geological test site (including alternating soft sand pits and hard reflective glass panels), the posture stability and hazard avoidance response of the same quadruped robot were tested when the system of the present invention was turned off (control group, relying only on conventional vision and passive touch) and when the system of the present invention was turned on (experimental group).

[0108] Ground contact detection and hazard avoidance response time: After stepping into the soft sand pit, the control group relied on a large tilt of the body before triggering the passive instability recovery strategy, with an average response delay of 350ms; while the experimental group completed the geological stiffness verification through the frequency response function within 50ms of the ground contact virtual load stage and directly blocked the transfer of the body's center of mass, shortening the hazard avoidance response time by 85.7%.

[0109] Aircraft attitude stability control (filter effect): When the experimental group performed high-frequency physical vibration flaw detection and ground exploration, the original mixed signal of the airborne inertial measurement unit was superimposed with a large number of high-frequency spikes. Without the decoupling control module of this invention, the attitude calculation would diverge. After the decoupling filter was turned on, the smooth macroscopic attitude data maintained a pure low-frequency trajectory, and the maximum pitch angle error fluctuation of the aircraft converged from ±4.5° to within ±0.8°.

[0110] See attached document Figure 6 , Figure 6 The horizontal axis represents the running time, and the vertical axis represents the body pitch angle. The interval from 1s to 3s in the figure is marked as the high-frequency perturbation physical exploration window, which corresponds to the stage in the aforementioned embodiment where the robot performs virtual load physical excitation or oscillating phase sweep frequency excitation.

[0111] As shown by the gray lines in the figure (the original mixed state data without decoupling filtering), within this window, due to the direct mechanical transmission of the underlying physical excitation, the body's inertial data experienced violent high-frequency oscillations and locally exceeded the attitude instability safety threshold represented by the black dashed line. This directly led to the collapse of the top-level motion control solution.

[0112] The bold black line representing the processing result of this invention (smooth macroscopic pose data after decoupling filtering is enabled) maintained a smooth macroscopic kinematic low-frequency trajectory consistent with the non-exploration period during the exploration window. This intuitively demonstrates that after the decoupling control module performs adaptive notch filtering using the reference feedforward signal, it can effectively remove co-frequency interference components, enabling decoupling and compatibility verification between the multi-source sensing and detection system and the overall macroscopic dynamic balance control.

Claims

1. A quadruped robot inspection and risk warning system based on multi-source perception, characterized in that, include: The feature quantization module is used to calculate the material uncertainty of the target area based on the acquired environmental data, and output the ground-penetrating trigger flag when the material uncertainty is greater than the set confidence level safety threshold. The frequency domain detection module is used to respond to the ground-penetrating trigger flag and send a superimposed torque command to the target leg. It calculates the contact parameters by combining the synchronously acquired vibration response signal. When the contact parameters are lower than the physical safety threshold, it outputs a path replanning command. The fault monitoring module is used to issue a sweep torque command when the target leg is in the air, generate a degradation index characterizing the mechanical state by combining the synchronously acquired vibration response signal, and output a hardware protection command based on the degradation index. The decoupling control module is used to filter out the same-frequency interference vibration components in the acquired original mixed state data sequence by using the superimposed torque command or the swept frequency torque command as a reference feedforward signal vector, and output smooth macroscopic pose data.

2. The quadruped robot inspection and risk warning system based on multi-source perception according to claim 1, characterized in that, The environmental data includes color environmental image data and three-dimensional depth point cloud data; The target area is the robot's estimated landing point; The feature quantization module outputs the ground exploration trigger flag when the material uncertainty corresponding to the estimated landing point is greater than the set confidence level safety threshold. The process by which the feature quantization module calculates the material uncertainty of the target region includes: Calculate the local variance of local pixels in the color environment image data to obtain texture divergence features; Extract the proportion of pixels with brightness greater than a preset brightness threshold from the color environment image data to obtain reflective feature parameters; The depth missing density feature is obtained by statistically analyzing the ratio of invalid data points within a local grid of the 3D depth point cloud data. Based on preset weighting coefficients, the texture dispersion feature, reflectivity feature parameter, and depth missing density feature are linearly superimposed to generate the material uncertainty.

3. The quadruped robot inspection and risk warning system based on multi-source perception according to claim 1, characterized in that, The frequency domain detection module contains a feature frequency mapping data table between visual feature types and mechanical vibration frequencies. When the ground-penetrating trigger flag is received, the frequency domain detection module determines the dominant feature type that causes the value to exceed the limit in the material uncertainty calculation and extracts the corresponding excitation frequency parameter from the feature frequency mapping data table. The superimposed torque command is a sinusoidal perturbation torque command with a frequency equal to the excitation frequency parameter.

4. The quadruped robot inspection and risk warning system based on multi-source perception according to claim 1, characterized in that, The frequency domain detection module issues the superimposed torque command within the time window when the target leg is under virtual load; The criteria for determining the virtual load time window are as follows: The normal contact reaction force of the foot end torque sensor corresponding to the target leg exceeds the contact judgment threshold of zero load base noise, and the center of mass of the robot chassis does not undergo a set displacement change in the direction of the supporting polygon corresponding to the target leg.

5. The quadruped robot inspection and risk warning system based on multi-source perception according to claim 4, characterized in that, The superimposed torque command includes a micro-amplitude high-frequency test torque signal, and the vibration response signal is a wideband vibration velocity response time series. The contact parameters include the equivalent contact stiffness and the estimated coefficient of dynamic friction; The specific process by which the frequency domain detection module calculates the contact parameters is as follows: A fast Fourier transform is performed synchronously on the time series corresponding to the test torque signal and the broadband vibration velocity response time series to calculate the frequency domain ratio and obtain the frequency response function of the current physical contact locality. Extract the resonance peak amplitude and the corresponding phase angle deviation data of the frequency response function within the test frequency band, calculate and output the equivalent contact stiffness using the resonance peak amplitude, and calculate and output the estimated dynamic friction coefficient using the phase angle deviation data.

6. The quadruped robot inspection and risk warning system based on multi-source perception according to claim 1, characterized in that, The levitation state is a pure levitation state; The criteria for determining the pure levitation state include: The robot's gait phase state machine data indicates that the target leg is in the swing phase stage of lifting off the ground, and the resultant force value of the three-dimensional force output by the foot end torque sensor corresponding to the target leg is less than the zero noise threshold for a continuous predetermined sampling period.

7. A quadruped robot inspection and risk warning system based on multi-source perception according to claim 6, characterized in that, The fault monitoring module generates the sweep torque command within the multiplexing test time window in the middle of the complete swing time period. The frequency sweep torque command is a broadband frequency sweep torque command, which is a continuous sine wave signal whose frequency increases linearly with time. The frequency band determined by the start frequency and end frequency of the continuous sine wave signal covers the inherent frequency range of the base of the target leg transmission mechanical component.

8. The quadruped robot inspection and risk warning system based on multi-source perception according to claim 1, characterized in that, The process by which the fault monitoring module generates the degradation index and outputs hardware protection commands is as follows: The ratio of frequency domain transformation of the time series corresponding to the frequency sweep torque command to the vibration response signal is calculated to obtain the real-time structural transfer function. The degradation index is generated by calculating the sum of squared deviations between the real-time structural transfer function and the health baseline function over the frequency band. Retrieve a list of graded degradation thresholds, including preset first-level warning thresholds, second-level warning thresholds, and third-level shutdown thresholds; When the degradation index is greater than or equal to the first-level warning threshold and less than the second-level warning threshold, an instruction to reduce the current maximum running speed of the joint is output. When the degradation index is greater than or equal to the level 2 warning threshold and less than the level 3 shutdown threshold, a command to limit the maximum output torque is output, and a request to modify the force polygon calculation logic of the support phase is sent to the gait planning program. When the degradation index is greater than or equal to the level 3 shutdown threshold, a shutdown command is output to shut off the stator power supply current of the motor and trigger the electromagnetic brake to close. The command to reduce the maximum operating speed of the current joint, limit the maximum output torque, or stop the machine are the hardware protection commands.

9. A quadruped robot inspection and risk warning system based on multi-source perception according to claim 1, characterized in that, The decoupling control module has a built-in adaptive notch filter based on the least mean square algorithm; The adaptive notch filter uses the reference feedforward signal vector and the real-time updated adaptive weight vector to calculate the estimated interference component, and subtracts the estimated interference component from the original mixed state data sequence to output the smooth macroscopic pose data.

10. A quadruped robot inspection and risk warning system based on multi-source perception according to claim 9, characterized in that, It also includes a whole-body motion control algorithm model; The decoupling control module inputs the smoothed macroscopic pose data as the body's center of mass state variable into the whole-body motion control algorithm model; When the path replanning instruction exists, the exclusion zone parameter in the path replanning instruction is extracted as an additional spatial inequality constraint and substituted into the quadratic programming optimization solution of the whole-body motion control algorithm model to calculate and output the final motor execution instruction that maintains the dynamic balance of the body.

Images