Self-stabilizing step-down retreat method, system, medium and electronic device for foot-type robot

Abstract

This invention discloses a self-stabilizing weighted retreat method, system, medium, and electronic device for a legged robot. The method includes: acquiring acceleration, angular velocity, and foot contact force data of the legged robot to generate a synchronization state data frame; calculating vibration intensity, foot contact force distribution uniformity, and the number of effective supporting feet, comparing each index with a preset threshold to determine its own stability state level; calculating and smoothing the weighting factor of each monitoring sensor channel based on the level and each index to obtain a smoothed weighting factor vector; using this vector to perform confidence weighting on the original monitoring data, constructing a multi-source evidence chain based on the obtained feature values, and performing a graded response decision in conjunction with its own stability state level. This invention effectively reduces the false alarm rate of monitoring data caused by robot instability, eliminates the monitoring vacuum window during the retreat process, avoids false disconnection due to instability, and improves the reliability and continuity of hot work monitoring.

Description

Technical Field

[0001] This invention relates to the field of robotics, and more specifically, to a self-stabilizing, weight-reducing retreat method, system, medium, and electronic device for legged robots. Background Technology

[0002] In heavy industrial hot work monitoring scenarios, legged robots are typically equipped with monitoring payloads such as infrared thermal imagers, visible light cameras, and combustible gas sensors to perform temperature monitoring, gas detection, and scene monitoring tasks along a preset patrol route.

[0003] Existing legged robots are generally equipped with inertial measurement units (IMUs) and foot force sensors for body posture estimation and motion control. When the body posture angle or acceleration exceeds a preset threshold, anti-fall adjustments are triggered (such as lowering the center of gravity, expanding the support polygon, or switching to a static stable gait) to ensure the robot's walking stability and safety in complex terrain. Meanwhile, in terms of multi-sensor data fusion, existing systems typically use Kalman filtering, Bayesian inference, or DS evidence theory to fuse multi-source sensor information. The evaluation of the quality of individual sensor data is generally based on the sensor's own health status (such as communication timeout, signal saturation, and self-test anomalies) to mark whether the data is usable.

[0004] However, the aforementioned existing technologies have three drawbacks: First, the IMU and foot force data are only used for motion control loops and no mapping relationship is established with the confidence assessment of monitoring sensor data. When the robot becomes unstable due to external mechanical impact (such as hammering vibration, hoisting and landing) or non-rigid support surface (such as grid board, oily mud surface), the infrared thermal imager will cause pixel drift and temperature measurement deviation due to micro-vibration, and the gas sensor will cause reading spikes due to airflow disturbance. However, the system still trusts all collected data with equal weight, resulting in a high false alarm rate. Second, when the robot's stationary position is no longer safe or stable, the existing solution only triggers anti-fall posture control. There is no mechanism to inform the upper-level monitoring system of the duration and range of the temporary blind spot, no path planning capability to automatically retreat to a stable stationary point, and no function to automatically restore the monitoring topology after stabilization, resulting in an uncontrollable vacuum window in monitoring. Third, the robot stability event and monitoring alarm response subsystems are independent of each other and do not communicate with each other. When the robot generates a false alarm due to external disturbances and triggers a safety cut-off (such as accidentally closing the air valve), it is impossible to trace whether the false alarm is related to the robot's unstable state, which seriously affects the decision-making credibility of the monitoring system. Summary of the Invention

[0005] The purpose of this invention is to provide a self-stabilizing, weight-reducing retreat method, system, medium, and electronic device for legged robots, which solves the technical problems in the prior art such as high false alarm rate of monitoring data caused by robot instability, monitoring vacuum window generated during retreat, and inability to distinguish between real danger and misjudgment leading to false cut-off.

[0006] The first aspect of this invention provides a self-stabilizing, weighted retreat method for a legged robot, comprising the following steps: Acquire basic data and preprocess it to obtain a synchronization state data frame. The basic data includes at least acceleration data, angular velocity data and foot contact force data. Based on the synchronous state data frame, a multidimensional stability state index is calculated. The multidimensional stability state index includes at least a vibration intensity index, a foot contact force distribution uniformity index, and an effective supporting foot number index. The multidimensional stability index is compared with a preset threshold to determine the current stability level of the legged robot. The weighting factor of each monitoring sensor channel is calculated based on the self-stability level and the multi-dimensional stability index, and the weighting factor is smoothed to obtain a smoothed weighting factor vector. The original monitoring data of each monitoring sensor channel are weighted by confidence using the smoothing weighting factor vector to obtain the weighted monitoring feature value; A multi-source evidence chain is constructed based on the monitoring feature value, and a graded response decision is executed according to the multi-source evidence chain and the self-stability level.

[0007] In this solution, the step of obtaining basic data and preprocessing it to obtain synchronization status data frames specifically includes: The acceleration data, angular velocity data, and foot contact force data are interpolated and aligned according to a unified time reference to generate the original data packet after time synchronization. A low-pass filter is applied to the acceleration and angular velocity data in the original time-synchronized data packet to remove noise. The foot contact force data is filtered and denoised using a medium-range filter. The filtered and denoised digital signal is converted into a physical unit quantity to obtain the synchronization state data frame, wherein the acceleration data and the angular velocity data come from the inertial measurement unit of the legged robot, and the foot contact force data comes from each foot contact force sensor of the legged robot.

[0008] In this scheme, the calculation of the multidimensional stable state index based on the synchronization state data frame specifically includes: Obtain the synchronization status data frame at the current moment and read the historical data cache of the sliding window, wherein the length of the sliding window is 0.5 seconds to 2.0 seconds; The composite amplitude is calculated and the peak value is taken for the acceleration sequence within the sliding window, and the composite amplitude is calculated and the peak value is taken for the angular velocity sequence; The acceleration sequence is subjected to Fourier transform to extract the energy proportion of the characteristic frequency band, and the composite amplitude peak of acceleration, the composite amplitude peak of angular velocity, and the energy proportion of the characteristic frequency band are weighted and fused to obtain the vibration intensity index; The deviation of the difference between the maximum and minimum vertical forces at each foot end from the mean is calculated to obtain the foot end contact force distribution uniformity index. The number of feet with vertical forces exceeding a preset ground contact force threshold is counted to obtain the effective support foot count index.

[0009] In this solution, comparing the multidimensional stability index with a preset threshold to determine the current stability level of the legged robot specifically includes: The self-stability level includes at least a stability level, a minor disturbance level, and a severe instability level, wherein, When the vibration intensity index is less than the first vibration threshold, the foot contact force distribution uniformity index is greater than the first uniformity threshold, and the effective supporting foot number index is greater than or equal to the first number, the self-stability state level is determined to be the stability level. When the vibration intensity index is greater than or equal to the first vibration threshold and less than the second vibration threshold, or the foot contact force distribution uniformity index is greater than the second uniformity threshold and less than or equal to the first uniformity threshold, or the effective supporting foot number index is equal to the first number, the self-stability state level is determined to be the mild disturbance level, wherein the second vibration threshold is greater than the first vibration threshold and the second uniformity threshold is less than the first uniformity threshold. When the vibration intensity index is greater than or equal to the second vibration threshold, or the foot contact force distribution uniformity index is less than or equal to the second uniformity threshold, or the effective supporting foot number index is less than or equal to the second number, the self-stability state level is determined to be the severe instability level.

[0010] In this scheme, the step of calculating the weighting factor of each monitoring sensor channel based on its own stable state level and the multidimensional stable state index, and smoothing the weighting factor to obtain a smoothed weighting factor vector, specifically includes: The monitoring sensing channel includes at least an infrared channel, a visible light channel, a combustible gas channel, and an acoustic channel, wherein... For the infrared channel, based on the relationship between the vibration intensity index and the first vibration threshold and the second vibration threshold, the infrared instantaneous weighting factor is calculated according to the infrared sensitivity coefficient; For the combustible gas channel, based on the relationship between the vibration intensity index and the first vibration threshold and the second vibration threshold, and the relationship between the foot contact force distribution uniformity index and the first uniformity threshold and the second uniformity threshold, the instantaneous gas weighting factor is calculated according to the gas sensitivity coefficient and the force unevenness influence coefficient. The infrared instantaneous weighting factor and the gas instantaneous weighting factor are smoothed by first-order low-pass filtering to obtain smoothed infrared weighting factor and smoothed gas weighting factor. The visible light weighting factor of the visible light channel and the acoustic weighting factor of the acoustic channel are set respectively, wherein the value of the visible light weighting factor at the severe instability level is smaller than the value at the mild disturbance level, and the value of the acoustic weighting factor at the severe instability level is larger than the value of the infrared weighting factor at the severe instability level.

[0011] In this scheme, the step of using the smoothing weighting factor vector to perform confidence weighting on the original monitoring data of each monitoring sensor channel to obtain weighted monitoring feature values ​​specifically includes: The monitoring characteristic values ​​include infrared-weighted characteristic values, gas-weighted characteristic values, visible light-weighted characteristic values, and acoustic-weighted characteristic values, wherein... The infrared thermal image of the infrared channel is obtained, the hot spot temperature rise value is calculated, and the hot spot temperature rise value is weighted by the infrared weighting factor in the smooth weighting factor vector to obtain the infrared weighted feature value. Obtain the gas concentration value of the combustible gas channel, and weight the gas concentration value with the gas weighting factor in the smooth weighting factor vector to obtain the gas weighted feature value; The visible light image of the visible light channel is acquired, and the person recognition result is weighted by the visible light weighting factor in the smoothing weighting factor vector to obtain the visible light weighted feature value. Acoustic signals from the acoustic channel are acquired, and the abnormal acoustic event identification results are weighted by the acoustic weighting factor in the smooth weighting factor vector to obtain acoustic weighted feature values.

[0012] A second aspect of the present invention also provides a self-stabilizing weight reduction retreat system for a legged robot, comprising a memory and a processor. The memory includes a program for a self-stabilizing weight reduction retreat method for a legged robot. When the processor executes the program for the self-stabilizing weight reduction retreat method for a legged robot, it performs the following steps: Acquire basic data and preprocess it to obtain a synchronization state data frame. The basic data includes at least acceleration data, angular velocity data and foot contact force data. Based on the synchronous state data frame, a multidimensional stability state index is calculated. The multidimensional stability state index includes at least a vibration intensity index, a foot contact force distribution uniformity index, and an effective supporting foot number index. The multidimensional stability index is compared with a preset threshold to determine the current stability level of the legged robot. The weighting factor of each monitoring sensor channel is calculated based on the self-stability level and the multi-dimensional stability index, and the weighting factor is smoothed to obtain a smoothed weighting factor vector. The original monitoring data of each monitoring sensor channel are weighted by confidence using the smoothing weighting factor vector to obtain the weighted monitoring feature value; A multi-source evidence chain is constructed based on the monitoring feature value, and a graded response decision is executed according to the multi-source evidence chain and the self-stability level.

[0013] A third aspect of the present invention provides a computer-readable storage medium comprising a program for a self-stabilizing weight reduction and backoff method for a legged robot, wherein when the program is executed by a processor, it implements the steps of the self-stabilizing weight reduction and backoff method for a legged robot as described in any of the preceding claims.

[0014] The fourth aspect of the present invention provides a computer program product, the computer program product including computer program code, which, when run on a computer, causes the computer to implement the steps of a self-stabilizing weight reduction and backoff method for a legged robot as described in any of the preceding claims.

[0015] The fifth aspect of the present invention provides an electronic device, the electronic device comprising: a processor and a memory; wherein the memory is used to store a computer program, and the processor is used to execute the computer program stored in the memory to cause the electronic device to perform the steps of a self-stabilizing weight reduction and backoff method for a legged robot as described in any of the preceding claims.

[0016] The present invention discloses a self-stabilizing weight reduction retreat method, system, medium, and electronic device for a legged robot, which has the following beneficial effects: 1. Unlike existing solutions that use IMU data solely for fall prevention control, this invention establishes for the first time a causal mapping relationship between robot instability and the reduced reliability of monitoring sensor data. Through a sensor channel-specific weighting factor function, it explains at the physical level how instability affects the measurement quality of different types of sensors and adjusts decision weights accordingly. Compared to existing solutions that indiscriminately trust all collected data, this invention can reduce the false alarm rate of hot work monitoring.

[0017] 2. Existing solutions either stubbornly resist when the robot is unstable (risk of falling) or retreat without a plan (leaving a blind spot for monitoring). This invention ensures that the robot does not lose its basic perception of the fire point during retreat by scoring the field of view coverage of the retreat target point and maintaining the minimum perception during the retreat process. After reaching the stable point, it automatically replans and returns, while accurately reporting the blind spot window to the platform.

[0018] 3. This invention incorporates the E.robotself dimension into the hierarchical response decision-making process. It explicitly stipulates that when the robot is in SSS-3 and the confidence of the perception data is significantly reduced, the L2 hard cut-off requires more stringent multi-channel cross-validation before it can be executed. This mechanism fundamentally avoids catastrophic misoperations such as accidentally closing the acetylene valve due to robot shaking, which is a safety guarantee that existing solutions do not have at all.

[0019] 4. Wheeled hot work monitoring robots cannot enter areas such as dock grating areas and steep slopes due to terrain limitations. However, this invention enables legged robots to reliably perform monitoring tasks in these complex terrains through a self-stabilizing assessment and retreat mechanism. They monitor normally when the terrain is stable, retreat safely when the terrain is unstable, and automatically return after recovery. This greatly expands the deployment range of robot hot work monitoring from flat workshops to all terrains in the dock area.

[0020] 5. By triggering the collaborative coverage of neighboring robots during retreat, the present invention maintains monitoring continuity during the retreat of a single robot, avoiding the problem in existing solutions where robot withdrawal is equivalent to entering a blind zone. Attached Figure Description

[0021] Figure 1 The diagram illustrates the steps of a self-stabilizing weight reduction and backoff method for a legged robot according to the present invention. Figure 2 A flowchart illustrating the calculation of multidimensional stability indexes for a self-stabilizing weighted backoff method for a legged robot according to the present invention is shown. Figure 3 A schematic diagram illustrating the determination of the self-stabilization and weight reduction retreat method of a legged robot according to the present invention is shown. Figure 4 This diagram illustrates the generation of monitoring feature values ​​for a self-stabilizing, weighted retreat method for a legged robot according to the present invention. Figure 5 A block diagram of a self-stabilizing weight reduction and backoff system for a legged robot according to the present invention is shown. Detailed Implementation

[0022] To better understand the above-mentioned objectives, features, and advantages of the present invention, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. It should be noted that, unless otherwise specified, the embodiments of the present invention and the features thereof can be combined with each other.

[0023] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein, and therefore the scope of protection of the invention is not limited to the specific embodiments disclosed below.

[0024] Currently, the closest existing technical solution to this invention is as follows: During hot work monitoring patrols, the legged robot senses changes in its body posture through an IMU mounted on its body. When the detected posture angle or acceleration exceeds a preset safety threshold, it triggers the robot's anti-fall posture adjustment (such as lowering the center of gravity, expanding the support polygon, switching to a static stable gait, etc.). At the same time, the legged robot uses foot force sensors to detect the ground contact state of each foot for gait phase judgment and terrain adaptation.

[0025] The key difference from this invention is that in the existing solution, the IMU and foot force sensor data are only used for the robot's motion control loop (fall prevention, gait adjustment). It does not establish a mapping relationship between its own stable state, the confidence level of the sensor data, and the quality of monitoring decisions, nor does it couple the retreat behavior in unstable states with the continuous management of monitoring tasks. Specifically, when the robot becomes unstable due to external disturbances (such as deck hammering vibrations or grid panel shaking), the existing solution only adjusts its gait / posture without making any corrections to the confidence level of monitoring sensor data such as infrared thermometry and combustible gas detection. This results in the monitoring data collected in unstable states being equally trusted, easily leading to false alarms. Furthermore, the existing solution lacks a task-oriented retreat mechanism: the robot attempts to maintain its original monitoring position in unstable states, which is neither safe (potential for falling) nor reliable (the collected data is no longer trustworthy). It does not consider a closed-loop strategy of retreating to a stable point, reducing the weight of transitional data, and replanning to return to monitoring.

[0026] The technical shortcomings of the existing solutions are summarized as follows: One technical drawback is the lack of identification and compensation for the systematic interference of the robot's own instability on the monitoring sensor data. Specifically, when legged robots perform hot work monitoring tasks in complex industrial environments such as shipyard decks, they often encounter external mechanical impacts (such as being struck by a sledgehammer in a nearby area or a steel plate being hoisted and dropped) or non-rigid support surfaces (such as grid ramps or oily / muddy surfaces). These disturbances are transmitted through the robot body to loads such as infrared thermal imagers and gas sensors, causing pixel drift and temperature measurement deviations due to micro-vibrations in the infrared thermal imager, and reading spikes in gas sensors (especially electrochemical / catalytic combustion types) due to airflow disturbances. Existing solutions do not establish a causal chain between robot instability, reduced sensor data reliability, and the corresponding adjustment of data weight in monitoring decisions. Mechanically trusting all collected data leads to a high false alarm rate and reduced availability of the monitoring system.

[0027] Therefore, one of the technical problems to be solved by this invention is how to quantitatively assess the degree of disturbance of the robot's multidimensional self-state information to the quality of various monitoring sensor data, and accordingly adaptively adjust the confidence weight of sensor data in monitoring decisions.

[0028] The second technical drawback is the lack of a safe retreat and return mechanism to ensure the continuity of monitoring tasks. Specifically, when the legged robot's stationary position becomes unsafe or unstable, existing solutions only focus on the robot's fall prevention control, neglecting the continuity of the monitoring task. If the robot is forced to leave the monitoring position due to instability, existing solutions lack a mechanism to inform the upper-level monitoring system of the duration and extent of the temporary blind spot, to automatically plan a safe retreat path to a stable point, and to automatically restore the monitoring topology after the instability is resolved. This results in uncontrollable monitoring vacuums.

[0029] Therefore, the second technical problem to be solved by this invention is how to automatically plan a safe retreat path to a rigid and stable foothold when the robot body is unstable, maintain a minimum perception capability during the retreat process, report a temporary blind spot window to the platform, and automatically replan and return to monitoring after stabilization.

[0030] The third technical drawback is the lack of a closed-loop linkage between stability, data quality, and tiered response. In existing solutions, robot stability events and monitoring alarm responses are two completely independent subsystems that do not communicate with each other. When a robot generates a false alarm due to external disturbances, triggering a safety shutdown (such as accidentally closing an air valve), it is impossible to trace whether the false alarm is related to the robot's unstable state. Consequently, it is impossible to distinguish between real dangerous events and misjudgments caused by robot instability, severely impacting the reliability of the monitoring system's decisions.

[0031] Therefore, the third technical problem that this invention aims to solve is how to incorporate the robot's own stable state as one of the evidence dimensions of a multi-source evidence chain, and integrate it into the monitoring evidence chain alongside dimensions such as spatial verification, identity verification, and danger determination, so that the graded response decision can perceive and distinguish between real danger and perception deviation caused by the robot's own instability.

[0032] Specifically, Figure 1 The diagram illustrates the steps of a self-stabilizing weight reduction and backoff method for a legged robot according to the present invention.

[0033] like Figure 1 As shown, this invention discloses a self-stabilizing weighted retreat method for legged robots, comprising the following steps: S102, acquire basic data and preprocess it to obtain a synchronization state data frame. The basic data includes at least acceleration data, angular velocity data and foot contact force data. S104, calculate multi-dimensional stability index based on the synchronous state data frame. The multi-dimensional stability index includes at least vibration intensity index, foot contact force distribution uniformity index and effective supporting foot number index. S106, compare the multidimensional stability index with a preset threshold to determine the current stability level of the legged robot. S108, calculate the weighting factor of each monitoring sensor channel based on the self-stability level and the multi-dimensional stability index, and smooth the weighting factor to obtain a smoothed weighting factor vector. S110, the original monitoring data of each monitoring sensor channel is weighted by confidence using the smoothing weighting factor vector to obtain the weighted monitoring feature value; S112, construct a multi-source evidence chain based on the monitoring feature value, and perform a graded response decision according to the multi-source evidence chain and the self-stability level.

[0034] It should be noted that, in this embodiment, the proposed backoff method is applied to a custom backoff architecture, specifically including four core functional layers: Perception layer: The robot's multi-dimensional self-state data is collected in real time through the IMU, foot force sensor array, body vibration sensor and optional ground material recognition module.

[0035] Evaluation layer: Performs fusion analysis on multi-dimensional self-state data to generate the robot's self-steady state level (SSS Level) and outputs the weighting factor vectors of each monitoring sensor channel (infrared, visible light, combustible gas, acoustic).

[0036] Decision-making level: Based on its current stable state level and monitoring task context, it decides on the hierarchical execution of the demotion strategy, the triggering conditions for retreat and replanning, and the content of blind spot reports to the monitoring platform.

[0037] Execution layer: performs confidence-weighted correction of sensor data, backoff path planning and motion control, and collaborative communication with the upper-level monitoring system.

[0038] Accordingly, the entire retreat method of this invention operates collaboratively between the end-side computing unit of the legged robot and the cloud / edge monitoring platform. Specifically, firstly, the acceleration, angular velocity, and foot contact force data of the legged robot are acquired, and after time synchronization, filtering and denoising, and physical quantity conversion, a synchronization state data frame is generated. Then, based on this data frame, multi-dimensional stability state indices are calculated, including vibration intensity indices, foot contact force distribution uniformity indices, and effective supporting foot number indices. Next, each index is compared with a preset threshold to determine the current stability state level. Then, based on the current level and each index, the weighting factors of infrared, visible light, combustible gas, and acoustic channels are calculated according to the physical characteristics of each monitoring sensor channel, and a smooth weighting factor vector is obtained through first-order low-pass smoothing. Then, the original monitoring data of each channel is weighted by confidence using this vector to obtain weighted monitoring feature values, quantifying the interference of unstable state on sensor data as confidence correction. Finally, a multi-source evidence chain is constructed based on the weighted monitoring feature values, incorporating the stability state as an independent evidence dimension, and performing graded response decisions in combination with the stability state level.

[0039] According to an embodiment of the present invention, the step of obtaining basic data and preprocessing it to obtain a synchronization state data frame specifically includes: The acceleration data, angular velocity data, and foot contact force data are interpolated and aligned according to a unified time reference to generate the original data packet after time synchronization. A low-pass filter is applied to the acceleration and angular velocity data in the original time-synchronized data packet to remove noise. The foot contact force data is filtered and denoised using a medium-range filter. The filtered and denoised digital signal is converted into a physical unit quantity to obtain the synchronization state data frame, wherein the acceleration data and the angular velocity data come from the inertial measurement unit of the legged robot, and the foot contact force data comes from each foot contact force sensor of the legged robot.

[0040] It should be noted that, in this embodiment, considering that the acceleration and angular velocity signals collected by the inertial measurement unit are easily mixed with high-frequency electronic noise and mechanical resonance interference, and that the foot contact force signal will generate pulse-like spike noise at the moment of gait switching or foot landing of the legged robot, if the raw data is not filtered and directly used for subsequent stable state evaluation, it will lead to the calculation deviation of vibration intensity index and misjudgment of foot contact force distribution uniformity index. Therefore, this embodiment uses a low-pass filter for acceleration and angular velocity data to retain the low-frequency motion components of the body and filter out high-frequency noise, and uses a median filter for foot contact force data to eliminate the pulse spikes at the moment of contact without affecting the actual force trend of the foot, thereby ensuring the accuracy and reliability of the synchronous state data frame. In addition, since the sampling frequency and clock source of each sensor are different, interpolation alignment is used to ensure that all data have a unified time base, providing a time-consistent input for the subsequent calculation of multi-dimensional indexes within the sliding window.

[0041] According to embodiments of the present invention, such as Figure 2 As shown, the calculation of the multidimensional stable state index based on the synchronization state data frame specifically includes: Obtain the synchronization status data frame at the current moment and read the historical data cache of the sliding window, wherein the length of the sliding window is 0.5 seconds to 2.0 seconds; The composite amplitude is calculated and the peak value is taken for the acceleration sequence within the sliding window, and the composite amplitude is calculated and the peak value is taken for the angular velocity sequence; The acceleration sequence is subjected to Fourier transform to extract the energy proportion of the characteristic frequency band, and the composite amplitude peak of acceleration, the composite amplitude peak of angular velocity, and the energy proportion of the characteristic frequency band are weighted and fused to obtain the vibration intensity index; The deviation of the difference between the maximum and minimum vertical forces at each foot end from the mean is calculated to obtain the foot end contact force distribution uniformity index. The number of feet with vertical forces exceeding a preset ground contact force threshold is counted to obtain the effective support foot count index.

[0042] In this embodiment, the vibration intensity index is obtained by weighted fusion of the peak value of the composite acceleration amplitude, the peak value of the composite angular velocity amplitude, and the proportion of energy in the characteristic frequency band of acceleration within a sliding window. The window length of 0.5 seconds to 2.0 seconds balances the response speed and resistance to transient interference. The foot contact force distribution uniformity index characterizes the support uniformity by the degree of deviation of the vertical force of each foot from the mean. The effective support foot number index counts the number of feet with vertical force greater than the preset ground contact force threshold. The combination of the two can detect early signs of instability such as elastic deformation of the grid plate and slippage on the oily surface in advance by abnormal foot force before the robot's posture changes significantly, thus achieving early warning.

[0043] Furthermore, in this embodiment, the temporal amplitude of vibration is captured by synthesizing the peak amplitude of acceleration and the peak amplitude of synthesizing angular velocity. At the same time, the frequency domain characteristics of vibration are captured by extracting the energy proportion of characteristic frequency bands through Fourier transform. After weighted fusion, the vibration intensity index can not only reflect the severity of vibration, but also distinguish the essential difference between external hammer impact vibration and vibration generated by the robot's active gait through the spectrum distribution, avoiding misjudging normal body swaying in active gait as unstable disturbance. The multi-dimensional index fusion evaluation system can comprehensively cover different types of unstable disturbances such as high-frequency impact, low-frequency swaying, support surface deformation, and foot slippage, providing multi-dimensional state input for the accurate calculation of subsequent weighting factors.

[0044] According to embodiments of the present invention, such as Figure 3 As shown, comparing the multidimensional stability index with a preset threshold to determine the current stability level of the legged robot specifically includes: The self-stability level includes at least a stability level, a minor disturbance level, and a severe instability level, wherein, When the vibration intensity index Less than the first vibration threshold The foot contact force distribution uniformity index Greater than the first uniformity threshold Furthermore, when the number of effective support feet is greater than or equal to the first number, the self-stability level is determined to be the stability level. When the vibration intensity index is greater than or equal to the first vibration threshold And less than the second vibration threshold Or the foot contact force distribution uniformity index Greater than the second uniformity threshold And less than or equal to the first uniformity threshold Or, when the effective support foot count index is equal to the first count, the self-stability level is determined to be the mild disturbance level, wherein the second vibration threshold is greater than the first vibration threshold, and the second uniformity threshold is less than the first uniformity threshold. When the vibration intensity index is greater than or equal to the second vibration threshold Or the foot contact force distribution uniformity index Less than or equal to the second uniformity threshold If the number of effective support feet is less than or equal to the second number, the self-stability level is determined to be the severe instability level.

[0045] In this embodiment, the first vibration threshold and the second vibration threshold divide the vibration intensity into a stable range, a mild disturbance range, and a severe instability range. The first uniformity threshold and the second uniformity threshold divide the foot force uniformity into a uniform range, a mild uneven range, and a severe uneven range. The first number is 3, and the second number is 2. The thresholds together constitute a three-level stepped judgment system. The robot is judged to be stable only when the vibration intensity, foot force uniformity, and number of effective supporting feet simultaneously meet the stability level conditions, corresponding to SS1. If any indicator falls into the mild disturbance range, a mild disturbance level judgment is triggered, corresponding to SS2. If any indicator falls into the severe instability range, a severe instability level judgment is triggered, enabling timely response to abnormal disturbances in any single dimension and avoiding missed judgments when one indicator has severely exceeded the standard while other indicators are still within the normal range. Correspondingly, severe instability corresponds to SS3.

[0046] Furthermore, in this embodiment, when the severe instability level is determined, a continuous duration timer is started. Backoff is only triggered when the severe instability state continues for more than a preset time threshold, to avoid false triggering of backoff due to instantaneous impact spikes. When the stable state level switches from the current level to an adjacent level, a hysteresis comparison rule is used. The threshold for switching from a lower level to a higher level is higher than the threshold for switching from a higher level to a lower level, preventing frequent jumps in the stable state level caused by small fluctuations in the index at the threshold boundary. The attitude change rate index in the multidimensional stable state index is extracted from the IMU attitude calculation results. When determining the severe instability level, if the attitude change rate exceeds the limit and is not caused by active gait, it is used as an auxiliary criterion to further confirm the validity of the instability state and eliminate interference from the robot's active gait adjustment in the level determination.

[0047] According to an embodiment of the present invention, the step of calculating the weighting factor of each monitoring sensor channel based on the self-stability level and the multi-dimensional stability index, and smoothing the weighting factor to obtain a smoothed weighting factor vector, specifically includes: The monitoring sensing channel includes at least an infrared channel, a visible light channel, a combustible gas channel, and an acoustic channel, wherein... For the infrared channel, based on the relationship between the vibration intensity index and the first vibration threshold and the second vibration threshold, the infrared instantaneous weighting factor is calculated according to the infrared sensitivity coefficient; For the combustible gas channel, based on the relationship between the vibration intensity index and the first vibration threshold and the second vibration threshold, and the relationship between the foot contact force distribution uniformity index and the first uniformity threshold and the second uniformity threshold, the instantaneous gas weighting factor is calculated according to the gas sensitivity coefficient and the force unevenness influence coefficient. The infrared instantaneous weighting factor and the gas instantaneous weighting factor are smoothed by first-order low-pass filtering to obtain smoothed infrared weighting factor and smoothed gas weighting factor. The visible light weighting factor of the visible light channel and the acoustic weighting factor of the acoustic channel are set respectively, wherein the value of the visible light weighting factor at the severe instability level is smaller than the value at the mild disturbance level, and the value of the acoustic weighting factor at the severe instability level is larger than the value of the infrared weighting factor at the severe instability level.

[0048] In this embodiment, the sensitivity of different monitoring sensor channels to robot instability differs fundamentally: the infrared thermal imager is most affected by vibration due to pixel drift and temperature measurement deviation caused by micro-vibrations, therefore the infrared channel uses the highest infrared sensitivity coefficient and has the deepest weighting reduction; the combustible gas sensor experiences reading spikes due to airflow disturbances and is affected by both vibration and uneven force on the feet, therefore two independent variables, vibration intensity and foot force uniformity, are introduced simultaneously, and the weighting factor is calculated based on the gas sensitivity coefficient and the force unevenness influence coefficient; the visible light channel produces slightly blurred images under mild disturbances but can still be used for scene analysis, and only undergoes significant weighting reduction under severe instability; the acoustic channel has a certain resistance to the transmission of mechanical vibrations from the robot body, and only moderately reduces its weight under severe instability. The differentiated design of the weighting factors for each channel ensures that the weighting reduction magnitude matches the actual physical disturbance level of each sensor.

[0049] Furthermore, in this embodiment, a first-order low-pass filter is applied to the infrared instantaneous weighting factor and the gas instantaneous weighting factor to prevent the weighting factor from frequently changing with the instantaneous fluctuation of vibration intensity, which would cause the monitoring decision to switch repeatedly in a short period of time. At the same time, when its own stable state level recovers to the stable level, the weighting factor does not immediately jump back to 1.0 in a step manner, but gradually increases to 1.0 according to the preset recovery rate of each channel, so as to avoid the impact of sudden changes in confidence on the monitoring decision and ensure the continuity and stability of the data during the transition period. When its own stable state level is at the mild disturbance level, the system maintains the current weighting factor level unchanged to avoid unnecessary changes in the weighting factor caused by small fluctuations in the level determination within the same level, and ensure the temporal consistency of the weighting decision.

[0050] Specifically, one of the core innovations of this invention lies in establishing a family of mapping functions from the steady state to the confidence weighting factors of each sensing channel. For each sensing channel k (k∈{IR, VIS, GAS, ACOU}), a weighting factor is defined. Typically, the weighting factor function for each channel is designed as follows: Infrared Channel (IR): Infrared thermal imagers are most sensitive to micro-vibrations because these vibrations cause instantaneous shifts in the sampling position of pixels relative to the target within the field of view, resulting in fluctuations in temperature measurements. Weighting factor: ; in, This is the disturbance sensitivity coefficient for the infrared channel (typically 1.0 to 1.5, with a larger value for the infrared channel). This is the lowest confidence level (e.g., 0.2). This indicates the constraint function.

[0051] Gas Channel (GAS): Electrochemical / catalytic combustion gas sensors are prone to reading spikes under airflow disturbances. Therefore, the influence of uneven force on the foot end on airflow disturbances needs to be additionally included in the weighting factor. ; in, This is the influence coefficient of the unevenness of force distribution.

[0052] Visible light channel (VIS): Visible light cameras are sensitive to image blurring caused by vibration. The confidence level is appropriately lowered in SS2 and significantly lowered in SS3, because severe vibration makes the image unusable for people identification and scene analysis.

[0053] Acoustic Channel (ACOU): The acoustic sensor (microphone array) has a certain resistance to interference from the transmission of mechanical vibrations of the body. However, since the vibration may be transmitted through the body structure to generate low-frequency pseudo-sounds, it also needs to be deweighted in SS3, but the overall deweighting is less than that of the infrared and gas channels.

[0054] According to embodiments of the present invention, such as Figure 4 As shown, the process of using the smoothing weighting factor vector to perform confidence-weighted averaging on the original monitoring data of each monitoring sensor channel to obtain weighted monitoring feature values ​​specifically includes: The monitoring characteristic values ​​include infrared-weighted characteristic values, gas-weighted characteristic values, visible light-weighted characteristic values, and acoustic-weighted characteristic values, wherein... The infrared thermal image of the infrared channel is obtained, the hot spot temperature rise value is calculated, and the hot spot temperature rise value is weighted by the infrared weighting factor in the smooth weighting factor vector to obtain the infrared weighted feature value. Obtain the gas concentration value of the combustible gas channel, and weight the gas concentration value with the gas weighting factor in the smooth weighting factor vector to obtain the gas weighted feature value; The visible light image of the visible light channel is acquired, and the person recognition result is weighted by the visible light weighting factor in the smoothing weighting factor vector to obtain the visible light weighted feature value. Acoustic signals from the acoustic channel are acquired, and the abnormal acoustic event identification results are weighted by the acoustic weighting factor in the smooth weighting factor vector to obtain acoustic weighted feature values.

[0055] In this embodiment, the monitoring feature values ​​obtained by weighting the raw data from each monitoring sensor channel are no longer simple mappings of the original physical quantities, but rather confidence measures that incorporate the influence factors of the robot's own stable state. Among them, the infrared weighted feature value reflects the confidence of thermal anomalies after vibration correction, the gas weighted feature value reflects the confidence of concentration exceeding limits after vibration and uneven force correction, the visible light weighted feature value reflects the confidence of personnel identification after ambiguity correction, and the acoustic weighted feature value reflects the confidence of acoustic anomalies after mechanical noise correction. This weighting operation quantifies the physical interference of the robot's unstable state on various sensors into a systematic reduction of confidence, so that the deviation caused by disturbance in the raw data is identified and corrected before entering the evidence chain construction stage, thus suppressing the propagation of false alarm signals from the source.

[0056] Furthermore, in this embodiment, the weighted monitoring feature values ​​are retained within a confidence space of 0 to 1, unifying the expression of multi-source data with different dimensions and physical meanings, and providing a consistent input for subsequent multi-source evidence fusion based on DS evidence theory. When the robot is in a stable state, each weighting factor approaches 1.0, and the weighted monitoring feature values ​​reflect the actual measurement results of the sensors. When the robot is in an unstable state, the monitoring feature values ​​corresponding to channels that are more affected by disturbances are systematically suppressed, while the monitoring feature values ​​corresponding to channels that are less affected by disturbances are only moderately reduced. This ensures that the evidence dimensions on which the graded response decision is based always reflect the actual credibility level of each channel in the current stable state, avoiding the unreasonable weakening of the data value of originally credible channels due to the uniform weighting of each channel.

[0057] It is worth mentioning that, according to an embodiment of the present invention, the step of constructing a multi-source evidence chain based on the monitoring feature value and performing a graded response decision according to the multi-source evidence chain and the self-stability state level specifically includes: The weighted monitoring feature value is input into the multi-source evidence fusion module, and the DS evidence theory is used to fuse the danger determination dimensions to obtain the danger evidence value. Acquire evidence in the dimensions of identity verification, work permit, and spatial location; The robot's own stable state level and the smooth weighting factor vector are used as evidence of the robot's own state dimension, and together with the danger evidence value, the identity verification dimension evidence, the work permit dimension evidence and the spatial location dimension evidence, the multi-source evidence chain is constructed. A tiered response decision is made based on the multi-source evidence chain.

[0058] In this embodiment, the weighted monitoring feature values ​​of the three hazard-related channels—infrared, gas, and acoustic—are used as inputs to the hazard determination dimension. The DS evidence theory is employed for fusion to obtain the hazard evidence value. During the fusion process, the basic probability assignment of each channel is directly mapped by its weighted feature value, ensuring that the credibility of each channel, which has undergone stability correction, is preserved and transmitted at the evidence level. The robot's own stable state level and smoothing weighting factor vector serve as independent robot state dimension evidence, which are incorporated into the multi-source evidence chain alongside hazard evidence value, identity verification dimension evidence, work permit dimension evidence, and spatial location dimension evidence. This allows the graded response decision-making to simultaneously perceive information on both the environmental hazard state and the robot's own credibility. When the robot is in an unstable state, the decision-making layer can use this dimension to trace the confidence level of each monitoring evidence, thereby distinguishing between real hazards and perceptual deviations caused by the robot's own instability.

[0059] Furthermore, in this embodiment, the graded response decision executes differentiated responses based on the combined state of each evidence dimension in the multi-source evidence chain: when the danger evidence value is true, the self-stability state level is not at the severe instability level, and each weighting factor is not less than the preset high confidence threshold, the response action is executed according to the standard graded response procedure, indicating that the current danger perception is credible; when the danger evidence value is true, the self-stability state level is at the severe instability level, and the infrared weighting factor or gas weighting factor is less than the preset low confidence threshold, the hard cutoff response is temporarily suspended and multi-channel cross-validation is triggered, because under this condition, the danger evidence value may be caused by sensor deviation due to robot instability; when the danger evidence value is false and the self-stability state level is at the severe instability level for more than the preset retreat trigger time, the monitoring alarm response is not executed but retreat is executed, indicating that there is no real danger at present but the robot itself is in an unsafe state.

[0060] It is worth mentioning that, according to an embodiment of the present invention, the step of performing a hierarchical response decision based on the multi-source evidence chain specifically includes: When the danger evidence value is true, the self-stability level is not the severe instability level, and each weighting factor in the smooth weighting factor vector is greater than or equal to 0.7, at least one response action among L1 alarm, L2 hard cut-off, or L3 emergency evacuation shall be executed according to the standard graded response procedure. When the danger evidence value is true, the self-stability level is the severe instability level, and the infrared weighting factor or the gas weighting factor is less than 0.5, the L2 hard cut-off is temporarily suspended, only the L1 alarm is executed and labeled with a low confidence level, and a multi-channel cross-validation request is triggered. If the danger evidence value is false and the self-stability level is the severe instability level for more than the preset backoff trigger time, no monitoring alarm response will be executed.

[0061] In this embodiment, the tiered response decision-making does not rely solely on the authenticity of hazard evidence values. Instead, it comprehensively considers the robot's stability level and the weighting factor levels of each channel to determine the response strategy: When the robot is in a stable state and the confidence levels of each channel remain high, it indicates that the current sensing data is reliable, and the standard tiered response procedure is followed to execute L1 alarm, L2 hard cutoff, or L3 emergency evacuation; when the robot is severely unstable and the infrared or gas weighting factor is below 0.5, it indicates that the hazard evidence value may originate from sensor bias caused by the robot's instability rather than a genuine hot work hazard, in which case hard cutoff is temporarily suspended and cross-validation is triggered; when there is no hazard evidence but the robot itself is severely unstable, the system focuses on retreat rather than monitoring alarms. This differentiated response mechanism enables the decision-making layer to accurately distinguish between genuine hazards and sensing biases caused by instability, preventing false cutoffs.

[0062] Furthermore, in this embodiment, after the L1 alarm is labeled with a low confidence tag, the platform can identify the confidence level of the alarm based on the tag. In subsequent accident tracing, by searching the association between the low confidence tag and the robot's own stability level record during the same period, it can also determine whether the robot was in an unstable state when the alarm occurred, thereby distinguishing between real danger and false alarms caused by instability. The multi-channel cross-verification request requires that the visible light flame feature and infrared thermal anomaly simultaneously meet preset conditions before a hard cutoff can be executed. The authenticity of the danger is confirmed by mutual verification through sensor channels with different physical principles, avoiding misjudgment caused by deviation of a single channel due to robot instability. The preset retreat trigger time is 1.0 second, which excludes brief instability caused by instantaneous disturbances and ensures that retreat is triggered only under continuous instability, avoiding the impact of frequent retreat on the continuity of monitoring.

[0063] Specifically, backoff replanning is triggered when any of the following conditions are met: If the self-stable state level SS3 persists for more than the minimum duration (e.g., 1 second), transient disturbances are excluded; If the number of effective support feet drops to 2 or less, it indicates that the robot is at risk of tipping over. Uniformity of force on the foot Reduced to the safe lower limit the following; The robot receives a forced retreat instruction from the monitoring platform (e.g., the platform determines that a large hoisting device is about to pass through the patrol area).

[0064] The selection of the retreat target point follows the principles of proximity, accessibility, sufficient rigidity, and no loss of key field of view. The specific steps are as follows: Step 1: Selection of Candidate Rigid Stable Standing Points (RSSPs). Based on the pre-built or online-updated accessibility cost map (TCM), select ground grids that meet the following conditions as candidate RSSPs: the grid is marked as a rigid zone (steel plate / concrete surface), non-grid ramp area, oily / muddy area, or trench no-go zone; the grid area is sufficient to accommodate the robot's supporting polygon (typical requirement area for quadruped robots ≥ 0.5m²); the Euclidean distance from the grid to the current robot position is ≤ the maximum retreat distance (e.g., 3.0m, which can be configured according to the scenario).

[0065] Step 2: Ranking and scoring candidate RSSPs. Perform multi-objective scoring for each candidate RSSP: Score(p) = w1 (1 - dist(p, p_cur) / D_retreat_max) + w2 view_score(p) +w3 safety_margin(p); Where dist(p, p_cur) is the Euclidean distance from the candidate point to the current position; view_score(p) is the field-of-view coverage score of the candidate point to the original monitored target area (the higher the better, to avoid completely losing the monitored field of view after retreat); safety_margin(p) is the safety margin score around the candidate point (the farther away from the danger zone the better); w1, w2, and w3 are weighting coefficients (typical values ​​w1=0.4, w2=0.35, w3=0.25).

[0066] Step 3: Select the RSSP with the highest score as the retreat target point.

[0067] 5.4.3 Retreat Path Planning The planning of retreat routes differs from that of conventional patrol routes, prioritizing safety and rapid arrival over path smoothness: The path planning algorithm adopts an improved A or D Lite algorithm based on TCM, which significantly increases the passage cost in unstable areas (grid ramps, oily surfaces) in the cost function, ensuring that the robot retreats along rigid surfaces as much as possible; The gait selection for retreat path should prioritize the use of static stable gait (i.e., at least 3 feet in contact with the ground at any time) to reduce the risk of falling during the retreat process; If there is no safe retreat path (i.e., the current area is unstable in all directions), the robot will perform the minimum risk posture of lowering its center of gravity and locking its joints in place, and send an emergency distress signal to the platform.

[0068] 5.4.4 Minimal Perception Maintenance and Regression During the Withdrawal Process During the retreat, the robot does not completely shut down its monitoring and sensing capabilities, but rather maintains a minimum level of awareness: Infrared and visible light are still collected at a low frame rate (e.g., 1Hz, 5-10Hz in normal mode) to maintain basic thermal anomaly and flame detection; The combustible gas sensor maintains continuous sampling without reducing the sampling rate; Sensor data collected during the backoff period is labeled as backoff mode, and the data, along with a weighting factor label, is uploaded to the platform.

[0069] Upon arrival at the RSSP, the robot performs the following recovery procedure: 1. Stable positioning test: Confirm that the IMU vibration and foot force indicators have returned to the SSS-1 range; 2. Gradual recovery of confidence: The weighting factor of each channel gradually recovers from its current value to 1.0 according to the recovery rate; 3. Topology replanning: Based on the new stable location, the topology replanning module is invoked to regenerate the monitoring patrol topology (patrol point sequence, PTZ constraints, dwell time). 4. Platform Synchronization: Report the time window, cause, minimum perception summary during the retreat event, and new topology to the monitoring platform. The platform updates the temporary blind spot window and coordinates multi-robot collaboration accordingly.

[0070] Figure 5 A block diagram of a self-stabilizing weight reduction and backoff system for a legged robot according to the present invention is shown.

[0071] like Figure 5 As shown, this invention discloses a self-stabilizing weight reduction retreat system for a legged robot, including a memory and a processor. The memory includes a program for a self-stabilizing weight reduction retreat method for a legged robot. When the processor executes the program for the self-stabilizing weight reduction retreat method for a legged robot, it performs the following steps: Acquire basic data and preprocess it to obtain a synchronization state data frame. The basic data includes at least acceleration data, angular velocity data and foot contact force data. Based on the synchronous state data frame, a multidimensional stability state index is calculated. The multidimensional stability state index includes at least a vibration intensity index, a foot contact force distribution uniformity index, and an effective supporting foot number index. The multidimensional stability index is compared with a preset threshold to determine the current stability level of the legged robot. The weighting factor of each monitoring sensor channel is calculated based on the self-stability level and the multi-dimensional stability index, and the weighting factor is smoothed to obtain a smoothed weighting factor vector. The original monitoring data of each monitoring sensor channel are weighted by confidence using the smoothing weighting factor vector to obtain the weighted monitoring feature value; A multi-source evidence chain is constructed based on the monitoring feature value, and a graded response decision is executed according to the multi-source evidence chain and the self-stability level.

[0072] It should be noted that when the self-stabilizing weight reduction retreat system for legged robots disclosed in this invention is applied, the specific process corresponds to the self-stabilizing weight reduction retreat method for legged robots described in the above embodiments. Since the specific implementation details of the system application are consistent with the content of the self-stabilizing weight reduction retreat method for legged robots described above, no further details will be provided in this embodiment.

[0073] The above embodiments are merely illustrative of the principles and effects of the present invention and are not intended to limit the invention. Any person skilled in the art can modify or alter the above embodiments without departing from the spirit and scope of the present invention. Therefore, all equivalent modifications or alterations made by those skilled in the art without departing from the spirit and technical concept disclosed in the present invention should still be covered by the claims of the present invention.

Claims

1. A self-stabilizing, weighted retreat method for a legged robot, characterized in that, Includes the following steps: Acquire basic data and preprocess it to obtain a synchronization state data frame. The basic data includes at least acceleration data, angular velocity data and foot contact force data. Based on the synchronous state data frame, a multidimensional stability state index is calculated. The multidimensional stability state index includes at least a vibration intensity index, a foot contact force distribution uniformity index, and an effective supporting foot number index. The multidimensional stability index is compared with a preset threshold to determine the current stability level of the legged robot. The weighting factor of each monitoring sensor channel is calculated based on the self-stability level and the multi-dimensional stability index, and the weighting factor is smoothed to obtain a smoothed weighting factor vector. The original monitoring data of each monitoring sensor channel are weighted by confidence using the smoothing weighting factor vector to obtain the weighted monitoring feature value; A multi-source evidence chain is constructed based on the monitoring feature value, and a graded response decision is executed according to the multi-source evidence chain and the self-stability level.

2. The self-stabilizing weight reduction and backoff method for a legged robot according to claim 1, characterized in that, The process of obtaining basic data and preprocessing it to obtain synchronization state data frames specifically includes: The acceleration data, angular velocity data, and foot contact force data are interpolated and aligned according to a unified time reference to generate the original data packet after time synchronization. A low-pass filter is applied to the acceleration and angular velocity data in the original time-synchronized data packet to remove noise. The foot contact force data is filtered and denoised using a medium-range filter. The filtered and denoised digital signal is converted into a physical unit quantity to obtain the synchronization state data frame, wherein the acceleration data and the angular velocity data come from the inertial measurement unit of the legged robot, and the foot contact force data comes from each foot contact force sensor of the legged robot.

3. The self-stabilizing weight reduction and backoff method for a legged robot according to claim 2, characterized in that, The calculation of the multidimensional stable state index based on the synchronization state data frame specifically includes: Obtain the synchronization status data frame at the current moment and read the historical data cache of the sliding window, wherein the length of the sliding window is 0.5 seconds to 2.0 seconds; The composite amplitude is calculated and the peak value is taken for the acceleration sequence within the sliding window, and the composite amplitude is calculated and the peak value is taken for the angular velocity sequence; The acceleration sequence is subjected to Fourier transform to extract the energy proportion of the characteristic frequency band, and the composite amplitude peak of acceleration, the composite amplitude peak of angular velocity, and the energy proportion of the characteristic frequency band are weighted and fused to obtain the vibration intensity index; The deviation of the difference between the maximum and minimum vertical forces at each foot end from the mean is calculated to obtain the foot end contact force distribution uniformity index. The number of feet with vertical forces exceeding a preset ground contact force threshold is counted to obtain the effective support foot count index.

4. The self-stabilizing weight reduction and backoff method for a legged robot according to claim 3, characterized in that, The step of comparing the multidimensional stability index with a preset threshold to determine the current stability level of the legged robot specifically includes: The self-stability level includes at least a stability level, a minor disturbance level, and a severe instability level, wherein, When the vibration intensity index is less than the first vibration threshold, the foot contact force distribution uniformity index is greater than the first uniformity threshold, and the effective supporting foot number index is greater than or equal to the first number, the self-stability state level is determined to be the stability level. When the vibration intensity index is greater than or equal to the first vibration threshold and less than the second vibration threshold, or the foot contact force distribution uniformity index is greater than the second uniformity threshold and less than or equal to the first uniformity threshold, or the effective supporting foot number index is equal to the first number, the self-stability state level is determined to be the mild disturbance level, wherein the second vibration threshold is greater than the first vibration threshold and the second uniformity threshold is less than the first uniformity threshold. When the vibration intensity index is greater than or equal to the second vibration threshold, or the foot contact force distribution uniformity index is less than or equal to the second uniformity threshold, or the effective supporting foot number index is less than or equal to the second number, the self-stability state level is determined to be the severe instability level.

5. The self-stabilizing weight reduction and backoff method for a legged robot according to claim 4, characterized in that, The calculation of the weighting factor for each monitoring sensor channel based on its own stable state level and the multidimensional stable state index, and the smoothing of the weighting factor to obtain a smoothed weighting factor vector, specifically includes: The monitoring sensing channel includes at least an infrared channel, a visible light channel, a combustible gas channel, and an acoustic channel, wherein... For the infrared channel, based on the relationship between the vibration intensity index and the first vibration threshold and the second vibration threshold, the infrared instantaneous weighting factor is calculated according to the infrared sensitivity coefficient; For the combustible gas channel, based on the relationship between the vibration intensity index and the first vibration threshold and the second vibration threshold, and the relationship between the foot contact force distribution uniformity index and the first uniformity threshold and the second uniformity threshold, the instantaneous gas weighting factor is calculated according to the gas sensitivity coefficient and the force unevenness influence coefficient. The infrared instantaneous weighting factor and the gas instantaneous weighting factor are smoothed by first-order low-pass filtering to obtain smoothed infrared weighting factor and smoothed gas weighting factor. The visible light weighting factor of the visible light channel and the acoustic weighting factor of the acoustic channel are set respectively, wherein the value of the visible light weighting factor at the severe instability level is smaller than the value at the mild disturbance level, and the value of the acoustic weighting factor at the severe instability level is larger than the value of the infrared weighting factor at the severe instability level.

6. The self-stabilizing weight reduction and backoff method for a legged robot according to claim 5, characterized in that, The process of using the smoothing weighting factor vector to perform confidence-weighted analysis on the original monitoring data of each monitoring sensor channel to obtain weighted monitoring feature values ​​specifically includes: The monitoring characteristic values ​​include infrared-weighted characteristic values, gas-weighted characteristic values, visible light-weighted characteristic values, and acoustic-weighted characteristic values, wherein... The infrared thermal image of the infrared channel is obtained, the hot spot temperature rise value is calculated, and the hot spot temperature rise value is weighted by the infrared weighting factor in the smooth weighting factor vector to obtain the infrared weighted feature value. Obtain the gas concentration value of the combustible gas channel, and weight the gas concentration value with the gas weighting factor in the smooth weighting factor vector to obtain the gas weighted feature value; The visible light image of the visible light channel is acquired, and the person recognition result is weighted by the visible light weighting factor in the smoothing weighting factor vector to obtain the visible light weighted feature value. Acoustic signals from the acoustic channel are acquired, and the abnormal acoustic event identification results are weighted by the acoustic weighting factor in the smooth weighting factor vector to obtain acoustic weighted feature values.

7. A self-stabilizing, weighted retreat system for a legged robot, characterized in that, The system includes a memory and a processor. The memory contains a program for a legged robot's self-stabilizing, weighted, and retreat method. When executed by the processor, the program performs the following steps: Acquire basic data and preprocess it to obtain a synchronization state data frame. The basic data includes at least acceleration data, angular velocity data and foot contact force data. Based on the synchronous state data frame, a multidimensional stability state index is calculated. The multidimensional stability state index includes at least a vibration intensity index, a foot contact force distribution uniformity index, and an effective supporting foot number index. The multidimensional stability index is compared with a preset threshold to determine the current stability level of the legged robot. The weighting factor of each monitoring sensor channel is calculated based on the self-stability level and the multi-dimensional stability index, and the weighting factor is smoothed to obtain a smoothed weighting factor vector. The original monitoring data of each monitoring sensor channel are weighted by confidence using the smoothing weighting factor vector to obtain the weighted monitoring feature value; A multi-source evidence chain is constructed based on the monitoring feature value, and a graded response decision is executed according to the multi-source evidence chain and the self-stability level.

8. A computer-readable storage medium, characterized in that, The computer-readable storage medium includes a program for a legged robot self-stabilizing weight reduction and retreat method. When the program is executed by a processor, it implements the steps of a legged robot self-stabilizing weight reduction and retreat method as described in any one of claims 1 to 6.

9. A computer program product, characterized in that, The computer program product includes computer program code, which, when run on a computer, causes the computer to implement the steps of a self-stabilizing weight reduction and backoff method for a legged robot as described in any one of claims 1 to 6.

10. An electronic device, characterized in that, The electronic device includes a processor and a memory; wherein the memory is used to store a computer program, and the processor is used to execute the computer program stored in the memory to cause the electronic device to perform the steps of a self-stabilizing weight reduction and backoff method for a legged robot as described in any one of claims 1 to 6.

Images