Multi-floor robot inspection system and navigation method thereof

Abstract

The invention discloses a multi-floor robot inspection high-precision positioning system, and the system is characterized in that the system comprises a wheeled robot body which carries a plurality of sensors and is used for executing an inspection task; the multi-sensor positioning system is used for realizing accurate positioning and navigation; the environment sensing system is used for environment sensing and obstacle avoidance; the elevator interaction module is used for realizing autonomous switching among floors; the multi-sensor positioning system comprises a UWB positioning module used for providing a global positioning reference; the visual positioning module is used for providing local accurate positioning; and the inertial navigation module is used for providing continuous attitude and motion information. The technical scheme provided by the invention has remarkable advantages in the aspects of technical innovation, application value, economic benefit, industry influence, social benefit, market competition and the like, and huge comprehensive benefit is created. The technical problem in the field of multi-floor robot inspection is solved, industrial technical progress is promoted, and considerable economic and social benefits are created.

Description

Technical Field

[0001] The present invention relates to the technical field of intelligent robots, and more specifically, to a multi-floor robot inspection system and a navigation method thereof. Background Art

[0002] With the development of industrial automation and intelligence, inspection robots are increasingly used in industrial sites. With the increase in demand for intelligent inspection, robots need to not only inspect on a single plane, but also transfer and inspect between multiple floors. However, their application in multi-floor scenarios still faces many challenges.

[0003] At present, the inspection robot systems on the market mainly adopt the following technical solutions: single-floor mapping and navigation technology, single-floor mapping based on laser SLAM, path planning based on grid map, Monte Carlo positioning algorithm, conventional floor switching method, manual assisted floor switching, simple marker point recognition, and fixed path navigation; however, the above technologies still have the following shortcomings:

[0004] 1) Limitations of the positioning system: First, it is easy to lose positioning during the floor switching process. The existing robot positioning system mainly relies on a single laser SLAM technology. During the floor switching process, the positioning system is prone to failure due to the rapid changes in the environment and the limitations of the sensor's field of view. Especially when entering the stairs or elevators, due to the narrow space and single features, the environmental information collected by the laser sensor is sharply reduced, resulting in the SLAM algorithm being unable to accurately estimate the robot's position. Secondly, there is a lack of association between maps of different floors. The traditional system independently builds maps and locates each floor, and no effective spatial association relationship is established between floors. In addition, the positioning accuracy is significantly reduced in scenes such as elevators and stairs. The metal walls around the elevator cause laser reflection interference, and the inclined surface of the stairs causes the horizontal laser scanning effect to deteriorate. The positioning accuracy in these scenes often drops from the centimeter level of the plane scene to the meter level, which seriously affects the navigation performance. Finally, it is impossible to realize automatic identification of the floor, and it still relies on manually preset floor information. Real-time tracking cannot be achieved during the floor switching process, and the position jump causes the task to be interrupted, requiring manual confirmation of the floor information.

[0005] 2) Map management issues: First, multi-floor maps are stored independently, making management complex. Second, there is a lack of topological association between floors, a lack of description of the connection relationship between floors, an inability to express the properties of vertical channels, and a lack of corresponding points between floor plans, making it impossible to generate optimal cross-floor paths. Third, map updates and maintenance are difficult, requiring all tasks to be stopped. The update process is long and requires recalibration and verification. Incremental updates are not possible, and small changes require rebuilding the entire map.

[0006] 3) Navigation planning defects: It is impossible to achieve global path planning across floors. The stair navigation strategy is simple, the success rate is low, it lacks dynamic obstacle avoidance capabilities, and it cannot handle path replanning in emergency situations.

[0007] 4) Low efficiency of task execution: Cross-floor tasks require manual intervention, and it is impossible to achieve optimal scheduling of multi-floor tasks. In addition, the interruption mechanism during task execution is imperfect and there is a lack of effective task recovery mechanism.

[0008] Based on the above-mentioned defects of the prior art, the inventor of the present application proposed a high-precision robot inspection system and inspection method suitable for multi-floor scenarios. Summary of the invention

[0009] This application aims to address the technical problems that existing robot inspection systems have low floor recognition accuracy and unstable switching process in multi-floor scenarios, and traditional single sensor positioning solutions are prone to failure in complex environments, especially during floor switching, where positioning accuracy and reliability are difficult to guarantee. By adopting multi-source data fusion technology, high-precision positioning is achieved by fusing data from multiple sensors such as lidar, vision, IMU, and barometer.

[0010] The present application provides a multi-floor robot inspection high-precision positioning system, characterized by comprising: a wheeled robot body equipped with a variety of sensors for performing inspection tasks;

[0011] A multi-sensor positioning system is used to achieve precise positioning and navigation; an environmental perception system is used for environmental perception and obstacle avoidance; an elevator interaction module is used to achieve autonomous switching between floors; the multi-sensor positioning system includes: a UWB positioning module for providing a global positioning reference; a visual positioning module for providing local precise positioning; and an inertial navigation module for providing continuous posture and motion information.

[0012] The present invention also provides a multi-floor robot inspection high-precision positioning method, which is characterized by comprising the following steps:

[0013] S1. Build a multi-floor environment map, use laser SLAM technology to establish a basic map, integrate visual information to build a semantic map, deploy a UWB base station network on each floor, and calibrate all sensor parameters at the same time;

[0014] S2. Develop multi-floor inspection routes and add navigation landmarks at the elevator entrances and exits on each floor and key inspection points;

[0015] S3, the positioning system performs real-time positioning through multi-sensor data fusion, establishes a hierarchical positioning framework from global to local, and realizes precise navigation of the robot;

[0016] S4, the environmental perception system and obstacle avoidance module work synchronously to perform dynamic path planning and obstacle avoidance control;

[0017] S5. When the robot needs to switch floors, it switches floors through the elevator interaction module.

[0018] Preferably, in step S1, a multi-floor environment map is constructed using a hierarchical structure, the base layer uses laser SLAM technology to build a two-dimensional grid map, the middle layer fuses depth camera information to build a three-dimensional feature map, and the top layer combines semantic information to build a navigation map.

[0019] Preferably, the calibration of the sensor parameters includes external parameter calibration of the lidar and depth camera, position calibration of the UWB base station network, and time synchronization calibration of the IMU and vision system. The calibration data is stored in a system configuration file for subsequent multi-sensor fusion positioning.

[0020] Preferably, the data processing of the laser radar comprises the following steps:

[0021] First, the lidar data is adaptively median filtered to remove outliers;

[0022] Then, invalid data is filtered out by a set distance threshold, and dynamic obstacles are identified and filtered based on a density clustering algorithm;

[0023] Finally, a laser data confidence assessment mechanism is established to ensure data quality. During the processing, the filtering parameters are adaptively adjusted according to the complexity of the environment to ensure the real-time and effectiveness of data processing.

[0024] Preferably, in step S3, the multi-sensor data fusion adopts an improved Kalman filter algorithm, a dynamic weight adjustment mechanism and data credibility evaluation, and the improved Kalman filter algorithm includes:

[0025] 1) State vector:

[0026]

[0027] Where: x k is the position, υ k is velocity, θ k is attitude, b k is the sensor bias.

[0028] 2) Nonlinear measurement equation:

[0029] z k =h(x k )+v k

[0030] Where: h(·) is the nonlinear measurement function, v k To measure the noise, assume that R k is the noise covariance matrix.

[0031] 3) Prediction Model:

[0032] 1. Status prediction

[0033] x k|k-1 =F k x k-1 +B k u k

[0034] 2. Covariance prediction

[0035]

[0036] Among them: F k is the state transition matrix, B k is the control input matrix, u k is the control vector, P k-1 is the state covariance matrix at time k-1, Q k is the process noise covariance matrix (process noise covariancematrix).

[0037] Preferably, in step 3, the hierarchical positioning framework from global to local adopts a multi-layer error compensation method, a smooth estimation algorithm and a feedback correction mechanism;

[0038] The multi-layer error compensation method includes a systematic error compensation method and a layered compensation for random errors;

[0039] The system error compensation method compensates for the sensor installation error through an accurate calibration method, establishes a time delay compensation model based on a timestamp, and realizes dynamic compensation of temperature drift; the steps include:

[0040] 1) Time synchronization calibration

[0041] Establish a timestamp alignment mechanism to map the data of each sensor to a unified time axis. Use the sliding window method to detect the time drift between sensors;

[0042] 2) Dynamic compensation model

[0043] Tcorrected =T measured +△T

[0044] Among them, ΔT is the compensation time delay, which is estimated by the dynamic adjustment algorithm; by analyzing the impact of temperature on sensor clock drift, the temperature compensation term is introduced:

[0045] △T=α·△T base +β·f(T) where f(T) is the temperature drift function and α, β are weights;

[0046] The hierarchical compensation of random errors adopts the optimal estimation strategy by analyzing the noise characteristics and error propagation law to effectively suppress the influence of random errors;

[0047] The optimal estimation strategy includes the following steps:

[0048] 1) Noise modeling:

[0049] Establish a probability distribution model for the noise characteristics of each sensor; introduce a dynamic weight adjustment mechanism to dynamically adjust the sensor weight according to the data credibility:

[0050] Among them, σ i is the noise variance of the i-th sensor;

[0051] 2) Optimal fusion strategy, using an extended framework based on Kalman filtering:

[0052]

[0053] Among them, K k is the Kalman gain, combined with weight adjustment;

[0054] The smoothing estimation algorithm comprises:

[0055] 1) Based on the improved RTS algorithm, continuity constraints are introduced in the backward smoothing process:

[0056]

[0057] Among them, the continuity of velocity and acceleration is considered;

[0058] 2) Smooth posture, using quaternion interpolation method to avoid gimbal lock problem:

[0059] q smooth =Slerp(q k ,q k+1 , t);

[0060] Among them, Slerp is spherical linear interpolation, t is the interpolation factor;

[0061] The feedback correction mechanism adopts a closed-loop feedback strategy to achieve error back propagation and adaptive adjustment of parameters by evaluating the quality of the fusion result; the closed-loop feedback strategy includes:

[0062] 1) Feedback framework: The feedback gain is calculated by evaluating the quality of the fusion result:

[0063] G feedback =f(E error ), where f(E error ) is the error evaluation function;

[0064] 2) Parameter adaptive adjustment: Use feedback gain to adjust model parameters, dynamically adjust sensor weights, and optimize data fusion effects: Q adjusted =Q+G feedback ·ΔQ.

[0065] Preferably, in step S5, the elevator interaction module performs floor switching including S5.1 feature fusion identification step, S5.2 time series correlation analysis step and S5.3 historical information optimization step, wherein the S5.1 feature fusion identification step includes:

[0066] S5.1.1 Floor feature extraction, the floor features include air pressure height feature extraction, environmental structure feature extraction and visual identification feature extraction;

[0067] S5.2.2 Feature fusion recognition, using a feature fusion method based on confidence weighting, effectively fuses the pressure altitude features, environmental structure features and visual identification features described in step 5.1, dynamically adjusts the weights according to the reliability of different features, and establishes an adaptive feature importance evaluation mechanism; the feature fusion method based on confidence weighting includes:

[0068] 1) Assign confidence C to each feature 特征 , dynamically calculated based on the current environment: where σ 特征 is the noise variance of the feature, Δ o is the magnitude of change;

[0069] 2) Weighted fusion, weighting the confidence of comprehensive features and calculating the final fusion value Fi represents the observed value of a single feature;

[0070] 3) Dynamic weight adjustment, real-time adjustment of weights based on the environment and feature quality

[0071] The step 5.2 of time series correlation analysis is to establish a time series state transfer model, analyze the characteristic change law in the floor switching process, use an improved hidden Markov model to describe the floor switching process, and combine the particle filter algorithm to realize probabilistic floor state estimation; the steps include:

[0072] Step 5.2.1 State Modeling

[0073] Define the state space S = {s1, s2, ..., s N}, each state corresponds to a floor number;

[0074] Define observation space O = {o1, o2, ..., o M}, including air pressure, structure and visual characteristics;

[0075] Observation value O t =[o 气压 , o 结构 , o 视觉 ];

[0076] Step S5.2.2 dynamically adjusts the value of A in conjunction with the state transfer matrix, based on the trend and characteristic change rate of the historical observation sequence:

[0077] Where Δo represents the magnitude of feature change, and λ is the weight factor;

[0078] Step S5.2.3 Multi-feature joint observation probability B:

[0079] B j (O t )=P(O t |s t =s j )=w 气压 ·B 气压 +w 结构 ·B 结构 +w 视觉 ·B 视觉 :Weight w 气压 ,w 结构 ,w 视觉 Adjust based on real-time confidence;

[0080] The particle filter algorithm steps described in step S5.2.4 include:

[0081] 1) Particle filter initialization: particle set in is the state represented by the particle, is the weight; initial particle distribution:

[0082] 2) State prediction: Use the HMM state transition model to predict particle distribution:

[0083]

[0084] 3) Weight update: Update the particle weight according to the current observation value Ot:

[0085] 4) Particle resampling: Avoid particle degradation through resampling:

[0086] 5) State estimation: The final state estimation is the weighted average of the particles:

[0087] S5.3 Historical information optimization step: Establish a historical information database for floor identification, optimize the current identification strategy by analyzing historical identification results, and use online learning methods to continuously update and improve the identification model.

[0088] Preferably, step S5 also includes abnormality detection classification, sensor abnormality detection and positioning navigation abnormality detection;

[0089] The positioning and navigation anomaly detection includes positioning accuracy anomaly, navigation deviation anomaly and path planning anomaly. An uncertainty evaluation model for pose estimation is established to monitor the reliability of the positioning system in real time. An improved RANSAC algorithm is used for outlier detection. Combined with historical trajectory analysis, abnormal behavior identification during navigation is achieved. A positioning verification mechanism based on map matching is established to ensure the continuous and reliable operation of the navigation system.

[0090] The uncertainty assessment model real-time monitoring positioning system comprises the following steps:

[0091] 1) Input data: input multi-sensor data, current pose estimation result data and historical trajectory data;

[0092] 2) Uncertainty quantification:

[0093] Gaussian distribution modeling: Assuming that the positioning error conforms to the multivariate normal distribution, construct the covariance matrix ∑:

[0094]

[0095] in Respectively represent the uncertainty in each direction, and represent the covariance;

[0096] Bayesian update: Update the uncertainty distribution of the pose based on the observed data:

[0097] P(x|z)∝P(z|x)P(x)

[0098] Where P(x|z) is the posterior probability, P(z|x) is the observed probability, and P(x) is the prior probability;

[0099] 3) Uncertainty measurement indicators:

[0100] Position reliability: Based on the determinant calculation of the covariance matrix:

[0101] C = det(∑), the larger the C, the higher the uncertainty;

[0102] Ellipsoid radius: Use principal component analysis to calculate the major and minor axes of the positioning error ellipsoid and evaluate the positioning range;

[0103] 4) Anomaly Detection: Define the Uncertainty Threshold C threshold , if C>C threshold It is judged as positioning abnormality.

[0104] Preferably, the improved RANSAC algorithm for outlier detection includes:

[0105] 1) Adaptive iteration number:

[0106] Dynamically adjust the number of iterations N: Where p is the success probability of the model, w is the inlier ratio, k is the number of samples, and the inlier ratio w is estimated online to reduce the computational complexity;

[0107] 2) Weighted inlier scoring: assign weight w to inliers i , the weights are related to the residuals:

[0108] where r i is the residual, σ is the noise variance;

[0109] 3) Multi-model verification: Calculate the confidence of different hypothesis models and select the best model: where σ m is the standard deviation of the model residuals;

[0110] 4) Combined with particle filtering:

[0111] A particle filter is used to update the inlier distribution after each iteration to improve the robustness in dynamic scenes.

[0112] The technical solution provided by this application has significant advantages in terms of technological innovation, application value, economic benefits, industry impact, social benefits and market competition, creating huge comprehensive benefits. The present invention not only solves the technical problems in the field of multi-floor robot inspection, but also promotes the technological progress of the industry and creates considerable economic and social benefits. Specifically, it includes:

[0113] 1. Technological innovation effect:

[0114] 1) Significantly improved positioning accuracy: Using a multi-source data fusion algorithm, positioning accuracy is increased from the original ±10cm to ±3cm; the floor recognition accuracy reaches 99.9%, an increase of 40% over the traditional solution; the repositioning time is shortened from an average of 15 seconds to 3 seconds, an increase of 80%; positioning stability in dynamic environments is improved by 60%.

[0115] 2) Significantly enhanced navigation capabilities: fully automatic floor switching is achieved without manual intervention; the success rate of stair navigation is increased from 85% to 98%; the dynamic obstacle avoidance reaction time is shortened by 50%; the path planning efficiency is increased by 35%, and the average time is reduced by 2 seconds.

[0116] 3) Improved system stability: The system's continuous working time is extended from 8 hours to 24 hours; the task interruption rate is reduced from 5% to 0.1%; the abnormal self-recovery capability is improved by 75%; and the battery utilization efficiency is improved by 30%.

[0117] 4) Enhanced scalability: Supports simultaneous management of building maps with more than 100 floors; can be expanded to access multiple types of sensors; is compatible with robot platforms of different brands; and supports flexible configuration of inspection strategies.

[0118] 2. Application value advantages:

[0119] 1) Reduced operation and maintenance costs: reduced the input of manual inspection personnel by 70%, equipment maintenance costs by 45%, training costs by 60%, and energy consumption by 25%;

[0120] 2) Improved work efficiency: inspection efficiency increased by 200%, data collection accuracy increased by 50%, problem discovery timeliness increased by 80%, and equipment utilization increased by 40%.

[0121] 3) Enhanced safety: Eliminates the risk of manual operations in high-risk environments, shortens emergency response time to abnormal situations by 75%, increases the accuracy of safety warnings by 65%, and improves accident prevention capabilities by 80%.

[0122] 3. Economic Benefits:

[0123] 1) Direct economic benefits: Annual labor cost savings of approximately RMB 1 million, equipment maintenance costs reduced by approximately RMB 400,000, energy consumption reduced by approximately RMB 150,000, and investment payback period shortened by 50%.

[0124] 2) Indirect economic benefits: Discover equipment hidden dangers in advance, reduce failure losses, extend equipment service life by 20%, improve equipment availability by 15%, and reduce unplanned downtime by 80%.

[0125] 4. Industry Impact:

[0126] 1) Technology leadership: It fills the technical gap in the field of multi-floor inspection, promotes the formulation of industry standards, drives the progress of related technologies, and promotes industry upgrades.

[0127] 2) Application and promotion value: It is suitable for various multi-story building inspection scenarios, solves common technical problems in the industry, has broad market application prospects, and drives the development of the industrial chain.

[0128] V. Social Benefits:

[0129] 1) Safe production: Reduce manual operations in dangerous environments, improve the inherent safety level, reduce the incidence of safety accidents, and ensure the safety of operators.

[0130] 2) Energy saving and environmental protection: reduce energy consumption, reduce environmental pollution, improve resource utilization efficiency, and promote green development.

[0131] 3) Employment upgrading: Promote the improvement of industrial workers’ skills, create high-tech jobs, improve the working environment, and enhance the value of work. BRIEF DESCRIPTION OF THE DRAWINGS

[0132] Figure 1 This is an architecture diagram of a multi-floor robot inspection system according to an embodiment of the present invention.

[0133] Figure 2 The navigation workflow diagram of the embodiment of the present invention.

[0134] Figure 3 This is a system module structure diagram of an embodiment of the present invention. DETAILED DESCRIPTION

[0135] The present invention proposes a multi-floor robot inspection system, such as Figure 1 As shown, it includes a wheeled robot body, a multi-sensor positioning system, an environmental perception system, and an elevator interaction module. The wheeled robot body is equipped with a laser radar, a depth camera, and an ultrasonic sensor array; the multi-sensor positioning system includes a UWB positioning module, a visual positioning module, and an inertial navigation module; the elevator interaction module includes at least an elevator state recognition unit, a call control unit, and an elevator internal positioning unit.

[0136] Working principle of this system: First, the system completes environmental mapping and sensor calibration during the deployment phase. Multi-floor maps are stored in a hierarchical structure for quick access and update. The deployment of the UWB base station network needs to consider signal coverage and accuracy requirements. In actual operation, the system achieves reliable positioning through multi-sensor fusion, and dynamic path planning ensures navigation safety. When it is necessary to switch floors, the elevator interaction module is responsible for autonomous calling and taking the elevator. During the execution of the inspection task, the system will record environmental changes in real time and update the map information as needed.

[0137] In order to better understand the above technical solution, the above technical solution will be described in detail below in conjunction with the accompanying drawings and specific implementation methods.

[0138] like Figure 2-Figure 3 As shown: This embodiment provides a navigation method for a multi-floor robot inspection system, comprising the following steps:

[0139] S1. Build a multi-floor environment map, use laser SLAM technology to establish a basic map, integrate visual information to build a semantic map, deploy a UWB base station network on each floor, and calibrate all sensor parameters at the same time;

[0140] S2. Develop multi-floor inspection routes and add navigation landmarks at the elevator entrances and exits on each floor and key inspection points;

[0141] S3, the positioning system performs real-time positioning through multi-sensor data fusion, establishes a hierarchical positioning framework from global to local, and realizes precise navigation of the robot;

[0142] S4, the environmental perception system and obstacle avoidance module work synchronously to perform dynamic path planning and obstacle avoidance control;

[0143] S5. When the robot needs to switch floors, it switches floors through the elevator interaction module.

[0144] Preferably, in step S1, since the traditional single sensor positioning solution is prone to failure in complex environments, especially in the process of floor switching, the positioning accuracy and reliability are difficult to guarantee. Therefore, this embodiment adopts multi-source data fusion technology to achieve high-precision positioning by fusing multiple sensor data such as laser radar, vision, IMU, barometer, etc.

[0145] More specifically, the construction of the multi-floor environment map in this embodiment adopts a hierarchical structure. The base layer uses laser SLAM technology to build a two-dimensional grid map, the middle layer integrates depth camera information to build a three-dimensional feature map, and the top layer combines semantic information to build a navigation map. The deployment of UWB base stations must ensure that each area is covered by at least 3 base stations.

[0146] The sensor calibration process mainly includes the external parameter calibration of the lidar and depth camera, the position calibration of the UWB base station network, and the time synchronization calibration of the IMU and vision system. The calibration data is stored in the system configuration file for subsequent multi-sensor fusion positioning.

[0147] LiDAR data processing: This embodiment first performs adaptive median filtering on the LiDAR data to remove outliers, then filters invalid data by setting a reasonable distance threshold, identifies and filters dynamic obstacles based on a density clustering algorithm, and finally establishes a laser data confidence assessment mechanism to ensure data quality. During the processing, the filter parameters are adaptively adjusted according to the complexity of the environment to ensure the real-time and effectiveness of data processing.

[0148] Furthermore, in step S3, the multi-sensor fusion positioning adopts an improved Kalman filter algorithm. The algorithm dynamically adjusts the weight of each sensor data to improve positioning accuracy and reliability. In areas with poor signals, the system will automatically switch to a positioning method based on visual features.

[0149] The state vector designed in this embodiment includes state quantities such as position, speed, and attitude, while taking into account the influence of sensor bias and drift. By establishing a nonlinear measurement equation and considering the correlation of multi-sensor measurement noise, the measurement model is optimized. Kinematic constraints and environmental information constraints are introduced into the prediction model to improve the accuracy of state prediction. The algorithm can adaptively adjust process noise and measurement covariance to achieve optimal state estimation. The following is a mathematical formula expression of the relevant content:

[0150] State vector:

[0151]

[0152] Where: x k is the position, υ k is velocity, θ k is attitude, b k is the sensor bias.

[0153] Nonlinear measurement equation:

[0154] z k =h(x k )+v k

[0155] Where: h(·) is the nonlinear measurement function, v k To measure the noise, assume that R k is the noise covariance matrix.

[0156] Prediction Model:

[0157] 1. Status prediction

[0158] x k|k-1 =F k x k-1 +Bk u k

[0159] 2. Covariance prediction

[0160]

[0161] Among them: F k is the state transition matrix, B k is the control input matrix, u k is the control vector, P k-1 is the state covariance matrix at time k-1, Q k is the process noise covariance matrix (process noise covariancematrix).

[0162] Preferably, in step 3, the hierarchical positioning framework from global to local adopts a multi-layer error compensation method, a smooth estimation algorithm and a feedback correction mechanism;

[0163] The multi-layer error compensation method includes a systematic error compensation method and a layered compensation for random errors;

[0164] The system error compensation method compensates for the sensor installation error through an accurate calibration method, establishes a time delay compensation model based on a timestamp, and realizes dynamic compensation of temperature drift; the steps include:

[0165] 1) Time synchronization calibration

[0166] Establish a timestamp alignment mechanism to map the data of each sensor to a unified time axis. Use the sliding window method to detect the time drift between sensors;

[0167] 2) Dynamic compensation model

[0168] T corrected =T measured +△T

[0169] Among them, ΔT is the compensation time delay, which is estimated by the dynamic adjustment algorithm; by analyzing the impact of temperature on sensor clock drift, the temperature compensation term is introduced:

[0170] △T=α·△T base +β·f(T) where f(T) is the temperature drift function and α, β are weights;

[0171] The hierarchical compensation of random errors adopts the optimal estimation strategy by analyzing the noise characteristics and error propagation law to effectively suppress the influence of random errors;

[0172] When multi-sensor data is fused, this embodiment designs an optimal estimation strategy to deal with data noise and sensor inconsistency. The optimal estimation strategy includes the following steps:

[0173] 1) Noise modeling:

[0174] Establish a probability distribution model of the noise characteristics of each sensor, such as Gaussian distribution; introduce a dynamic weight adjustment mechanism to dynamically adjust the sensor weight according to the data credibility:

[0175] Among them, σ i is the noise variance of the i-th sensor;

[0176] 2) Optimal fusion strategy, using an extended framework based on Kalman filtering:

[0177]

[0178] Among them, K k is the Kalman gain, combined with weight adjustment;

[0179] The trajectory and attitude estimation of the robot in motion may have jitters, and the trajectory smoothness needs to be optimized. This embodiment uses an improved smoothing estimation algorithm to consider the motion continuity constraint in the position estimation, optimize the trajectory smoothness, and effectively eliminate the jump phenomenon of the estimation result. In terms of attitude estimation, a quaternion interpolation smoothing method is used, combined with angular velocity constraints to ensure the continuity and accuracy of attitude estimation while avoiding universal joint lock. The improved smoothing estimation algorithm of this embodiment includes:

[0180] 1) Based on the improved RTS algorithm, continuity constraints are introduced in the backward smoothing process:

[0181]

[0182] Among them, the continuity of velocity and acceleration is considered;

[0183] 2) Smooth posture, using quaternion interpolation method to avoid gimbal lock problem:

[0184] q smooth =Slerp(q k ,q k+1 , t);

[0185] Among them, Slerp is spherical linear interpolation, t is the interpolation factor;

[0186] The feedback correction mechanism adopts a closed-loop feedback strategy to dynamically adjust sensor parameters and system models using data fusion results. By evaluating the quality of the fusion results, error back propagation and parameter adaptive adjustment are achieved; the closed-loop feedback strategy includes:

[0187] 1) Feedback framework: The feedback gain is calculated by evaluating the quality of the fusion result:

[0188] G feedback =f(E error ), where f(E error ) is the error evaluation function;

[0189] 2) Parameter adaptive adjustment: Use feedback gain to adjust model parameters, dynamically adjust sensor weights, and optimize data fusion effects: Q adjusted =Q+G feedback ·ΔQ.

[0190] Preferably, in step S5, the elevator interaction module performs floor switching including S5.1 feature fusion identification step, S5.2 time series correlation analysis step and S5.3 historical information optimization step, wherein the S5.1 feature fusion identification step includes:

[0191] S5.1.1 Floor feature extraction, the floor features include air pressure height feature extraction, environmental structure feature extraction and visual identification feature extraction; air pressure height feature extraction is to collect air pressure data in real time, and establish an accurate air pressure-altitude conversion model to achieve a preliminary estimate of the floor height. The model adopts an adaptive parameter correction method, analyzes the trend of air pressure changes, establishes a floor switching detection mechanism, and provides basic data support for subsequent identification. The relationship between air pressure and altitude in this embodiment is based on the International Standard Atmospheric Model (ISA), and its formula is:

[0192] Where: h: height (meters), P: air pressure at the measuring point (Pascals), P0: reference air pressure (standard atmospheric pressure at sea level, 101325Pa), T0: reference temperature (standard temperature at sea level, 288.15K), L: temperature lapse rate (usually 0.0065K / m), R: universal gas constant (287.05J / (kg·K)), g: acceleration due to gravity (9.80665m / s2), M: molar mass of air (0.02896kg / mol).

[0193] Environmental structure feature extraction is to extract the unique environmental structure features of the floor through the point cloud data collected by lidar; establish a spatial structure recognition algorithm based on point cloud segmentation, extract the fixed structure features of walls, doors, and columns, and construct feature descriptors. At the same time, a topological relationship analysis method based on graph matching is used to establish a structured description of the floor environment.

[0194] Visual sign feature extraction uses the visual sensors carried by the robot to identify the visual features of floor signs and road signs in the environment through deep learning algorithms; to achieve accurate recognition of digital signs, text signs and special signs; to establish a visual feature library, combined with online learning methods, to continuously improve the adaptability and robustness of the recognition system.

[0195] S5.2.2 Feature fusion recognition, using a feature fusion method based on confidence weighting, effectively fuses the pressure altitude features, environmental structure features and visual identification features described in step 5.1, dynamically adjusts the weights according to the reliability of different features, and establishes an adaptive feature importance evaluation mechanism; the feature fusion method based on confidence weighting includes:

[0196] 1) Feature extraction:

[0197] Air pressure characteristics: altitude difference, rate of change of air pressure.

[0198] Structural features: spatial characteristics of walls, columns and door frames.

[0199] Visual features: Classification and location of floor signs.

[0200] 2) Assign confidence C to each feature 特征 , dynamically calculated based on the current environment:

[0201] where σ 特征 is the noise variance of the feature, Δ o is the magnitude of change;

[0202] 3) Weighted fusion, weighting the confidence of comprehensive features and calculating the final fusion value Fi represents the observed value of a single feature;

[0203] 4) Dynamic weight adjustment, real-time adjustment of weights based on the environment and feature quality

[0204] In the step 5.2 of temporal correlation analysis, the traditional HMM model may be sensitive to noise and lack real-time performance when processing a multi-feature complex dynamic environment (such as the fusion of air pressure, structure and visual features). This embodiment establishes a temporal state transfer model, analyzes the feature change law in the floor switching process, and uses an improved hidden Markov model to describe the floor switching process; the steps include:

[0205] Step 5.2.1 State Modeling

[0206] Define the state space S = {s1, s2, ..., s N}, each state corresponds to a floor number;

[0207] Define observation space O = {o1, o2, ..., o M}, including air pressure, structure and visual characteristics;

[0208] Observation value O t =[o 气压 , o 结构 , o 视觉 ];

[0209] Step S5.2.2 The transition probability matrix AAA of the standard HMM is static:

[0210] A ij =P(s t+1 =s j |s t =s i )

[0211] This embodiment dynamically adjusts the value of A in conjunction with the state transfer matrix, based on the trend and characteristic change rate of the historical observation sequence:

[0212] Where Δo represents the amplitude of feature change, and λ is the weight factor;

[0213] Step S5.2.3 Multi-feature joint observation probability B:

[0214] B j (O t )=P(O t |s t =s j )=w 气压 ·B 气压 +w 结构 ·B 结构 +w 视觉 ·B 视觉 ; Weight w 气压 ,w 结构 ,w 视觉 Adjust based on real-time confidence;

[0215] Step S5.2.4: The inference accuracy of HMM is limited when facing high-dimensional and nonlinear problems. Particle filtering can improve the robustness of floor state estimation. This embodiment combines the particle filtering algorithm to realize probabilistic floor state estimation. The particle filtering algorithm steps include:

[0216] 1) Particle filter initialization: particle set in is the state represented by the particle, is the weight; initial particle distribution:

[0217] 2) State prediction: Use the HMM state transition model to predict particle distribution:

[0218]

[0219] 3) Weight update: Update particle weight according to the current observation value Ot:

[0220] 4) Particle resampling: Avoid particle degradation through resampling:

[0221] 5) State estimation: The final state estimation is the weighted average of the particles:

[0222] S5.3 Historical information optimization step: Establish a historical information database for floor identification, optimize the current identification strategy by analyzing historical identification results, and use online learning methods to continuously update and improve the identification model.

[0223] Preferably, step S5 further includes abnormality detection classification, sensor abnormality detection and positioning navigation abnormality detection;

[0224] The positioning and navigation anomaly detection includes positioning accuracy anomaly detection, navigation deviation anomaly detection and path planning anomaly detection. An uncertainty evaluation model for pose estimation is established, the reliability of the positioning system is monitored in real time, an improved RANSAC algorithm is used for outlier detection, and historical trajectory analysis is combined to realize abnormal behavior identification during navigation. A positioning verification mechanism based on map matching is established to ensure the continuous and reliable operation of the navigation system.

[0225] The uncertainty assessment model real-time monitoring positioning system comprises the following steps:

[0226] 1) Input data: Input multi-sensor data such as lidar, IMU, UWB, visual sensor, current pose estimation result data (position, attitude) and historical trajectory data;

[0227] 2) Uncertainty quantification:

[0228] Gaussian distribution modeling: Assuming that the positioning error conforms to the multivariate normal distribution, construct the covariance matrix ∑:

[0229]

[0230] in Respectively represent the uncertainty in each direction, and represent the covariance;

[0231] Bayesian update: Update the uncertainty distribution of the pose based on the observed data:

[0232] P(x|z)∝P(z|x)P(x)

[0233] Where P(x|z) is the posterior probability, P(z|x) is the observed probability, and P(x) is the prior probability;

[0234] 4) Uncertainty metrics:

[0235] Position reliability: Based on the determinant calculation of the covariance matrix:

[0236] C = det(∑), the larger the C, the higher the uncertainty;

[0237] Ellipsoid radius: Use principal component analysis (PCA) to calculate the major and minor axes of the positioning error ellipsoid and evaluate the positioning range;

[0238] 4) Anomaly Detection: Define the Uncertainty Threshold C threshold , if C>C threshold It is judged as positioning abnormality.

[0239] As a preference, the standard RANSAC (Random Sampling Consensus) uses iterative random sampling to find the set of inliers in the data that best fits the hypothesized model for outlier detection. However, it also has limitations, such as insufficient adaptability to noise and dynamic scenes. This embodiment uses an improved RANSAC algorithm for outlier detection, including:

[0240] 1) Adaptive iteration number:

[0241] Dynamically adjust the number of iterations N: Where p is the success probability of the model, w is the inlier ratio, k is the number of samples, and the inlier ratio w is estimated online to reduce the computational complexity;

[0242] 2) Weighted inlier scoring: assign weight w to inliers i , the weights are related to the residuals:

[0243] where r i is the residual, σ is the noise variance;

[0244] 3) Multi-model verification: Calculate the confidence of different hypothesis models and select the best model:

[0245] where σ m is the standard deviation of the model residuals;

[0246] 4) Combined with particle filtering:

[0247] A particle filter is used to update the inlier distribution after each iteration to improve the robustness in dynamic scenes.

[0248] This embodiment also includes task execution anomaly detection: a hierarchical task execution monitoring framework is adopted to promptly detect abnormal situations in the task execution process through real-time monitoring of task progress, execution quality and resource consumption. Combined with task completion evaluation indicators, accurate identification of task execution anomalies is achieved. At the same time, by analyzing the dependencies between tasks, the risk of anomaly propagation is evaluated to prevent chain failures.

[0249] b) Processing strategy

[0250] Fast fault diagnosis: Adopting intelligent diagnosis methods based on knowledge graphs, establishing a complete fault mode library and diagnosis rule library, realizing automatic extraction and classification of fault features through multi-source data analysis and deep learning algorithms, using improved Bayesian networks for fault reasoning, combining expert experience and knowledge to quickly locate the cause of the fault, and establishing an online learning mechanism for diagnostic knowledge, continuously optimizing the diagnostic model, and improving the diagnostic accuracy;

[0251] Hierarchical processing mechanism: Design a multi-level fault handling strategy, respond in a hierarchical manner according to the severity, impact scope and handling difficulty of the fault, establish a complete handling plan library, formulate corresponding handling processes for different types of faults, establish a fault escalation mechanism, and ensure that major faults are handled in a timely manner;

[0252] Automatic recovery process: Adopt intelligent fault recovery mechanism, establish state backtracking and progressive recovery strategy to ensure that the system can safely recover from fault state, adopt transaction management mechanism to ensure the atomicity and consistency of recovery process, establish recovery effect evaluation mechanism, ensure the reliability of recovery through multiple rounds of verification, provide manual intervention interface, and support manual recovery operation in special circumstances;

[0253] c) Security protection mechanism

[0254] Emergency obstacle avoidance strategy: A multi-level emergency obstacle avoidance system is adopted to achieve all-round perception of environmental obstacles by fusing multi-sensor data, establish an obstacle avoidance decision-making model based on risk assessment, and combine it with a dynamic path planning algorithm to achieve real-time obstacle avoidance control. A hierarchical control architecture is adopted to ensure the safety of obstacle avoidance while maintaining the smoothness of the motion trajectory. At the same time, a collision prediction model is established to achieve early warning of potential dangers.

[0255] Safe parking control: Adopt intelligent safe parking control strategy, select the best parking plan through comprehensive analysis of current status and environmental conditions; establish a multi-level braking control model and adopt different braking strategies according to the degree of urgency.

[0256] Remote takeover mechanism: Utilize the remote takeover system to establish a real-time data transmission channel, adopt multi-level authority management, strictly control the authorization process of remote takeover, establish an audit mechanism for operation behavior, and record all remote operations for later analysis and tracing; adopt a rich remote monitoring interface to support operators to perform precise control and status monitoring.

[0257] By fusing multi-sensor data, we can sense obstacles in the environment, assess their risk levels, and develop real-time obstacle avoidance strategies that take into account both safety and smoothness of motion trajectories. We have thus established an obstacle avoidance decision model based on risk assessment. Multi-sensor data fusion includes lidar, which detects the distance and shape of obstacles; cameras, which identify dynamic and static obstacles; ultrasound, which detects close-range obstacles; and IMU, which provides motion status data.

[0258] Risk assessment indicators: Obstacle distance d: The distance between the obstacle and the robot's closest point. Obstacle speed υ obs : Relative speed of the obstacle. Obstacle size S obs : The area of ​​the obstacle. Current speed υ robot : Robot speed.

[0259] Risk assessment formula: Use a multi-factor weighted model to calculate the obstacle risk level R obs :

[0260] Among them, w1, w2, and w3 are weight coefficients, which are dynamically adjusted according to task requirements.

[0261] Obstacle avoidance decision: Set risk level threshold based on Robs:

[0262]

[0263] Dynamic path planning adjustment: Low risk: Maintain the original path. Medium risk: Plan a local detour path to avoid obstacles. High risk: Emergency brake and re-plan the path.

[0264] The collision prediction model predicts potential collision risks in advance and triggers emergency obstacle avoidance or braking measures. The time-to-collision (TTC) is based on the relative speed between the obstacle and the robot. rel and the relative distance d rel calculate:

[0265] υ rel =υ robot -υ obs

[0266] When TTC≤T threshold , triggering a collision warning;

[0267] TTC is the time to collision, which indicates how long it will take for a collision to occur if both parties continue to move at the current speed based on the current relative speed and distance.

[0268] T threshold is a preset threshold used to determine whether to trigger a collision warning; when TTC≤T threshold When the vehicle is approaching the limit, the system considers that there is a risk of an imminent collision and therefore triggers the warning mechanism.

[0269] Collision probability prediction: combined with the movement direction of the obstacle θ obs and the robot path θ robot : σ θ is the directional sensitivity coefficient.

[0270] The multi-level braking control model dynamically adjusts the braking strategy according to the degree of urgency to ensure safe parking.

[0271] Normal braking: Smooth deceleration in low-risk scenariosa brake =-k·υ robot ; k is the deceleration coefficient,

[0272] Emergency Braking: Stopping a vehicle quickly in high-risk situations: d safe It is a safe stopping distance.

[0273] Forced braking: Triggering forced braking when a collision is about to occur: a brake =-a max , a max is the maximum deceleration of the vehicle.

[0274] Brake switching conditions: According to risk level R obs And the collision time TTC:

[0275]

[0276] The control output is to adjust the brake signal through the PID controller:

[0277] Where e is the deviation between the current speed and the target speed.

[0278] The method of this embodiment realizes reliable multi-floor navigation and autonomous elevator interaction, significantly improving inspection efficiency. The system adopts modular design, which is easy to maintain and upgrade. The positioning accuracy is improved through multi-sensor fusion, and the task execution efficiency is optimized through intelligent scheduling.

[0279] The above description is only a preferred embodiment of the present invention. It should be pointed out that a person skilled in the art can make several improvements without departing from the principle of the present invention. These improvements should also be regarded as within the scope of protection of the present invention.

Claims

1. A multi-floor robot inspection high-precision positioning system, characterized in that: include: The wheeled robot body is equipped with a variety of sensors for performing inspection tasks; Multi-sensor positioning system for precise positioning and navigation; Environmental perception system, used for environmental perception and obstacle avoidance; Elevator interaction module, used to achieve autonomous switching between floors; The multi-sensor positioning system comprises: UWB positioning module, used to provide global positioning reference; Visual positioning module, used to provide local precise positioning; Inertial navigation module to provide continuous attitude and motion information.

2. A high-precision positioning method for multi-floor robot inspection, characterized in that: The steps include: S1. Build a multi-floor environment map, use laser SLAM technology to establish a basic map, integrate visual information to build a semantic map, deploy a UWB base station network on each floor, and calibrate all sensor parameters at the same time; S2. Develop multi-floor inspection routes and add navigation landmarks at the elevator entrances and exits on each floor and key inspection points; S3, the positioning system performs real-time positioning through multi-sensor data fusion, establishes a hierarchical positioning framework from global to local, and realizes precise navigation of the robot; S4, the environmental perception system and obstacle avoidance module work synchronously to perform dynamic path planning and obstacle avoidance control; S5. When the robot needs to switch floors, it switches floors through the elevator interaction module.

3. The multi-floor robot inspection high-precision positioning method according to claim 2 is characterized in that: In step S1, the construction of the multi-floor environment map adopts a hierarchical structure. The base layer uses laser SLAM technology to build a two-dimensional grid map, the middle layer fuses depth camera information to build a three-dimensional feature map, and the top layer combines semantic information to build a navigation map.

4. The multi-floor robot inspection high-precision positioning method according to claim 3 is characterized in that: The calibration of the sensor parameters includes the external parameter calibration of the lidar and depth camera, the position calibration of the UWB base station network, and the time synchronization calibration of the IMU and the visual system. The calibration data is stored in the system configuration file for subsequent multi-sensor fusion positioning.

5. The multi-floor robot inspection high-precision positioning method according to claim 4 is characterized in that: The data processing of the laser radar comprises the following steps: First, the lidar data is adaptively median filtered to remove outliers; Then, invalid data is filtered out by the set distance threshold, and dynamic obstacles are identified and filtered based on the density clustering algorithm; Finally, a laser data confidence assessment mechanism is established to ensure data quality. During the processing, the filtering parameters are adaptively adjusted according to the complexity of the environment to ensure the real-time and effectiveness of data processing.

6. The multi-floor robot inspection high-precision positioning method according to claim 2 is characterized in that: In step S3, the multi-sensor data fusion adopts an improved Kalman filter algorithm, a dynamic weight adjustment mechanism and data credibility evaluation, and the improved Kalman filter algorithm includes: 1) State vector: Where: x k is the position, v k is the speed, θ k For posture, b k is the sensor bias, 2) Nonlinear measurement equation: z k =h(x k )+v k Where: h(·) is the nonlinear measurement function, v k To measure the noise, assume that R k is the noise covariance matrix, 3) Prediction Model:

1. Status prediction x k|k-1 =F k x k-1 +B k u k 2. Covariance prediction Among them: F k is the state transfer matrix, B k is the control input matrix, u k is the control vector, P k-1 is the state covariance matrix at time k-1, Q k is the process noise covariance matrix.

7. The multi-floor robot inspection high-precision positioning method according to claim 2 is characterized in that: In step 3, the hierarchical positioning framework from global to local adopts a multi-layer error compensation method, a smooth estimation algorithm and a feedback correction mechanism; The multi-layer error compensation method includes a systematic error compensation method and a layered compensation for random errors; The system error compensation method compensates for the sensor installation error through an accurate calibration method, establishes a time delay compensation model based on a timestamp, and realizes dynamic compensation of temperature drift; the steps include: 1) Time synchronization calibration Establish a timestamp alignment mechanism to map the data of each sensor to a unified time axis, and use the sliding window method to detect the time drift between sensors; 2) Dynamic compensation model T corrected =T measured +△T Among them, ΔT is the compensation time delay, which is estimated by the dynamic adjustment algorithm; by analyzing the impact of temperature on sensor clock drift, the temperature compensation term is introduced: ΔT=α·ΔT base +β·f(T); where f(T) is the temperature drift function, and α, β are weights; The hierarchical compensation of random errors adopts the optimal estimation strategy by analyzing the noise characteristics and error propagation law to effectively suppress the influence of random errors; The optimal estimation strategy includes the following steps: 1) Noise modeling: Establish a probability distribution model for the noise characteristics of each sensor; introduce a dynamic weight adjustment mechanism to dynamically adjust the sensor weight according to the data credibility: Among them, σ i is the noise variance of the i-th sensor; 2) Optimal fusion strategy, using an extended framework based on Kalman filtering: Among them, K k is the Kalman gain, combined with weight adjustment; The smoothing estimation algorithm comprises: 1) Based on the improved RTS algorithm, continuity constraints are introduced in the backward smoothing process: Among them, the continuity of velocity and acceleration is considered; 2) Smooth posture, using quaternion interpolation method to avoid gimbal lock problem: q smooth =Slerp(q k ,q k+1 ,t); Among them, Slerp is spherical linear interpolation, t is the interpolation factor; The feedback correction mechanism adopts a closed-loop feedback strategy to achieve error back propagation and adaptive adjustment of parameters by evaluating the quality of the fusion result; the closed-loop feedback strategy includes: 1) Feedback framework: The feedback gain is calculated by evaluating the quality of the fusion result: G feedback =f(E error ), where f(E error ) is the error evaluation function; 2) Parameter adaptive adjustment: Use feedback gain to adjust model parameters, dynamically adjust sensor weights, and optimize data fusion effects: Q adjusted =Q+G feedback ·ΔQ.

8. The multi-floor robot inspection high-precision positioning method according to claim 2 is characterized in that: In step S5, the elevator interaction module performs floor switching including S5.1 feature fusion identification step, S5.2 time series correlation analysis step and S5.3 historical information optimization step, wherein the S5.1 feature fusion identification step includes: S5.1.1 Floor feature extraction, the floor features include air pressure height feature extraction, environmental structure feature extraction and visual identification feature extraction; S5.2.2 Feature fusion recognition, using a feature fusion method based on confidence weighting, effectively fuses the pressure altitude features, environmental structure features and visual identification features described in step 5.1, dynamically adjusts the weights according to the reliability of different features, and establishes an adaptive feature importance evaluation mechanism; the feature fusion method based on confidence weighting includes: 1) Assign confidence C to each feature 特征 , dynamically calculated based on the current environment: where σ 特征 is the noise variance of the feature, Δ o is the magnitude of change; 2) Weighted fusion, weighting the confidence of comprehensive features and calculating the final fusion value Fi represents the observed value of a single feature; 3) Dynamic weight adjustment, real-time adjustment of weights based on the environment and feature quality The step 5.2 of time series correlation analysis is to establish a time series state transfer model, analyze the characteristic change law in the floor switching process, use an improved hidden Markov model to describe the floor switching process, and combine the particle filter algorithm to realize probabilistic floor state estimation; the steps include: Step 5.2.1 State Modeling Define the state space S = {s1, s2, ..., s N }, each state corresponds to a floor number; Define observation space O = {o1, o2, ..., o M }, including air pressure, structure and visual characteristics; Observation value O t =[o 气压 , o 结构 , o 视觉 ]; Step S5.2.2 dynamically adjusts the value of A in conjunction with the state transfer matrix, based on the trend and characteristic change rate of the historical observation sequence: Where Δo represents the amplitude of feature change, and λ is the weight factor; Step S5.2.3 Multi-feature joint observation probability B: B j (O t )=P(O t |s t =s j )=w 气压 ·B 气压 +w 结构 ·B 结构 +w 视觉 ·B 视觉 ; Weight w 气压 ,w 结构 ,w 视觉 Adjust based on real-time confidence; The particle filter algorithm steps described in step S5.2.4 include: 1) Particle filter initialization: particle set in is the state represented by the particle, is the weight; initial particle distribution: 2) State prediction: Use the HMM state transition model to predict particle distribution: 3) Weight update: Update the particle weight according to the current observation value Ot: 4) Particle resampling: Avoid particle degradation through resampling: 5) State estimation: The final state estimation is the weighted average of the particles: S5.3 Historical information optimization step: Establish a historical information database for floor identification, optimize the current identification strategy by analyzing historical identification results, and use online learning methods to continuously update and improve the identification model.

9. The multi-floor robot inspection high-precision positioning method according to claim 2, characterized in that: The step S5 also includes abnormality detection classification, sensor abnormality detection and positioning navigation abnormality detection; The positioning and navigation anomaly detection includes positioning accuracy anomaly detection, navigation deviation anomaly detection and path planning anomaly detection, establishes an uncertainty assessment model for pose estimation, monitors the reliability of the positioning system in real time, uses an improved RANSAC algorithm for outlier detection, combines historical trajectory analysis to achieve abnormal behavior identification during navigation, and establishes a positioning verification mechanism based on map matching to ensure the continued reliable operation of the navigation system; The uncertainty assessment model is used to monitor the positioning system in real time, and the steps include: 1) Input data: input multi-sensor data, current pose estimation result data and historical trajectory data; 2) Uncertainty quantification: Gaussian distribution modeling: Assuming that the positioning error conforms to the multivariate normal distribution, construct the covariance matrix ∑: in Respectively represent the uncertainty in each direction, and represent the covariance; Bayesian update: Update the uncertainty distribution of the pose based on the observed data: P(x|z)∝P(z|x)P(x) Where P(x|z) is the posterior probability, P(z|x) is the observed probability, and P(x) is the prior probability; 3) Uncertainty measurement indicators: Position reliability: Based on the determinant calculation of the covariance matrix: C = det(∑), the larger the C, the higher the uncertainty; Ellipsoid radius: Use principal component analysis to calculate the major and minor axes of the positioning error ellipsoid and evaluate the positioning range; 4) Anomaly Detection: Define the Uncertainty Threshold C threshold , if C>C threshold It is judged as positioning abnormality.

10. The multi-floor robot inspection high-precision positioning method according to claim 9, characterized in that: The improved RANSAC algorithm for outlier detection includes: 1) Adaptive iteration number: Dynamically adjust the number of iterations N: Where p is the success probability of the model, w is the inlier ratio, k is the number of samples, and the inlier ratio w is estimated online to reduce the computational complexity; 2) Weighted inlier scoring: assign weight ω to inliers i , the weights are related to the residuals: where r i is the residual, σ is the noise variance; 3) Multi-model verification: Calculate the confidence of different hypothesis models and select the best model: where σ m is the standard deviation of the model residuals; 4) Combined with particle filtering: A particle filter is used to update the inlier distribution after each iteration to improve the robustness in dynamic scenes.

Images