A robot laser localization method that integrates temporal echo reliability and attention

Abstract

This invention provides a robot laser localization method that integrates temporal echo reliability and attention fusion, belonging to the field of autonomous navigation and localization technology for mobile robots. The method maintains a cross-frame temporal sliding window within each laser azimuth sector, extracts ranging variance, intensity variance, inter-frame intensity jumps, motion compensation consistency, and joint ranging intensity discrimination features, and calculates the reliability of regular echoes. It employs three mechanisms—spatial memory attention, inter-frame self-attention, and scanned map cross-attention—to fuse and correct the reliability of regular echoes. It performs dual-cluster distribution detection on the ranging sequence to identify candidate hole facade types and forcibly retains their residual constraints. It sets a residual information matrix based on echo type differences to construct an optimization objective function to solve for the pose. Based on the pose, residuals, and initial estimates, it detects abnormal patterns, removes suspicious residual blocks after triggering, and re-optimizes. This invention significantly improves the localization accuracy and robustness of data center inspection robots.

Description

Technical Field

[0001] This invention belongs to the field of autonomous navigation and positioning technology for mobile robots, and more specifically relates to a robot laser positioning method that integrates temporal echo reliability and attention. Background Technology

[0002] With the large-scale development of data centers, server racks are densely arranged, and the cold and hot aisle environments are complex. Traditional manual inspection methods are inefficient and risky, making them insufficient for the daily operation and maintenance needs of data centers. Inspection robots are gradually becoming an alternative. Data center inspection robots are typically equipped with single-line LiDAR for positioning and navigation. Single-line LiDAR acquires distance information of the surrounding environment through rotational scanning. Combined with an odometry or inertial measurement unit, pose estimation is performed using nonlinear optimization methods that take the laser matching residual and odometry residual as constraints. Each LiDAR echo contains ranging values ​​and reflection intensity information. The reflection intensity is related to the material, color, and incident angle of the target object and can serve as an auxiliary basis for judging the echo quality. However, the presence of various objects in data center server rooms that affect the reliability of laser echoes poses unique challenges to positioning tasks in this scenario.

[0003] Specifically, highly reflective objects such as glass partitions and polished metal surfaces are prone to specular reflection and multipath effects, causing laser echoes to contain false distance information. Low-reflective objects such as black-painted cabinets have strong light absorption capabilities, resulting in low echo intensity and large fluctuations in distance measurement values ​​when glancing at large angles. Furthermore, the high degree of repetition in the appearance of cabinet rows can easily lead to matching ambiguities. Cabinet front doors often use ventilation mesh or perforated plate structures; on these surfaces, the laser distance measurement value at the same azimuth angle will regularly jump between the cabinet door surface and the depth of the holes. This phenomenon is fundamentally different from the random false echoes produced by glass and other materials. The diverse sources and mechanisms of these interferences make it difficult to judge the reliability of laser echoes using a single criterion.

[0004] To address the aforementioned issues, existing laser positioning methods typically employ two approaches. The first is a uniform weighting strategy, assigning the same optimization weight to all laser points. Stable and unstable echoes receive equal weight in the optimization process, and the residual terms corresponding to unstable echoes may dominate the optimization direction, leading to pose estimation bias. The second is a single-frame intensity threshold strategy, using whether the reflection intensity exceeds a threshold as the filtering criterion. However, this strategy struggles to effectively identify the core problem of ranging instability in low-reflectivity objects and may also exclude valid echoes with high intensity but stable ranging. Both strategies rely on single-frame information for judgment, failing to utilize the evolution of echoes across multiple frames to distinguish different interference patterns. Furthermore, historical observation information accumulated from multiple passes through the same area is not utilized, requiring re-evaluation when interference patterns recur in certain scenarios. In addition, existing methods generally fail to distinguish between regular ranging fluctuations on perforated plate structures and random spurious echoes generated by glass, etc. Simply treating the former as unstable echoes results in the loss of significant effective constraints. Summary of the Invention

[0005] To address the above issues, the present invention aims to provide a robot laser localization method that integrates temporal echo reliability and attention. This method extracts multi-dimensional statistical features through a cross-frame temporal sliding window and integrates three mechanisms—spatial memory, inter-frame self-attention, and scan map cross-attention—for reliability correction. Combined with dual-cluster detection and differentiated residual weighting, it effectively distinguishes between dual-cluster jumping and random interference in perforated plate structures, solving the problem of erroneous rejection of lateral constraints in cold aisles, and significantly improving the localization accuracy and robustness of data center inspection robots.

[0006] To achieve the above objectives, the present invention employs the following technical solution: In a first aspect, embodiments of this application provide a robot laser localization method that fuses temporal echo reliability and attention, comprising the following steps: S101: The robot uses a single-line lidar to scan and obtain an initial point set, while also acquiring odometry or inertial measurement unit data to provide an initial estimate of the positioning attitude; each frame of the initial point set includes azimuth, range, and reflection intensity. S102: Maintain a sliding window of fixed length for each laser azimuth sector, record the ranging sequence and reflection intensity sequence of the sector in the past multiple consecutive frames, and calculate the motion compensation residual sequence between frames based on the initial pose estimate. S103: Calculate the temporal ranging fluctuation characteristics based on the ranging sequence, calculate the temporal intensity fluctuation characteristics and inter-frame intensity abrupt change characteristics based on the reflection intensity sequence, calculate the motion compensation consistency characteristics based on the motion compensation residual sequence, and calculate the ranging-intensity joint discrimination characteristics based on the temporal ranging fluctuation characteristics and the temporal intensity fluctuation characteristics. Combine all the calculated features into a regular feature vector and obtain the regular echo confidence through nonlinear mapping. S104: Perform dual-cluster distribution detection on the ranging sequence to determine whether there is a dual-cluster pattern caused by the hole structure, and determine whether the dual-cluster pattern satisfies the kinematic consistency constraint based on the motion compensation residual sequence. If it does, determine that the sector is a candidate type of hole facade and generate a hole facade identifier. At the same time, record the near-end cluster center value obtained by dual-cluster detection. S105: Read relevant data from the pre-built two-layer map, and use spatial memory attention mechanism, inter-frame self-attention mechanism and scan map cross-attention mechanism to fuse and correct the credibility of the rule echo to obtain the final echo credibility; S106: Determine the echo type of each sector based on the hole facade identifier and the rule echo confidence level, and adopt differentiated residual block establishment strategy and weight configuration strategy for different echo types; the echo types include at least stable type, hole facade type and abnormal type; S107: Selectively establish laser matching residual blocks and set corresponding weight coefficients according to the echo type and final echo confidence. Combine the established laser matching residual blocks with the odometry residual term to form an optimization objective function for solving, and obtain the robot's optimal pose and residual information of each sector. S108: Based on the robot's optimal pose, the residual information of each sector, and the initial pose estimate, detect whether an abnormal mode is triggered. If triggered, remove the laser matching residual block of the corresponding suspicious sector and adjust the weight coefficient of the remaining residual block. Use the initial pose estimate as the initial value to re-optimize and update the robot's optimal pose. S109: Update the geometric features to the double-layer map according to the echo type of each sector, and update the statistical information in this observation to the double-layer map at the same time.

[0007] In an optional implementation, step S101 includes: The robot uses a single-line lidar to continuously scan during the inspection process, obtaining a set of laser points per frame. Each laser point includes an azimuth angle. Distance value and reflection intensity , forming the original point set azimuth angle Indicates the first The scanning angle of each laser point in the polar coordinate system of the lidar, and the ranging value. This indicates the measured distance and reflection intensity corresponding to the laser point. This indicates the amplitude of the echo signal received by the laser point; Simultaneously, the robot's relative motion information is acquired through a wheeled odometry system or an inertial measurement unit, and the initial estimated value of the pose at the current moment is calculated by combining it with the pose at the previous moment. , representing the initial pose estimation of the robot in a two-dimensional plane.

[0008] In an optional implementation, step S102 includes: The maintenance length for each laser azimuth sector is A sliding window records the sector in the most recent Ranging sequence in the frame and reflection intensity sequence ; Based on the initial pose estimate, calculate the motion compensation residual sequence between each frame. subscript Indicates the frame number; Motion Compensation Residual The calculation formula is:

[0009] in, This indicates the use of the initial pose estimate. The relative change between adjacent frames will The distance measurement value of this sector at that time Projected onto the odometer The predicted distance is obtained in the robot's coordinate system at that time.

[0010] In an optional implementation, step S103 includes: Based on ranging sequence The distance measurement variance is calculated using the following formula. As a characteristic of time-series ranging fluctuations:

[0011] in, for Frame ranging mean; Based on reflection intensity sequence The strength variance is calculated using the following formula. As a characteristic of temporal intensity fluctuation:

[0012] in, for Mean frame reflection intensity; The inter-frame intensity jump is calculated using the following formula. As a feature of inter-frame intensity abrupt change:

[0013] Based on motion-compensated residual sequence The motion compensation consistency feature is calculated using the following formula:

[0014] in, Let be the standard deviation of the motion compensation residual sequence. To preset the standard deviation threshold of motion compensation residuals, The range of values ​​is ; The joint feature of ranging intensity is calculated using the following formula:

[0015] in, To prevent small quantities from being divided by zero, Used to distinguish different interference modes; The above five features are combined into the rule feature vector:

[0016] Through nonlinear activation function The regular echo confidence level is obtained by performing nonlinear mapping. ; in This is the current laser azimuth sector identifier. This represents a preliminary assessment of the sector's echo reliability based solely on timing rule characteristics.

[0017] In an optional implementation, step S104 includes: Clustering algorithm is used to analyze the ranging sequence. Clustering is performed on all values ​​to obtain the proximal cluster center values. and remote cluster center value The proximal cluster center values ​​correspond to the center values ​​of clusters with smaller distances, and the distal cluster center values ​​correspond to the center values ​​of clusters with larger distances. The sum of the proportions of the two main clusters is calculated using the following formula:

[0018] in and These represent the proportions of data points belonging to the near-end cluster and the far-end cluster, respectively. This indicates the proportion of bi-clustered data to the total data points; if If so, it is determined that a dual-cluster pattern exists, where The preset threshold for the proportion of dual clusters; According to the motion compensation residual sequence If the overall change in distance measurement in this sector during the robot's translation process is monotonically changed with distance travel, and the kinematic consistency constraint is satisfied, it is determined to be a candidate type of hole facade and the hole facade identifier is generated. ; This indicates that the sector is marked as a candidate type for a perforated facade. This indicates that it has not been marked.

[0019] In an optional implementation, step S105 includes: Geometric stability layer from a pre-built two-layer map Read geometric reference data from the middle and from the suspected layer of unstable echo. The system reads historical statistics, including the number of times the inspection passed through. Number of historical anomalies The distribution ratio of historical echo types, the median of historical representative distances, and the mean confidence level of the previous regular echo; The execution flow of the spatial memory attention mechanism is as follows: Divide the environment into grids It then concatenates summary information on the current location encoding, channel identifier, and current rule echo reliability to generate a query vector. ; From the suspected layer of unstable echo Read the historical statistical information of the current grid and its neighborhood; Spatial memory attention weights are calculated using the following formula. :

[0020] in, The key matrix is ​​formed by encoding the historical statistical information of the current grid and its neighborhood. The dimension of the key vector; Spatial memory attention weights are obtained by weighted summation of attention values. ;in The value matrix is ​​formed by encoding the historical statistical information of the current grid and its neighborhood. It is a scalar, and its value range is... This indicates that the spatial memory attention mechanism makes a priori assessment of the credibility of the current raster echo based on historical statistical information; The execution flow of the inter-frame self-attention mechanism is as follows: The current frame and all frames before and after it. Frames form a timing window, in which The preset number of frames per side; Extract the ranging value of each sector in each frame Reflection intensity Motion compensation residuals Confidence of regular echoes The echo type encoding is concatenated to form a feature vector. and form the input sequence. ;in This is the length of the timing window; The query matrix is ​​obtained through three linear transformations. Key matrix Sum matrix ;in The weight matrix is ​​a learnable weight matrix; The inter-frame self-attention weights are calculated using the following formula. :

[0021] Using the attention weights For the value matrix We perform a weighted summation to obtain the self-attention weighted feature representation matrix. ; The feature representation matrix Stacking After the attention layer, the echo confidence correction value for each sector is mapped through a fully connected layer. ; The execution flow of the cross-attention mechanism for the scanned map is as follows: Extract the feature representation of the current sector, including the sector azimuth angle. Distance value Reflection intensity and the credibility of regular echoes Encode it to construct a query vector From the geometrically stable layer The matching window reads the stable distance values ​​and intensity statistics of the sector during historical stable matches, and from the suspected unstable echo layer Historical anomaly statistics for the region were read from the spatial memory record, and the above data was encoded to construct a key vector. Read data from the corresponding historical matching scores or anomaly weights and encode it to construct a value vector. ; The cross-attention weights of the scanned map are calculated using the following formula:

[0022] Using the attention weights For the value vector Weighted summation is performed to obtain a context vector that integrates historical map information. This vector is then mapped to the location of each sector via a fully connected layer. Figure 1 Consistency correction value ; The final echo confidence level is generated using the following formula. :

[0023] in To preset the weighting coefficients, To preset a minimum confidence level, This is a truncation function that restricts the output value to a certain range. Within the range.

[0024] In an optional implementation, step S106 includes: like and Higher than the preset confidence threshold If it is, then it is determined to be of the stable type and denoted as the echo type. At this point, the laser matching residual block is established normally and the normal residual information matrix is ​​used. As weight; like If the motion compensation is consistent, then it is determined to be the aforementioned hole facade type and recorded as the echo type. At this point, the proximal cluster center value is taken. As the representative distance Preserve laser matching residual blocks and use medium- and low-weight residual information matrices. Huber robust loss function; the lower limit of echo confidence for the perforated facade type is taken as the preset minimum confidence threshold for the perforated facade type. ; If neither the conditions for the stable type nor the conditions for the perforated facade type are met, then it is judged to be the abnormal type and recorded as the echo type. At this point, either laser matching residual blocks are not established or the minimum residual information matrix is ​​used. As weights.

[0025] In an optional implementation, step S107 includes: According to the echo type and the credibility of the final echo Set the residual information matrix subscript For sector index, Indicates the first Residual information matrix of each sector: when hour, ;in The preset weight scaling index; when hour, ; when hour, ; From the geometrically stable layer Read the corresponding raster map distance value; register the representative distance with the map distance value of the corresponding raster in the geometrically stable layer to construct a residual function. ; The laser-matched residual block and the odometer residual block are combined to form the following nonlinear least squares problem:

[0026] in Let odometer residual function be used. Huber robust loss function; With the initial pose estimate The robot's optimal pose is obtained by iteratively optimizing the initial value. Simultaneously, it records the residual information of each sector.

[0027] In an optional implementation, step S108 includes: The optimal pose of the robot With the initial pose estimate The pose deviation between the two is compared with a preset pose deviation threshold. If it exceeds the preset pose deviation threshold, an abnormal mode is triggered. Total cost of laser matching Compare with a preset cost threshold; if the cost exceeds the preset cost threshold, an abnormal mode is triggered. The residuals of each sector and the suspected unstable echo layers are compared. The location of historical abnormal areas is compared. If high residual points are concentrated in the historical abnormal areas, an abnormal mode is triggered. exception types The percentage of sectors is compared with a preset abnormal percentage threshold. If the percentage exceeds the preset abnormal percentage threshold, an abnormal mode is triggered. After triggering the abnormal mode, sectors with residuals higher than a preset residual threshold are marked as suspicious sectors and removed from the optimization problem. The residual information matrix of the remaining laser matching residual blocks is multiplied by a shrinkage factor. ,in The preset weighting contraction coefficient is used, with the initial pose estimate as the initial value. Re-execute step S107 for iterative optimization to update the robot's optimal pose using the initial values. And update the unstable echo suspicious layer. Record.

[0028] In an optional implementation, step S109 includes: For each sector, if the ranging variance of that sector over the past multiple consecutive frames... Below the variance threshold and the echo type If the anomaly does not belong to the aforementioned anomaly type, then the geometric feature points of the current frame for that sector are stored in the geometric stabilization layer. ; For each grid Update the number of times the inspection has been completed. If the echo type Then update the historical anomaly occurrence count. And update the echo type distribution ratio; the unstable echo suspected layer It does not participate in the calculation of scan matching scores, but is only used for attention mechanism queries and conflict detection.

[0029] As can be seen from the above technical solutions, the present invention has the following advantages: The robot laser localization method based on temporal echo reliability and attention fusion provided in this application extracts ranging variance, intensity variance, inter-frame intensity jump, motion compensation consistency, and ranging-intensity joint discrimination features by constructing a cross-frame temporal sliding window. It then uses three attention mechanisms—spatial memory attention, inter-frame self-attention, and scan map cross-attention—to fuse and correct the reliability of regular echoes, achieving accurate evaluation of the reliability of echoes in each laser sector. Simultaneously, it effectively distinguishes between regular double-cluster jumps and random multipath interference in hole facades through dual-cluster distribution detection, forcibly retaining effective constraints of hole facade types and avoiding the problem of cold aisle lateral constraints being completely eliminated. Based on this, it adaptively sets a differentiated residual information matrix according to the echo type, ensuring that stable echoes participate normally in optimization, hole facade echoes retain constraints in a reduced-weight manner, and abnormal echoes are reduced in weight or eliminated. Combined with the separation and maintenance of the geometrically stable layer and the unstable echo suspicious layer in the dual-layer map and the re-optimization mechanism after abnormal triggering, it significantly improves the accuracy and robustness of laser localization for inspection robots in complex data center environments.

[0030] This application extracts ranging variance, intensity variance, inter-frame intensity jumps, motion compensation consistency, and ranging-intensity joint discrimination features by constructing a cross-frame temporal sliding window. These five regular features are combined into a regular feature vector and mapped to regular echo reliability via a nonlinear activation function, achieving a preliminary quantitative assessment of the echo reliability for each laser sector. Compared to traditional methods that rely on single-frame intensity thresholds or equal weighting of all echoes, this application fully utilizes the statistical evolution of ranging values, reflection intensity, and motion compensation residuals over time. It can effectively identify echo characteristics under different interference modes, such as unstable ranging timing but low intensity in black cabinets, abnormal intensity jumps caused by glass multipath, and high intensity but stable ranging in metal components. This fundamentally overcomes the shortcomings of single-frame intensity threshold strategies, which have both high false negative and false positive rates.

[0031] This application employs a spatial memory attention mechanism, an inter-frame self-attention mechanism, and a scan map cross-attention mechanism to intelligently fuse and correct the credibility of rule-based echoes. Specifically, spatial memory attention calculates prior confidence weights based on the number of inspection passes, historical anomaly occurrences, and echo type distribution ratios stored in the unstable echo suspicion layer. Inter-frame self-attention performs self-attention calculations on the temporal features of consecutive frames to output a corrected echo credibility value. Scan map cross-attention cross-compares the current sector features with the geometrically stable layer and the unstable echo suspicion layer to output a corrected map matching score value. The combined outputs of these three attention mechanisms are weighted and fused to obtain the final echo credibility. This application constructs a complete information utilization link from single frame to multiple frames and from the current instance to history, enabling echo credibility assessment to possess adaptive learning and memory capabilities.

[0032] This application employs bi-cluster clustering analysis on the ranging sequence, calculates the sum of the proportions of the two main clusters, and checks whether the motion compensation residuals satisfy the channel geometric constraints during robot translation. Sectors exhibiting regular bi-cluster fluctuations are identified as perforated facade types rather than directly eliminated, and their laser matching residual constraints are forcibly retained. In data center cold aisles, rack front doors often have ventilation mesh or perforated panel structures. Regular fluctuations in ranging values ​​between the distance to the rack door and the depth of the perforations are a common physical phenomenon in such scenarios. Traditional methods that simply classify these as unstable and filter them out entirely would result in insufficient lateral positioning constraints. This application accurately distinguishes between the regular bi-cluster fluctuations of perforated panel structures and random multipath interference from glass surfaces, ensuring the usability of lateral positioning in perforated rack facade scenarios.

[0033] This application employs a differentiated residual information matrix setting strategy for laser matching residual blocks based on echo type: for stable types, the weights of the normal residual information matrix are scaled exponentially by the final echo confidence; for hole / facade types, a medium-to-low weight matrix is ​​multiplied by a confidence scaling factor and used in conjunction with the Huber robust loss function for optimization, and these types are not eliminated entirely regardless of spatial memory attention weights; for abnormal types, no residual blocks are created or they are assigned the minimum weight. Compared to the drawbacks of a uniform weight strategy where unreliable and reliable echoes have equal status in optimization, easily dominating the optimization direction and getting trapped in local optima, this application achieves an optimal balance between reliability and constraint sufficiency during optimization through three levels of differentiated weight configuration. This ensures that stable echoes dominate the optimization direction, hole / facade echoes retain effective lateral constraints through weight reduction, and abnormal echoes are suppressed to the minimum impact.

[0034] This application employs a two-layer map separation and maintenance mechanism consisting of a geometrically stable layer and an unstable echo suspicion layer. Stability-tested geometric features are stored in the geometrically stable layer as a scanning matching benchmark, while unstable echo statistics, such as the number of inspection passes, historical anomaly occurrences, and echo type distribution ratios, are stored in the unstable echo suspicion layer for attention mechanism queries and conflict detection. This is combined with re-optimization steps such as suspicious residual block removal after anomaly pattern triggering, weight reduction of remaining residual blocks, and re-optimization using odometry pose as the initial value. This mechanism completely decouples stable geometric references from unstable statistical information in terms of storage and functionality. Furthermore, the anomaly re-optimization process enhances the system's self-recovery capability when matching quality deteriorates, effectively reducing the risk of matching ambiguity caused by geometric repetition in parallel rack columns in data centers. This provides comprehensive assurance for positioning consistency and robustness in long-term inspection tasks. Attached Figure Description

[0035] To more clearly illustrate the technical solution of the present invention, the accompanying drawings used in the description will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0036] Figure 1 A flowchart illustrating the robot laser localization method based on temporal echo reliability and attention fusion provided in this application.

[0037] Figure 2 A flowchart illustrating another robot laser localization method based on temporal echo reliability and attention fusion provided in this application. Detailed Implementation

[0038] Various embodiments of this disclosure will be described more fully in the following detailed description of the specific steps of the robot laser localization method based on temporal echo reliability and attention fusion. This disclosure may have various embodiments, and adjustments and changes may be made therein. However, it should be understood that there is no intention to limit the various embodiments of this disclosure to the specific embodiments disclosed herein, but rather this disclosure should be understood to cover all adjustments, equivalents, and / or alternatives falling within the spirit and scope of the various embodiments of this disclosure.

[0039] In the following, the terms “comprising” or “may include”, which may be used in various embodiments of this disclosure, indicate the presence of the disclosed functions, operations, or elements, and do not limit the addition of one or more functions, operations, or elements. Furthermore, as used in various embodiments of this disclosure, the terms “comprising,” “having,” and their cognates are intended only to indicate a particular feature, number, step, operation, element, component, or combination of the foregoing, and should not be construed as primarily excluding the presence of one or more other features, numbers, steps, operations, elements, components, or combinations of the foregoing, or the possibility of adding one or more combinations of the foregoing.

[0040] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0041] Please see Figure 1 The diagram shows a flowchart of a robot laser localization method based on temporal echo reliability and attention fusion in a specific embodiment. The method includes: S101: The robot uses a single-line lidar to scan and obtain an initial point set, while also acquiring odometry or inertial measurement unit data to provide an initial estimate of the positioning attitude; each frame of the initial point set includes azimuth, range, and reflection intensity.

[0042] In a specific implementation, the robot, equipped with a single-line LiDAR, moves along a planned path for inspection within the cold or hot aisles of the data center. The LiDAR continuously rotates and scans at a fixed frequency, acquiring a set of discrete laser point cloud data per frame. Each laser point contains three dimensions of information: azimuth angle... This indicates the scanning angle of the laser point in the polar coordinate system of the lidar, and the ranging value. This indicates the measured distance from the laser radar to the point of reflection from the target object, and the reflection intensity. This represents the amplitude of the echo signal received by the laser point, which is related to factors such as the surface material, color, and angle of incidence of the target object. All laser points together constitute the original point set of this frame. .

[0043] The robot records the number of wheel rotations and their direction changes using a wheel-mounted odometer, or collects acceleration and angular velocity information using an inertial measurement unit. This relative motion information, combined with the optimized and validated pose from the previous moment, allows the initial estimate of the pose at the current moment to be calculated. The initial estimate of the pose. Belongs to the two-dimensional special Euclidean group Including robots in a two-dimensional plane coordinate, The coordinates, orientation angle, and three degrees of freedom serve as the initial values ​​for subsequent motion compensation and iterative optimization. In this embodiment, the scanning frequency of the lidar is... This means acquiring 10 frames of point cloud data per second, with each frame containing approximately 1080 laser points, and an azimuth resolution of approximately [missing information]. .

[0044] S102: Maintain a sliding window of fixed length for each laser azimuth sector, record the ranging sequence and reflection intensity sequence of the sector in the past multiple consecutive frames, and calculate the motion compensation residual sequence between frames based on the initial pose estimate.

[0045] In a specific implementation, the entire scanning field of view of the lidar is spaced at preset angles. It is divided into several discrete laser azimuth sectors. For example, the scanning field of view is divided into 360 sectors, each sector covering The range of angles, representing the azimuth angle. Take the center value of the angle interval of this sector. Each original laser point is determined based on its azimuth angle. It is assigned to the corresponding sector, which satisfies... It belongs to that sector.

[0046] Maintain a separate database of length for each sector. A sliding window. This window records the sector in the past consecutive... The ranging values ​​and reflection intensities of all laser points received in the frame are used to form a ranging sequence. and reflection intensity sequence subscript Indicates the frame number. The oldest frame in the corresponding window. Corresponding to the current frame. Window length. The value of directly affects the statistical significance of time series features. If If the value is too small, the statistical sample is insufficient, and the features are easily affected by noise; if If the value is too large, it may introduce outdated environmental information. In this embodiment, it is preferred that... This means that the sliding window covers the historical data of the previous 20 frames, ensuring statistical stability while also taking into account the speed of response to environmental changes.

[0047] Meanwhile, based on the initial pose estimate The relative change between adjacent frames, relative to the ranging value of the previous frame. Perform motion compensation. The distance measurement value of this sector at that time The corresponding spatial point is projected onto the odometry relative transformation. The predicted distance is obtained in the robot's coordinate system at that time. The predicted distance differs from the actual distance measured at the current moment. The difference between them is the motion compensation residual. .

[0048] The sign of the value reflects the direction of deviation between the predicted distance and the actual distance, and its magnitude reflects the degree of deviation. The closer the value is to 0, the better the consistency between the laser point and the odometer motion prediction, and the more reliable the echo.

[0049] S103: Calculate the temporal ranging fluctuation characteristics based on the ranging sequence, calculate the temporal intensity fluctuation characteristics and inter-frame intensity abrupt change characteristics based on the reflection intensity sequence, calculate the motion compensation consistency characteristics based on the motion compensation residual sequence, and calculate the ranging-intensity joint discrimination characteristics based on the temporal ranging fluctuation characteristics and the temporal intensity fluctuation characteristics. Combine all the calculated features into a regular feature vector and obtain the regular echo confidence through nonlinear mapping.

[0050] In this specific implementation, this step calculates five time-series statistical features with clear physical meanings based on the three time-series sequences maintained in step S102, serving as the quantitative basis for the reliability of the regular echoes. The specific calculation process is as follows: First, the variance of the distance measurement This feature quantifies the dispersion of the sector's ranging value in the time dimension, and the calculation formula is as follows:

[0051] in for Frame ranging mean, When the laser scans a stable metal baseboard or cabinet door frame, the ranging value fluctuation over 20 consecutive frames is typically within... Within this range, the distance measurement variance is... The value of approaches 0. When the laser is incident on the surface of the glass partition, specular reflection and multipath effects cause the ranging value to jump from tens of centimeters to several meters within a few frames. A significant increase. The larger this characteristic value, the more unstable the ranging is, and the less reliable the echo is.

[0052] Second, intensity variance This feature quantifies the dispersion of the sector's reflection intensity over time, and the calculation formula is as follows:

[0053] in for Mean frame reflection intensity Specular reflection from the glass surface can cause the intensity signal to rise abnormally in certain frames and then drop rapidly. The value is relatively large. The black-painted cabinet surface absorbs light, resulting in a persistently low intensity signal, which will also lead to... Increase. When a laser scans a uniformly diffuse reflective surface, such as a white wall or cement floor, the reflection intensity signal remains stable over time. The value is relatively small.

[0054] Third, inter-frame intensity jump This feature reflects the average variation in reflection intensity between adjacent frames, and is calculated using the following formula:

[0055] This feature is related to the intensity variance. The difference lies in the fact that intensity variance reflects the overall dispersion, while inter-frame intensity jumps are more sensitive to instantaneous changes within a single frame. The instantaneous specular reflection produced by a glass door at a specific angle often only affects a single frame or two consecutive frames, manifesting as a one-time intensity jump rather than a continuous increase in variance. Able to effectively capture such transient anomalies This may be due to the averaged effect caused by most frames being normal within the window. For persistent, unstable echoes, such as persistent low-intensity fluctuations on the surface of a black cabinet, both intensity variance and inter-frame intensity jumps will be reflected, but the former will show a more significant increase.

[0056] Fourth, consistency of motion compensation This feature measures consistency based on the standard deviation of the motion-compensated residual sequence, and the calculation formula is as follows:

[0057] in Let be the standard deviation of the motion compensation residual sequence. This is a preset threshold for the standard deviation of the motion compensation residuals. The meaning of this feature is: when the standard deviation of the motion compensation residual sequence... Much smaller than the preset threshold When, the ratio Approaching 0, then A value close to 1 indicates that the residual sequence is highly concentrated, motion compensation is consistent, and the echo is reliable. When the standard deviation of the residual sequence approaches or exceeds 1, it indicates that the residual sequence is highly concentrated, motion compensation is consistent, and the echo is reliable. At this time, the eigenvalue approaches 0. For example, when a robot travels along a straight path, the motion compensation residual of a certain sector is close to zero for multiple consecutive frames. If the sector is affected by multipath interference, causing the residual to fluctuate significantly and irregularly, then . The value of directly affects the sensitivity of this feature. In this embodiment, the preferred value is... That is, when the standard deviation of the residuals is less than 5 cm, the consistency is considered to be good.

[0058] Fifth, joint characteristics of ranging intensity This feature is the ratio of the distance variance to the intensity variance, calculated using the following formula:

[0059] in To prevent the possibility of dilution to zero, this embodiment preferably uses a small amount of material. The core function of this feature is to distinguish between two interference modes that have similar ranging performance but completely different physical mechanisms. The first mode is glass multipath interference, characterized by large ranging variance and large intensity variance. Both the numerator and denominator are relatively large, and the ratio tends to be at a moderate level. The second mode is black cabinet grazing interference or multipath interference, characterized by a large ranging variance but a small intensity variance. The numerator is much larger than the denominator, resulting in a significantly increased ratio. This ratio characteristic can effectively distinguish between the two types of interference sources, providing a more targeted basis for subsequent correction of attention mechanisms.

[0060] The above five features are combined into a regular feature vector:

[0061] Through a pre-trained nonlinear activation function Map the vector to The interval is used to obtain the confidence level of the regular echo. .

[0062] The activation function A multilayer perceptron model can be used for training offline using labeled echo samples. The training dataset should cover various echo samples from typical data center scenarios, including stable metal surfaces, glass partitions, black-coated cabinets, and perforated cabinet surfaces. As a preliminary assessment of the reliability of the echo in this sector, a value closer to 1 indicates a more reliable echo, while a value closer to 0 indicates a less reliable echo. When all five temporal characteristics exhibit stable behavior—that is, small ranging variance and intensity variance, small inter-frame jumps, and good motion compensation consistency—then... Output a high confidence value close to 1. When any feature exhibits abnormal fluctuations... The corresponding reduction.

[0063] S104: Perform dual-cluster distribution detection on the ranging sequence to determine whether there is a dual-cluster pattern caused by the hole structure, and determine whether the dual-cluster pattern satisfies the kinematic consistency constraint based on the motion compensation residual sequence. If it does, determine that the sector is a candidate type of hole facade and generate a hole facade identifier. At the same time, record the near-end cluster center value obtained by dual-cluster detection.

[0064] In a specific implementation, this step involves conducting specialized testing on the unique physical characteristics of the perforated panel structure at the front door of the data center cabinet.

[0065] Firstly adopt -Means clustering algorithm For ranging sequence All values ​​are subjected to binary clustering. The algorithm iteratively optimizes the division of all distance measurements into two clusters, ensuring that distance measurements within the same cluster are as close as possible and distance measurements between different clusters are as far apart as possible, ultimately obtaining the center values ​​of the two clusters. and These correspond to clusters on the cabinet doors that are closer together and clusters deep within the holes that are farther away.

[0066] After clustering is completed, the sum of the proportions of the two main clusters is calculated. .

[0067] in and These represent the proportions of data points belonging to the near-end cluster and the far-end cluster, respectively. This value reflects the proportion of data points exhibiting a bi-cluster distribution in the ranging sequence out of the total data points. If this proportion exceeds a preset bi-cluster proportion threshold... ,in In this embodiment, a value of 0.6 is preferred, which indicates that the sector exhibits a significant dual-cluster pattern.

[0068] Based on the establishment of the bi-cluster pattern, further examination is conducted to determine whether the bi-cluster pattern satisfies the kinematic consistency constraint. This is achieved using motion-compensated residual sequences. Determine whether the overall distance measurement of this sector changes monotonically with distance traveled during robot translation. If the jumping of the dual clusters is a physical normality caused by the perforated plate structure, the overall distance measurement should change regularly with the robot's movement, and the frequency of near and far clusters should smoothly transition with the relative position of the robot and the cabinet, rather than exhibiting chaotic lateral jumping. As the robot gradually moves from a position directly facing the cabinet to the end of the cabinet row, the proportion of echoes deep within the perforations should gradually decrease, and this change should be consistent with the robot's movement trajectory. If the dual cluster pattern of this sector simultaneously satisfies the proportion threshold condition and the kinematic consistency constraint, it is determined to be a candidate type of perforated facade, and a perforated facade identifier is generated. .

[0069] in This indicates that the sector is marked as a candidate type for a perforated facade. This indicates that the cluster is not labeled. The proximal cluster center values ​​are also recorded. As a candidate value representing the distance.

[0070] S105: Read relevant data from the pre-built two-layer map, and use spatial memory attention mechanism, inter-frame self-attention mechanism and scan map cross-attention mechanism to fuse and correct the credibility of the rule echo to obtain the final echo credibility.

[0071] In a specific implementation, this step employs three attention mechanisms to perform multi-level intelligent fusion correction of the credibility of rule echoes, with each mechanism having its own division of labor and working in concert.

[0072] 1. The specific execution process of the spatial memory attention mechanism is as follows: Geometric stability layer from a pre-built two-layer map Read geometric reference data from the middle and from the suspected layer of unstable echo. Historical statistical information is retrieved from the database. This mechanism divides the environment into square grids with sides of 0.5m, storing the number of times each grid has been inspected. Number of historical anomalies The query vector is generated by concatenating the raster coordinates embedding vector of the current raster, channel identifier, and summary information of the current rule echo confidence, along with the median historical representative distance and the mean confidence level of the previous rule echo. .from The system reads the historical records of the current raster and its 3×3 neighborhood of 9 rasters, and encodes various statistical information of these records into a key matrix. Sum matrix The key matrix Includes historical location encoding, historical echo type encoding, and historical regular echo confidence encoding, value matrix Includes historical anomaly weights and echo confidence statistics. Attention weights are calculated using the softmax function.

[0073] in Let be the dimension of the key vector. Used for scaling to ensure training stability.

[0074] Again We get the following by weighted summation:

[0075] This weight It is a scalar, and its value range is... If a grid has an abnormally frequent history, output a low value to indicate caution. If the history type is mainly perforated facades, maintain the perforated facade strategy. If it is the first time passing through, keep it neutral and close to 0.5.

[0076] 2. The specific execution flow of the inter-frame self-attention mechanism is as follows: The current frame and all frames before and after it. The frames form a timing window, totaling Frame, in which For each sector in each frame, extract the ranging value. Reflection intensity Motion compensation residuals Confidence of regular echoes Echo type encoding concatenates the information from these five dimensions into a feature vector. , dimension The feature vectors of all frames form the input sequence. Through three learnable linear transformation matrices. Will Mapped to query matrix Key matrix Sum matrix :

[0077] Calculate the self-attention weights:

[0078] Use this weight to We get the following by weighted summation:

[0079] Stacking After the attention layer, the echo confidence correction value for each sector is mapped through a fully connected layer. The core capability of this mechanism lies in its ability to identify specular flashes on glass surfaces as transient anomalies rather than persistent hole-like facade problems. When a glass door exhibits strong specular reflection in a particular frame, the intensity of that frame is abnormally high, but it returns to normal in adjacent frames. The inter-frame self-attention mechanism can identify this transient pattern and apply a low correction value.

[0080] 3. The specific execution flow of the cross-attention mechanism for map scanning is as follows: Extract the feature representation of the current sector, including the sector azimuth angle. Distance value Reflection intensity and the credibility of regular echoes Encode it to construct a query vector From the geometrically stable layer The matching window reads the stable distance values ​​and intensity statistics of the sector during historical stable matches, from the unstable echo suspicious layer. Historical anomaly statistics for the region were read from the spatial memory record, and the above data was encoded into a key vector. Data is read from the corresponding historical matching scores or anomaly weights and encoded to construct a value vector. Calculate the cross-attention weights:

[0081] right The context vector is obtained by weighted summation and then mapped to the map matching score correction value through a fully connected layer. For cases that are historically identified as interference areas and are not of the perforated facade type, this correction value is appropriately reduced, thereby decreasing the weight of echoes from that area in the final fusion.

[0082] At this point, all the correction values ​​used for weighted fusion have been generated. This includes the regular echo confidence level calculated in step S103. Based on the spatial memory attention mechanism and the unstable echo suspicious layer Spatial memory attention weights are calculated from historical statistical information. The echo confidence correction value is obtained by inter-frame self-attention mechanism through self-attention calculation and mapping of fully connected layers. The map matching score correction value is obtained through cross-attention calculation and fully connected layer mapping by the cross-attention mechanism of the scanned map. .

[0083] Ultimately, the credibility of the echo The weighted fusion yielded:

[0084] in To pre-set the weighting coefficients, this embodiment preferably uses... , , This allows rule branches to dominate the fusion process, while the attention mechanism provides effective correction. As a minimum confidence level, it ensures that echoes with extremely low confidence levels still retain limited optimization participation rights; This is a truncation function that restricts the output value to a certain range. Within the range. When commanding. At that time, the method degenerates into a pure rule-based method.

[0085] S106: Determine the echo type of each sector based on the hole facade identifier and the rule echo confidence level, and adopt differentiated residual block establishment strategy and weight configuration strategy for different echo types; the echo types include at least stable type, hole facade type and abnormal type.

[0086] In a specific implementation, this step determines each sector as one of three echo types based on the hole facade identification and the reliability of the regular echo, which is used to determine the subsequent residual block building strategy and weight configuration.

[0087] The decision logic is as follows: If and Higher than the preset confidence threshold ,in In this embodiment, a value of 0.7 is preferred, which is then classified as a stable type. .

[0088] This type indicates that the sector echo is reliable and does not exhibit holed double-cluster characteristics, allowing for normal laser matching residual block establishment and the use of a normal residual information matrix. It participates in the optimization as a weight.

[0089] like Furthermore, if the motion compensation is consistent, meaning the kinematic consistency verification has been passed in step S104, then it is determined to be a perforated facade type. .

[0090] This type indicates that although the sector exhibits bi-cluster jumping with a perforated plate structure, it does so regularly and constitutes a usable and effective constraint. In this case, the near-end cluster center value is taken. As a representative distance:

[0091] Preserve laser-matched residual blocks and use medium-to-low weighted residual information matrices And Huber robust loss function. The lower bound of the echo confidence for this type is set to a preset value. To ensure the existence of residual blocks, regardless of spatial memory attention weights None of them must be removed entirely. This design ensures that the lateral constraints of the server rack surface are not accidentally deleted in cold aisle scenarios.

[0092] If neither the stable type condition nor the hole facade type condition is met, it is judged as an abnormal type, that is... .

[0093] This type indicates that the echo in this sector is unreliable and does not conform to the regular double-cluster pattern of the aperture facade. In this case, no laser matching residual block will be established, or only the minimum residual information matrix will be used. As a weight, the impact of this sector on the optimization result is minimized.

[0094] S107: Selectively establish laser matching residual blocks and set corresponding weight coefficients according to the echo type and final echo confidence. Combine the established laser matching residual blocks with the odometry residual term to form an optimization objective function for solving, and obtain the robot's optimal pose and residual information for each sector.

[0095] In a specific implementation, the echo type and final echo confidence determined in the preceding steps are applied to the construction and optimization of the laser matching residual block.

[0096] For the Each sector, based on its echo type and the credibility of the final echo Set the residual information matrix :

[0097] in To preset the weight scaling index, this embodiment preferably uses... This allows residual blocks with higher reliability to receive higher optimization weights. When Time weighting ,when Time weighting .

[0098] From the geometrically stable layer Read the map distance reference value of the corresponding grid cell for this sector. , will represent distance Register with the map reference value and construct the residual function:

[0099] in Indicates the first The representative azimuth angle of each sector is transformed by pose. Project it onto the corresponding position in the map coordinate system.

[0100] The selected laser-matched residual blocks and the odometer residual blocks together constitute a nonlinear least squares problem:

[0101] in Let odometer residual function be used. is the Huber robust loss function with a threshold of 1.0, used to suppress the impact of outliers in a single frame on the optimization results.

[0102] The optimization problem uses the initial pose estimate from step S101. Using the initial values, the Levenberg-Marquardt algorithm is used iteratively with the Ceres solver. When the iteration step size is less than... Alternatively, convergence occurs when the number of iterations exceeds 100, yielding the optimal robot pose. At the same time, the residual information of each sector is recorded for subsequent anomaly detection.

[0103] S108: Based on the robot's optimal pose, the residual information of each sector, and the initial pose estimate, detect whether an abnormal mode is triggered. If triggered, remove the laser matching residual blocks of the corresponding suspicious sectors and adjust the weight coefficients of the remaining residual blocks. Use the initial pose estimate as the initial value to re-optimize and update the robot's optimal pose.

[0104] In a specific implementation, based on the robot's optimal pose and residual information for each sector output in step S107, and combined with the initial pose estimation value in step S101, a comprehensive detection is performed to determine whether an abnormal mode is triggered. The specific criteria include the following four items.

[0105] First, the robot's optimal pose. With initial pose estimate pose deviation between Compared with a preset pose deviation threshold, where For difference operators on Lie algebras, This corresponds to the norm. If the translation deviation exceeds 0.3m or the rotation deviation exceeds 5°, an abnormal mode is triggered.

[0106] Second, the total cost of laser matching The total cost is compared with a preset cost threshold. If the total cost exceeds three times the average cost during normal operation, an abnormal mode is triggered.

[0107] Third, the residuals of each sector and the suspected unstable echo layers are compared. The location of historical anomaly regions is compared. If points with residuals greater than 3 times the standard deviation are concentrated within historical anomaly regions, it indicates that a known unreliable echo pattern is being repeated, and an anomaly pattern is triggered.

[0108] Fourth, classify exception types The percentage of sectors is compared with a preset anomaly ratio threshold. If the anomaly ratio exceeds this threshold (preferably 30% in this embodiment), it indicates that the current frame is subject to large-area interference, triggering an anomaly mode.

[0109] An abnormal mode is triggered when any one of the above four criteria is met.

[0110] After triggering the abnormal mode, the re-optimization process is executed: Sectors with residuals exceeding a preset residual threshold are marked as suspicious sectors and removed from the optimization problem. The residual threshold is set to three times the standard deviation. The residual information matrix of the remaining laser matching residual blocks is multiplied by a shrinkage factor. ,in This is the preferred embodiment. ; with initial pose estimate Re-execute step S107 of iterative optimization with the initial value to update the robot's optimal pose. The location and triggering conditions of this anomaly will be recorded in the unstable echo suspicion layer. This is for use in subsequent inspections to retrieve spatial memory and attention information.

[0111] S109: Update the geometric features to the double-layer map according to the echo type of each sector, and update the statistical information in this observation to the double-layer map at the same time.

[0112] In a specific implementation, the purpose of this step is to update the observation results of the current frame to the two-layer map.

[0113] For geometrically stable layers The update conditions are checked sector by sector. If the range variance of the sector in the past multiple consecutive frames is... The variance is below a preset threshold; in this embodiment, the threshold is preferably 0.01, i.e., the ranging standard deviation is less than 10 cm, and the echo type... If the data does not belong to the abnormal type, the geometric features of the sector are determined to be stable and reliable, and the geometric feature points of the sector in the current frame are stored in the geometric stabilization layer. The geometrically stable layer serves as a baseline map for subsequent scanning and matching, storing only laser points that have passed stability testing.

[0114] For unstable echo suspected layer Update the statistics cell by cell. For each cell... Update the number of times the inspection has been conducted. .

[0115] If the raster contains sectors with an echo type of exception, that is... Then update the historical anomaly count. .

[0116] Simultaneously update the historical proportions of stable, porous facade, and anomalous types within the grid, and update the statistics of mean and variance of reflection intensity. Unstable echo suspected layer. It does not participate in the calculation of the scan matching score, but is only used for querying the spatial memory attention mechanism in step S105 and for reference in conflict detection in step S108.

[0117] This concludes the complete processing flow for the current frame. The robot continues moving to scan the next frame. The dual-layer map continuously accumulates and updates environmental knowledge during long-term inspections, forming a continuous optimization cycle of storage, retrieval, and updating. As the number of inspections increases, unstable echoes become suspicious. With increasingly rich statistical information, the prior confidence weights of spatial memory attention mechanisms are becoming more accurate, and the localization accuracy and robustness of the method are gradually improving over long-term operation.

[0118] In this embodiment, ranging variance, intensity variance, inter-frame intensity jump, motion compensation consistency, and ranging are extracted by constructing a cross-frame temporal sliding window. The system employs a joint intensity discrimination feature, utilizing three attention mechanisms—spatial memory attention, inter-frame self-attention, and scan map cross-attention—to perform multi-level intelligent fusion correction of the reliability of regular echoes, achieving a refined assessment of the reliability of echoes from each laser sector. Simultaneously, through dual-cluster distribution detection and kinematic consistency verification, it effectively distinguishes between the regular dual-cluster jumping of hole facades and random interference such as glass multipath. Differentiated residual information matrix setting strategies are adopted for stable types, hole facade types, and abnormal types, ensuring that stable echoes receive high weight for optimization, hole facade echoes are forcibly retained with reduced weight to retain effective lateral constraints, and abnormal echoes are suppressed or eliminated. Combined with the dual-layer map separation and maintenance of the geometrically stable layer and the unstable echo suspicion layer, as well as the re-optimization mechanism after anomaly triggering, the system significantly improves the accuracy and robustness of laser positioning for inspection robots in complex data center environments.

[0119] Furthermore, as a refinement and extension of the specific implementation methods described above, and to fully illustrate the specific implementation process in this embodiment, another robot laser localization method based on temporal echo reliability and attention fusion is provided. This method addresses the following typical data center scenarios: Densely perforated cabinet surface: The ranging shows a double-cluster jump, large variance but structural regularity, which is judged to be a perforated facade type, and the horizontal constraint of scanning matching is retained; Non-perforated areas of black-painted cabinets: Intensity is low, and the residuals of ranging and motion compensation during grazing are abnormal, which can be distinguished based on the results of dual-cluster analysis. Glass doors or partitions: Random and erratic fluctuations in distance and intensity are identified as anomalies; residual blocks should be removed or minimum weighting should be applied. ; Metal kickboards or cabinet door frames: Distance measurement is relatively stable, judged as a stable type, and will participate in matching normally; Parallel rack column geometric repetition: Matching is ambiguous, utilizing... Statistical and spatial memory attention mechanisms aid in judgment.

[0120] like Figure 2 As shown, the method includes the following steps: S1: Laser scanning and data acquisition.

[0121] The robot performs laser scanning and simultaneously acquires odometry or inertial measurement unit data to provide an initial estimate of the positioning attitude. Each frame scan yields a point set. Including azimuth Distance measurement and reflection intensity A single-line lidar provides only a constraint in the pitch plane; pose optimization is performed using a special Euclidean group in a two-dimensional plane. Mainly.

[0122] S2: Temporal feature extraction.

[0123] For each laser azimuth angle Sector, maintenance length is A sliding window records the past activity of that sector. Frame ranging sequence Intensity sequence and motion-compensated residual sequence Motion compensation residual It means The predicted distance of this sector after odometer transformation at that time is Real distance measurement at any time The difference reflects the consistency of laser matching.

[0124] S3: Calculation of the credibility of regular echoes.

[0125] For each laser azimuth angle For each sector, based on the time series sequence extracted in step 2, the following regular features are calculated: Distance variance Defined as a ranging sequence in the past Intra-frame variance, i.e. ,in for Frame ranging mean. This feature reflects the temporal stability of target distance: multi-frame jump anomalies such as ghost points will lead to a significant increase in variance; the ranging variance of stable geometric points (such as the ground and walls) is smaller.

[0126] Strength variance Defined as an intensity sequence The variance, i.e. ,in for Frame intensity mean. Glass or multipath interference can cause drastic fluctuations in intensity signals, resulting in a large variance; the intensity variance of stable object surfaces is usually small.

[0127] Inter-frame intensity jump Defined as the average of the absolute values ​​of intensity changes between adjacent frames, i.e. This feature is sensitive to single-frame specular flashes: instantaneous intensity jumps caused by reflections from the glass surface will be clearly reflected in this feature.

[0128] Motion compensation consistency Defined as motion-compensated residual sequence The consistency metric is calculated as follows: ,in The standard deviation of the residuals. This is a preset threshold. When the standard deviation of the residual sequence is small and the mean is close to zero, this value is close to 1, indicating that the echo is reliable; when the residuals fluctuate wildly, this value is small.

[0129] Joint characteristics of ranging intensity Defined as the ratio of distance variance to intensity variance, i.e. This is used to identify specific interference patterns. Patterns with large ranging variance and large intensity variance are commonly seen in glass multipath interference; patterns with large ranging variance but low intensity or large fluctuations are commonly seen in black cabinet grazing interference or multipath interference.

[0130] The above five rule features are combined into a feature vector: Through activation function (Through a nonlinear mapping function) mapped to The interval is used to obtain the confidence level of the regular echo:

[0131] in This indicates that the echo from that sector is completely reliable. This indicates that the information is completely unreliable. It should be noted that the reliability of the rule-based echo is only used as a preliminary basis for judgment, and the final echo type needs to be determined in conjunction with the hole pattern detection results in step S4.

[0132] S4: Hole detection.

[0133] Data center server racks typically have ventilation mesh or perforated panels on their front doors. When a single-line laser scans one side of the rack in the cold aisle, the same azimuth angle... Distance value The distance will fluctuate between two main clusters: one for short distances (cabinet doors) and the other for long distances (deep within the holes or inside the cabinet). This fluctuation in the distance measurement between the two clusters is a normal physical phenomenon in perforated structures and should not be confused with the random ghost distance of the glass. If areas with large distance measurement fluctuations are simply judged as unstable without establishing scan-matching residual blocks, only the ground and a few laser points on the opposite side will often remain in the cold aisle, resulting in insufficient effective constraints and rendering lateral positioning unusable.

[0134] Therefore, this step performs bi-cluster clustering analysis on the ranging sequences, and the specific method is as follows: The first step is to use the K-Means clustering algorithm (assuming...) ) for ranging sequence Cluster all values ​​and calculate the centroids of the two principal clusters, denoted as the nearest clusters. (Distance between cabinet doors) and long-distance clusters (Distance deep inside the hole).

[0135] The second step is to calculate the sum of the proportions of the two main clusters, i.e. ,in and These represent the probabilities or proportions of data points belonging to the nearest and farthest clusters, respectively. If... ( If the preset threshold is 0.6, then a dual-cluster pattern is considered to exist.

[0136] The third step is to examine the consistency characteristics of motion compensation. If the motion compensation residual in the dual-cluster mode still satisfies the channel geometry constraints during robot translation, i.e., the overall distance changes monotonically with mileage rather than lateral jumps, then it is determined to be a candidate type of perforated facade. This is because the dual-cluster jumps of the perforated facade are caused by the perforated structure. When the robot moves along the channel, the appearance of near and far clusters is regular and there will be no lateral jumps that contradict the robot's motion trajectory.

[0137] S5: Fusion of three attention mechanisms.

[0138] Three attention mechanisms are used to intelligently fuse and correct the credibility of rule echoes, thereby enabling the assessment of echo reliability.

[0139] Spatial memory and attention This module maintains a spatial memory table for cross-inspection, dividing the environment into grids. Or a chain-like route, storing the number of times the route has been inspected for each grid cell. (Total number of times the grid has been traversed), number of historical anomalies (The cumulative number of times an echo was identified as an anomalous echo within the grid), the distribution ratio of historical echo types (the historical proportions of stable types, void facade types, and anomalous types within the grid), and the historical representative distance. (Median of historical ranging representative values), mean confidence level of the previous rule echo Wait. When the robot's current pose falls into a certain grid... The specific attention calculation process is as follows: First, construct the query vector. This vector is encoded by the current position (via raster coordinates). The data includes embedded information, channel identifiers (used to distinguish different channel types, such as cold channel and hot channel), and the current rule echo confidence level. The summary information is pieced together. Then, the raster and its neighborhood (such as...) are extracted from the spatial memory table. The history of the neighborhood grid as key-value pairs , where key Includes historical location coding, historical echo type coding (one-hot vector), and history. Encoding, value Includes historical anomaly weights ( ) and echo reliability statistics (historical) (mean and variance). Next, calculate the attention weights: ,in Let be the dimension of the key vector. This is used for scaling to ensure training stability. Finally, the spatial memory weights are obtained through attention-weighted summation: , It is a scalar, with a range of values. . The physical meaning is: if a certain grid cell has a high number of inspection passes and a high number of historical anomalies, then... A low output value indicates a cautious approach to the echoes in that area, suggesting an early reduction in residual weights or preparation for removing residual blocks; if a raster's historical type is predominantly perforated facades, then... Maintain the perforated facade strategy to prevent accidental deletion of valid constraints due to repeated visits; if a grid is being visited for the first time, then... Keep it neutral (close to 0.5).

[0140] Inter-frame self-attention This module analyzes the current continuous... The relationship between frame sector features enables the identification of different echo types in the current observation. The specific implementation is as follows: First, construct a feature sequence. This involves combining the current frame with the features of its preceding and following frames... The frame composition length is The temporal window extracts the feature vector of each sector in each frame, including: ranging ,strength Motion compensation residuals Confidence of regular echoes Echo type (Type) (encoded as a one-hot vector, representing stable / holey facade / anomaly). These features are concatenated into a fixed-dimensional vector representation. Then, self-attention calculation is performed. Let the input sequence be... The query matrix is ​​obtained through three linear transformations. Key matrix Sum matrix ,in This is a learnable weight matrix. The attention weights are calculated as follows: The output of self-attention is: They can be stacked. A self-attention layer is added to enhance expressive power. Finally, the output representation is extracted and correction values ​​are calculated. The output of the self-attention layer is then processed. The echo confidence correction value for each sector is obtained through mapping using a fully connected layer. Meanwhile, the map matching probability is output through another branch. The key capability of this module is its ability to distinguish between different echo types in a single observation: for example, a specular flash on a glass surface may appear as an intensity within a single frame. The change is transient (affecting only adjacent frames) rather than a persistent hole facade problem (maintaining a dual-cluster mode for multiple consecutive frames) can be identified through inter-frame self-attention.

[0141] Scan map cross attention This module cross-validates the current laser frame with map information, using historical matching information from the map to help determine the reliability of the current echo. The specific implementation is as follows: query vector. Feature representation from the current sector (including sector azimuth) Distance measurement ,strength Confidence of regular echoes (etc.); key vector From geometrically stable layer The matching window (which stores features of stable matches in the sector's history, including stable distance values ​​and intensity statistics) and the unstable echo suspicion layer. Spatial memory records (store historical anomaly statistics for the area, including the location and type of anomalies); value vectors These are derived from the corresponding historical matching scores or anomaly weights. The attention weights are calculated as follows: Then, the context vector is obtained by attention-weighted summation: Finally, the context vector is mapped to the matching score correction value through a fully connected layer. This module appropriately lowers the matching score for interference areas in memory that are not perforated facade types.

[0142] Fusion of Three Attention Methods: Outputs of the Three Attention Methods and the Reliability of Rule Echoes Weighted fusion is performed to obtain the final echo confidence level. Used for subsequent laser matching residual block scheduling:

[0143] in , , The weighting coefficients can be adjusted through configuration; rule branches can run independently (let...). At this point, the system degenerates into a pure rule-based method; The function restricts the result to interval, This represents the preset minimum level of confidence.

[0144] S6: Echo type identification and processing strategy.

[0145] Based on the rule features and attention output, each sector is classified into three types: Stable type: Does not conform to the hole dual-cluster mode and has regular echo confidence. Above the threshold Normal laser matching residual blocks are established, using the normal residual information matrix. Weights.

[0146] Perforated facade type: The ranging sequence is distributed in two clusters and is consistent with the robot's motion, preserving the laser matching residual block. Representative distance is used. Participate in matching (take the nearest cluster center value of the ranging sequence) ), residual information matrix Take low to medium weight The Huber robust kernel function is used. The lower bound of the echo confidence level for the perforated facade type is the lowest confidence level for the perforated facade. This ensures the existence of residual blocks. Specifically, for sectors with perforated facades, regardless of spatial memory weights... Regardless of the circumstances, the perforated facade strategy should be adopted, and no part of the design should be rejected simply because the reliability of the regular echo is low.

[0147] Anomaly types: large ranging variance but not conforming to the hole dual-cluster pattern, or severely inconsistent with map matching residuals; update unstable echo suspected layer. Records indicate that no laser matching residual block is established for this sector, or the minimum residual information matrix is ​​used. Weights.

[0148] S7: Construction and optimization of laser matching residual blocks.

[0149] Select the representative distance With geometrically stable layer Register the corresponding raster and construct the residual function. And set the corresponding residual information matrix according to the echo type:

[0150] The laser matching residual block and the odometry residual block are combined to form a nonlinear least squares problem. Robust loss functions such as Huber are used to suppress the influence of outliers in a single frame, and the optimal pose of the robot is solved. :

[0151] S8: Echo anomaly mode handling.

[0152] Check whether an abnormal mode has been triggered. Criteria include: pose. With odometer prediction Does the deviation exceed the threshold? Is the total cost of laser matching abnormal? Are the residuals concentrated in certain areas? Mark the region; check if the unstable echo ratio exceeds the limit, etc. If an anomaly is detected, perform the following processing: remove the laser matching residual block of the suspicious sector; shrink the weight of the remaining residual block; and perform odometry pose estimation. Resolve for initial values; update the unstable echo suspected layer. Record.

[0153] S9: Two-layer map update and output.

[0154] This invention employs a two-layer map structure to manage environmental information, wherein the geometrically stable layer... Used to store verified geometric feature points as a baseline map for scan matching, only laser points that pass the stability test can update the geometric stability layer. Unstable echo suspected layer Statistical information used to record unstable echoes, including the mean and variance of the intensity of each grid cell, and the number of times the inspection has passed. Number of historical anomalies Echo type distribution ratio, etc. It is not involved in the scan matching score calculation, but is only used for conflict detection and spatial memory attention. Search.

[0155] This step outputs the optimized robot pose. Update the two-layer map. Geometric stability layer. Updated to the geometric features verified in the current frame; unstable echo suspected layer. Updated to include statistics on various unstable echoes detected this time (including the number of times the inspection passed by). Number of historical anomalies (echo type distribution ratio, etc.) for use in subsequent inspections to query spatial memory and attention.

[0156] The above description of the disclosed embodiments enables those skilled in the art to make or use the invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the invention. Therefore, the invention is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A robot laser localization method that fuses temporal echo reliability and attention, characterized in that, Includes the following steps: S101: The robot uses a single-line lidar to scan and obtain an initial point set, while also acquiring odometry or inertial measurement unit data to provide an initial estimate of the positioning attitude; each frame of the initial point set includes azimuth, range, and reflection intensity. S102: Maintain a sliding window of fixed length for each laser azimuth sector, record the ranging sequence and reflection intensity sequence of the sector in the past multiple consecutive frames, and calculate the motion compensation residual sequence between frames based on the initial pose estimate. S103: Calculate the temporal ranging fluctuation characteristics based on the ranging sequence, calculate the temporal intensity fluctuation characteristics and inter-frame intensity abrupt change characteristics based on the reflection intensity sequence, calculate the motion compensation consistency characteristics based on the motion compensation residual sequence, and calculate the ranging-intensity joint discrimination characteristics based on the temporal ranging fluctuation characteristics and the temporal intensity fluctuation characteristics. Combine all the calculated features into a regular feature vector and obtain the regular echo confidence through nonlinear mapping. S104: Perform dual-cluster distribution detection on the ranging sequence to determine whether there is a dual-cluster pattern caused by the hole structure, and determine whether the dual-cluster pattern satisfies the kinematic consistency constraint based on the motion compensation residual sequence. If it does, determine that the sector is a candidate type of hole facade and generate a hole facade identifier. At the same time, record the near-end cluster center value obtained by dual-cluster detection. S105: Read relevant data from the pre-built two-layer map, and use spatial memory attention mechanism, inter-frame self-attention mechanism and scan map cross-attention mechanism to fuse and correct the credibility of the rule echo to obtain the final echo credibility; S106: Determine the echo type of each sector based on the hole facade identifier and the rule echo confidence level, and adopt differentiated residual block establishment strategy and weight configuration strategy for different echo types; the echo types include at least stable type, hole facade type and abnormal type; S107: Selectively establish laser matching residual blocks and set corresponding weight coefficients according to the echo type and final echo confidence. Combine the established laser matching residual blocks with the odometry residual term to form an optimization objective function for solving, and obtain the robot's optimal pose and residual information of each sector. S108: Based on the robot's optimal pose, the residual information of each sector, and the initial pose estimate, detect whether an abnormal mode is triggered. If triggered, remove the laser matching residual block of the corresponding suspicious sector and adjust the weight coefficient of the remaining residual block. Use the initial pose estimate as the initial value to re-optimize and update the robot's optimal pose. S109: Update the geometric features to the double-layer map according to the echo type of each sector, and update the statistical information in this observation to the double-layer map at the same time.

2. The robot laser localization method based on temporal echo reliability and attention fusion according to claim 1, characterized in that, Step S101 includes: The robot uses a single-line lidar to continuously scan during the inspection process, obtaining a set of laser points per frame. Each laser point includes an azimuth angle. Distance value and reflection intensity , forming the original point set azimuth angle Indicates the first The scanning angle of each laser point in the polar coordinate system of the lidar, and the ranging value. This indicates the measured distance and reflection intensity corresponding to the laser point. This indicates the amplitude of the echo signal received by the laser point; Simultaneously, the robot's relative motion information is acquired through a wheeled odometry system or an inertial measurement unit, and the initial estimated value of the pose at the current moment is calculated by combining it with the pose at the previous moment. , representing the initial pose estimation of the robot in a two-dimensional plane.

3. The robot laser localization method based on temporal echo reliability and attention fusion according to claim 2, characterized in that, Step S102 includes: The maintenance length for each laser azimuth sector is A sliding window records the sector in the most recent Ranging sequence in the frame and reflection intensity sequence ; Based on the initial pose estimate, calculate the motion compensation residual sequence between each frame. subscript Indicates the frame number; Motion Compensation Residual The calculation formula is: in, This indicates the use of the initial pose estimate. The relative change between adjacent frames will The distance measurement value of this sector at that time Projected onto the odometer The predicted distance is obtained in the robot's coordinate system at that time.

4. The robot laser localization method based on temporal echo reliability and attention fusion according to claim 3, characterized in that, Step S103 includes: Based on ranging sequence The distance measurement variance is calculated using the following formula. As a characteristic of time-series ranging fluctuations: in, for Frame ranging mean; Based on reflection intensity sequence The strength variance is calculated using the following formula. As a characteristic of temporal intensity fluctuation: in, for Mean frame reflection intensity; The inter-frame intensity jump is calculated using the following formula. As a feature of inter-frame intensity abrupt change: Based on motion-compensated residual sequence The motion compensation consistency feature is calculated using the following formula: in, Let be the standard deviation of the motion compensation residual sequence. To preset the standard deviation threshold of motion compensation residuals, The range of values ​​is ; The joint feature of ranging intensity is calculated using the following formula: in, To prevent small quantities from being divided by zero, Used to distinguish different interference modes; The above five features are combined into the rule feature vector: Through nonlinear activation function The regular echo confidence level is obtained by performing nonlinear mapping. ; in This is the current laser azimuth sector identifier. This represents a preliminary assessment of the sector's echo reliability based solely on timing rule characteristics.

5. The robot laser localization method based on temporal echo reliability and attention fusion according to claim 4, characterized in that, Step S104 includes: Clustering algorithm is used to analyze the ranging sequence. Clustering is performed on all values ​​to obtain the proximal cluster center values. and remote cluster center value The proximal cluster center values ​​correspond to the center values ​​of clusters with smaller distances, and the distal cluster center values ​​correspond to the center values ​​of clusters with larger distances. The sum of the proportions of the two main clusters is calculated using the following formula: in and These represent the proportions of data points belonging to the near-end cluster and the far-end cluster, respectively. This indicates the proportion of bi-clustered data to the total data points; if If so, it is determined that a dual-cluster pattern exists, where The preset threshold for the proportion of dual clusters; According to the motion compensation residual sequence If the overall change in distance measurement in this sector during the robot's translation process is monotonically altered with distance traveled, and the kinematic consistency constraint is satisfied, it is determined to be a candidate type of hole facade and the hole facade identifier is generated. ; This indicates that the sector is marked as a candidate type for a perforated facade. This indicates that it has not been marked.

6. The robot laser localization method based on temporal echo reliability and attention fusion according to claim 5, characterized in that, Step S105 includes: Geometric stability layer from a pre-built two-layer map Read geometric reference data from the middle and from the unstable echo suspected layer The system reads historical statistics, including the number of times the inspection passed through. Number of historical anomalies The distribution ratio of historical echo types, the median of historical representative distances, and the mean confidence level of the previous regular echo; The execution flow of the spatial memory attention mechanism is as follows: Divide the environment into grids It then concatenates summary information on the current location encoding, channel identifier, and current rule echo reliability to generate a query vector. ; From the suspected layer of unstable echo Read the historical statistical information of the current grid and its neighborhood; Spatial memory attention weights are calculated using the following formula. : in, The key matrix is ​​formed by encoding the historical statistical information of the current grid and its neighborhood. The dimension of the key vector; Spatial memory attention weights are obtained by weighted summation of attention values. ;in The value matrix is ​​formed by encoding the historical statistical information of the current grid and its neighborhood. It is a scalar, and its value range is... This indicates that the spatial memory attention mechanism makes a priori assessment of the credibility of the current raster echo based on historical statistical information; The execution flow of the inter-frame self-attention mechanism is as follows: The current frame and all frames before and after it. Frames form a timing window, in which This is the preset number of frames per side; Extract the ranging value of each sector in each frame Reflection intensity Motion compensation residuals Confidence of regular echoes The echo type encoding is concatenated to form a feature vector. and form the input sequence. ;in This is the length of the timing window; The query matrix is ​​obtained through three linear transformations. Key matrix Sum matrix ;in The weight matrix is ​​a learnable weight matrix; The inter-frame self-attention weights are calculated using the following formula. : Using the attention weights For the value matrix We perform a weighted summation to obtain the self-attention weighted feature representation matrix. ; The feature representation matrix Stacking After the attention layer, the echo confidence correction value for each sector is mapped through a fully connected layer. ; The execution flow of the cross-attention mechanism for the scanned map is as follows: Extract the feature representation of the current sector, including the sector azimuth angle. Distance value Reflection intensity and the credibility of regular echoes Encode it to construct a query vector From the geometrically stable layer The matching window reads the stable distance values ​​and intensity statistics of the sector during historical stable matches, and from the suspected unstable echo layer Historical anomaly statistics for the region were read from the spatial memory record, and the above data was encoded to construct a key vector. Read data from the corresponding historical matching scores or anomaly weights and encode it to construct a value vector. ; The cross-attention weights of the scanned map are calculated using the following formula: Using the attention weights For the value vector We perform a weighted summation to obtain a context vector that integrates historical map information, and then map it through a fully connected layer to the map consistency correction value for each sector. ; The final echo confidence level is generated using the following formula. : in To preset the weighting coefficients, To preset a minimum confidence level, This is a truncation function that restricts the output value to a certain range. Within the range.

7. The robot laser localization method based on temporal echo reliability and attention fusion according to claim 6, characterized in that, Step S106 includes: like and Higher than the preset confidence threshold If it is, then it is determined to be of the stable type and denoted as the echo type. At this point, the laser matching residual block is established normally and the normal residual information matrix is ​​used. As weight; like If the motion compensation is consistent, then it is determined to be the aforementioned hole facade type and recorded as the echo type. At this point, the proximal cluster center value is taken. As the representative distance Preserve laser matching residual blocks and use medium- and low-weight residual information matrices. Huber robust loss function; the lower limit of echo confidence for the perforated facade type is taken as the preset minimum confidence threshold for the perforated facade type. ; If neither the conditions for the stable type nor the conditions for the perforated facade type are met, then it is judged to be the abnormal type and recorded as the echo type. At this point, either laser matching residual blocks are not established or the minimum residual information matrix is ​​used. As weights.

8. The robot laser localization method based on temporal echo reliability and attention fusion according to claim 7, characterized in that, Step S107 includes: According to the echo type and the credibility of the final echo Set the residual information matrix subscript For sector index, Indicates the first Residual information matrix of each sector: when hour, ;in The preset weight scaling index; when hour, ; when hour, ; From the geometrically stable layer Read the corresponding raster map distance value; register the representative distance with the map distance value of the corresponding raster in the geometrically stable layer to construct a residual function. ; The laser-matched residual block and the odometer residual block are combined to form the following nonlinear least squares problem: in Let odometer residual function be used. Huber robust loss function; With the initial pose estimate The robot's optimal pose is obtained by iteratively optimizing the initial value. Simultaneously, it records the residual information of each sector.

9. The robot laser localization method based on temporal echo reliability and attention fusion according to claim 8, characterized in that, Step S108 includes: The optimal pose of the robot With the initial pose estimate The pose deviation between the two is compared with a preset pose deviation threshold. If it exceeds the preset pose deviation threshold, an abnormal mode is triggered. Total cost of laser matching Compare with a preset cost threshold; if the cost exceeds the preset cost threshold, an abnormal mode is triggered. The residuals of each sector and the suspected unstable echo layers are compared. The location of historical abnormal areas is compared. If high residual points are concentrated in the historical abnormal areas, an abnormal mode is triggered. exception types The percentage of sectors is compared with a preset abnormal percentage threshold. If the percentage exceeds the preset abnormal percentage threshold, an abnormal mode is triggered. After triggering the abnormal mode, sectors with residuals higher than a preset residual threshold are marked as suspicious sectors and removed from the optimization problem. The residual information matrix of the remaining laser matching residual blocks is multiplied by a shrinkage factor. ,in The preset weighting contraction coefficient is used, with the initial pose estimate as the initial value. Re-execute step S107 for iterative optimization to update the robot's optimal pose using the initial values. And update the unstable echo suspicious layer. Record.

10. The robot laser localization method based on temporal echo reliability and attention fusion according to claim 9, characterized in that, Step S109 includes: For each sector, if the ranging variance of that sector over the past multiple consecutive frames... Below the variance threshold and the echo type If the anomaly does not belong to the aforementioned anomaly type, then the geometric feature points of the current frame for that sector are stored in the geometric stabilization layer. ; For each grid Update the number of times the inspection has been completed. If the echo type Then update the historical anomaly occurrence count. And update the echo type distribution ratio; the unstable echo suspected layer It does not participate in the calculation of scan matching scores, but is only used for attention mechanism queries and conflict detection.

Images