A robot autonomous charging and fault-tolerant control method and robot

Abstract

This application provides a method and robot for autonomous charging and fault-tolerant control. The method monitors the robot's energy consumption data in real time and locks onto the target charging coordinates; when approaching a charging station, it uses a visual docking operator to fine-tune the pose and generate an alignment trajectory with compensation attributes; and it monitors the health status of sensors through a perception and obstacle-stopping unit to determine a fault-tolerant compensation strategy vector and inject it into the trajectory. This application addresses the shortcomings of traditional solutions where docking failures are caused by pose deviations and sensor anomalies. By introducing visual fine-tuning compensation and link fault recovery mechanisms, it improves the reliability of charging docking. Compared to traditional methods, it can maintain a high docking success rate even when sensors are in a sub-healthy state, thus improving the robot's autonomous endurance.

Description

Technical Field

[0001] This application relates to the field of robot autonomous control technology, and more specifically, to a robot autonomous charging and fault-tolerant control method and a robot. Background Technology

[0002] With the widespread application of service robots in gardens, scenic areas, and industrial parks, their autonomous endurance has become crucial for enhancing their ability to perform long-duration, high-intensity tasks, especially in areas such as inspection, cleaning, and unmanned delivery, where automatic recharging is indispensable. To achieve energy replenishment, existing robot charging solutions typically employ a single method of far-field infrared guidance and near-field mechanical alignment. This approach first uses an infrared array to obtain the approximate location of the charging station; then, it drives the robot chassis to move towards the charging station; finally, a collision sensor detects physical contact to initiate the charging process.

[0003] However, this existing charging solution has significant technical flaws. Due to the complex environment and dynamic interference in the charging scenario, a single guidance method is difficult to adapt to changes in lighting or occlusion, which can easily lead to pose deviations during near-field docking that exceed mechanical tolerance. At the same time, the existing alignment process lacks real-time monitoring of the health status of the sensor data link. Once the sensor experiences sub-health issues such as increased sampling delay or decreased signal-to-noise ratio, the system often cannot make a timely fault-tolerant response, resulting in inaccurate docking actions or even triggering unexpected obstacle-stopping logic, which seriously reduces the reliability of the robot's autonomous charging. Summary of the Invention

[0004] To address the aforementioned technical problems, this application provides a robot autonomous charging and fault-tolerant control method and a robot, thereby at least alleviating the aforementioned technical problems.

[0005] A method for autonomous charging and fault-tolerant control of a robot includes: Step 1: Monitor the robot's energy consumption status data in real time. In response to the recognition of the charging request signal triggered by the energy consumption status data, use the relocation algorithm to lock the robot's pose in the preset global point cloud map, and perform charging pile reference position mapping processing to determine the target charging coordinates. Step 2: Plan the motion path between the robot's current pose and the target charging coordinates, and use the visual docking operator to perform pose fine-tuning when the robot approaches the target charging coordinates to a preset distance, so as to generate an autonomous alignment trajectory with spatial vector compensation attributes. Step 3: Use the obstacle sensing unit to monitor the health status of the sensors during the charging alignment action in real time, so as to generate the identification link abnormal component and determine the fault tolerance and recovery strategy vector. Then, inject the fault tolerance and recovery strategy vector into the autonomous alignment trajectory to trigger the charging circuit control signal after the physical touch lock is completed.

[0006] Optionally, identifying the charging request signal in step 1 includes: Extract the instantaneous voltage value and load current component from the energy consumption status data, and generate the remaining range entropy value by calculating the correlation distribution of the instantaneous voltage value and load current component; In response to the remaining range entropy value falling below a preset safety threshold, a charging task priority preemption operator is triggered to generate a charging request signal pointing to the target charging coordinates.

[0007] Optionally, robot pose locking in step 1 includes: The laser SLAM operator is used to capture the point cloud features of the current environment and then perform multi-scale feature matching with a pre-set global point cloud map to obtain the robot's initial pose in the global coordinate system. Inertial navigation data is retrieved to perform zero-bias drift compensation on the initial pose, thereby generating a robot pose locking result with coordinate reference attributes and performing robot pose locking accordingly.

[0008] Optionally, the pose fine-tuning process in step 2 includes: A visual docking operator is used to capture auxiliary positioning marks on the surface of the charging pile, and geometric center features in the image plane are extracted from the auxiliary positioning marks. The deviation angle and distance vector of the geometric center feature relative to the robot's visual center axis are calculated to generate pose correction commands that guide the moving mechanism to produce displacement for pose fine-tuning.

[0009] Optionally, the actions for generating the autonomous alignment trajectory include: Based on the pose correction command, calculate the pose transformation matrix of the robot chassis physical space; The robot's motion path is smoothed by using the pose transformation matrix to construct an autonomous alignment trajectory that satisfies dynamic constraints and has spatial three-dimensional pose alignment properties.

[0010] Optionally, monitoring the sensor health status in step 3 includes: Configure a state machine snapshot extraction agent for the sensors involved in the alignment process to capture the signal-to-noise ratio and sampling latency of the sensor data stream in real time; Based on the signal-to-noise ratio and sampling delay of the sensor data stream, sensor health status characteristic data is generated to monitor the sensor health status.

[0011] Optionally, generating components to identify link anomalies includes: Kernel density estimation is performed on the sensor health status feature data to construct a probability density distribution function, and anomaly deviation calculation is performed based on the probability density distribution function to generate sensor health deviation features. The distribution of sensor health deviation characteristics over time is statistically analyzed to calculate the feature drift amount and determine the feature drift weight accordingly. Simultaneously, data loss rate statistics are performed to determine the data packet loss rate. Based on feature drift weights and packet loss rate, an identification link anomaly component is generated. This component is used to characterize the fault type, such as sensor hardware failure, communication link congestion, or data acquisition unit malfunction.

[0012] Optionally, the fault tolerance mitigation strategy vector is determined, including: Based on the constructed robot fault-tolerant knowledge base and fault evolution model, the abnormal components of the identified link are subjected to inference mapping to extract fault evolution features, and the fault evolution features are subjected to feature clustering to generate fault propagation path descriptors. By performing correlation decision analysis on the fault propagation path descriptor to identify fault compensation needs, and based on this, performing optimal remedial strategy matching processing to generate alternative control compensation components. Calculate the compensation contribution of the alternative control compensation components to generate a fault-tolerant recovery strategy vector.

[0013] Optionally, in step 3, a fault-tolerant recovery strategy vector is injected into the autonomous alignment trajectory to trigger a charging circuit control signal after physical touch locking is completed, including: The autonomous alignment trajectory is spatiotemporally discretized to determine a set of discrete trajectory points, and action command control point matching is performed based on the set of discrete trajectory points to generate a reference action command sequence. Based on the fault-tolerant recovery strategy vector and the baseline action instruction sequence, fault-tolerant guidance control instructions are generated; Based on the fault-tolerant guidance control command, the robot is guided to connect with the charging pile at the target charging coordinate and the charging circuit control signal is triggered after the physical contact lock is completed.

[0014] A robot that autonomously recharges itself using any of the methods described above.

[0015] Technical advantages of the technical solution provided in this application First, by employing a relocation algorithm to perform pose locking on a pre-set global point cloud map and combining it with charging pile reference position mapping processing, this application solves the technical deficiency of inaccurate target location determination in the prior art. Traditional solutions often rely on single sensor signals, which are easily affected by environmental noise, leading to positioning drift. This application, however, locks the pose through relocation on a global point cloud map and further performs charging pile reference position mapping processing. This ensures that the determined target charging coordinates are no longer isolated detection points, but rather reference coordinates with high spatial determinism associated with the global map. Compared to traditional methods, these coordinates determined based on global mapping provide more stable spatial guidance for subsequent planning, helping to reduce initial path deviation.

[0016] Secondly, the technical solution described in this application, which utilizes a visual docking operator for pose fine-tuning to generate an autonomous alignment trajectory with spatial vector compensation attributes, solves the technical problem of physical alignment misalignment that easily occurs in near-field docking in the prior art. Traditional mechanical docking processes often result in uncontrollable attitude deflections due to motion control errors or uneven ground. This application captures real-time deviations using a visual docking operator and converts them into trajectory parameters with spatial vector compensation attributes. This means that the autonomous alignment trajectory can be dynamically corrected in the three-dimensional vector dimension according to the actual pose state. Compared to traditional solutions, this trajectory generation method with vector compensation attributes enables the robot to achieve smoother pose self-alignment when approaching the target, significantly improving the degree of physical alignment overlap.

[0017] Third, the technical solution described in this application, which utilizes a perception-based obstacle-stopping unit to monitor the health status of sensors, generates identification link anomaly components, and determines a fault-tolerant remedial strategy vector for injection into the trajectory, solves the technical deficiency of traditional solutions where the system is prone to direct collapse or action failure when sensors malfunction. Traditional solutions typically treat the perception link as a black box; once the link malfunctions, the alignment action becomes uncontrollable. This application extracts and identifies the anomaly components of the perception link, quantifies the specific impact of the perception link failure, and determines a fault-tolerant remedial strategy vector accordingly. Dynamically injecting this strategy vector into the autonomous alignment trajectory essentially performs logical correction on the deviation caused by sensor anomalies during execution. Compared to traditional methods, this application can maintain trajectory connectivity and execution logic integrity even when the perception hardware is in a sub-healthy state, thus possessing higher fault tolerance during the physical touch-locking process. Attached Figure Description

[0018] Figure 1 This is a flowchart of a robot autonomous charging and fault-tolerant control method according to an embodiment of this application.

[0019] Figure 2 This is a schematic diagram of a robot structure according to an embodiment of this application.

[0020] Figure 3 This is an electronic device according to an embodiment of the present application. Detailed Implementation

[0021] like Figure 1 As shown, this embodiment of the present application provides a method for autonomous charging and fault-tolerant control of a robot, comprising: Step 1: Monitor the robot's energy consumption status data in real time. In response to the recognition of the charging request signal triggered by the energy consumption status data, use the relocation algorithm to lock the robot's pose in the preset global point cloud map, and perform charging pile reference position mapping processing to determine the target charging coordinates. Step 2: Plan the motion path between the robot's current pose and the target charging coordinates, and use the visual docking operator to perform pose fine-tuning when the robot approaches the target charging coordinates to a preset distance, so as to generate an autonomous alignment trajectory with spatial vector compensation attributes. Step 3: Use the obstacle sensing unit to monitor the health status of the sensors during the charging alignment action in real time, so as to generate the identification link abnormal component and determine the fault tolerance and recovery strategy vector. Then, inject the fault tolerance and recovery strategy vector into the autonomous alignment trajectory to trigger the charging circuit control signal after the physical touch lock is completed.

[0022] Optionally, identifying the charging request signal in step 1 includes: Extract the instantaneous voltage value and load current component from the energy consumption status data, and generate the remaining range entropy value by calculating the correlation distribution of the instantaneous voltage value and load current component; In response to the remaining range entropy value falling below a preset safety threshold, a charging task priority preemption operator is triggered to generate a charging request signal pointing to the target charging coordinates.

[0023] Preferably, the energy consumption status data of the power supply is first collected in real time through the robot's built-in energy consumption monitoring module. The energy consumption status data specifically includes the instantaneous voltage value and load current component of the power supply. At the same time, auxiliary energy consumption data such as voltage fluctuation amplitude, current change frequency, and power supply temperature are collected simultaneously. The collection process adopts a fixed sampling period, which can be dynamically adjusted according to the robot's working scenario (e.g., the sampling period can be set to a shorter duration in industrial park inspection scenarios, and can be appropriately extended in small scenic area inspection scenarios). This ensures that the collected data is continuous, complete, and accurately reflects the real-time working status of the power supply, providing reliable raw data support for the subsequent extraction of instantaneous voltage value and load current component, and avoiding deviations in subsequent calculation results due to data distortion or breakage.

[0024] Preferably, the collected energy consumption status data undergoes preprocessing. First, an outlier removal algorithm is used to remove abnormal data caused by sensor momentary failures, electromagnetic interference, and power supply momentary fluctuations. Then, smoothing filtering is applied to reduce high-frequency noise in the voltage and current data, retaining valid data that conforms to the normal operating characteristics of the power supply. Subsequently, instantaneous voltage values ​​and load current components are accurately extracted from the preprocessed valid data. The instantaneous voltage value represents the instantaneous voltage output of the power supply, and the load current component represents the current consumption of various loads such as the robot drive chassis, sensing unit, and actuator. The extracted instantaneous voltage value and load current component will serve as the core basic data for subsequent calculation of their correlation distribution, ensuring the relevance and accuracy of the correlation distribution calculation.

[0025] Preferably, based on the extracted instantaneous voltage value and load current component, a statistical analysis method is used to calculate the correlation distribution between the two. Specifically, firstly, corresponding data pairs of instantaneous voltage value and load current component within a preset time period are statistically analyzed. Each data pair is bound to the same collection timestamp to ensure consistency between the two in the time dimension. Then, the matching degree of each data pair is calculated. The correlation coefficient between instantaneous voltage value and load current component is obtained through a correlation analysis algorithm, and then a correlation distribution model between the two is constructed. This correlation distribution model can clearly reflect the dynamic relationship between voltage and current. When the robot load increases, the load current component increases, and the instantaneous voltage value will fluctuate accordingly. This dynamic correlation relationship is directly related to the remaining endurance of the power supply, providing a core correlation basis for the subsequent generation of the remaining endurance entropy value.

[0026] Preferably, based on the constructed correlation distribution model of instantaneous voltage value and load current component, the remaining range entropy value is calculated. The remaining range entropy value is used to quantitatively characterize the remaining range status of the power supply. Its calculation process combines the dispersion of the correlation distribution and the data fluctuation characteristics, and is obtained by quantifying the uncertainty of the correlation distribution. Specifically, the smaller the dispersion of the correlation distribution and the smoother the data fluctuation, the more stable the power supply output and the stronger the remaining range capability, and the higher the remaining range entropy value is. When the remaining power of the power supply is insufficient, the dispersion of the correlation distribution of instantaneous voltage value and load current component increases, the data fluctuation intensifies, and the remaining range entropy value decreases accordingly. This remaining range entropy value can more comprehensively and accurately reflect the actual range status of the power supply, which is different from the traditional method of judging the range by a single voltage or current value, and avoids the one-sidedness of judging by a single parameter.

[0027] Preferably, the preset safety threshold is dynamically adapted to the robot's operating scenario, the rated capacity of the power supply, and the energy consumption requirements of the actuator. The preset safety threshold varies reasonably in different operating scenarios. For example, in an industrial park inspection scenario, due to the long operating time and large load fluctuations, the preset safety threshold is set to a higher level to ensure that the robot has sufficient endurance to complete the recharging action and avoids shutdown due to power depletion during recharging. In a small scenic area inspection scenario, the operating distance is short and the load is relatively stable, so the preset safety threshold can be appropriately reduced to balance endurance utilization efficiency and recharging reliability. The preset safety threshold needs to be calibrated through multiple tests and combined with the discharge characteristics of the power supply to ensure that it can accurately reflect the critical endurance state of the power supply and provide a clear and reliable judgment benchmark for triggering subsequent charging request signals.

[0028] Preferably, the calculated remaining endurance entropy value is compared with a preset safety threshold in real time. When the remaining endurance entropy value is lower than the preset safety threshold, it indicates that the remaining endurance of the power supply can no longer meet the robot's subsequent operation needs. At this time, the charging task priority preemption operator is triggered. The core function of this operator is to raise the priority of the charging task to higher than the currently executed inspection, cleaning, and other tasks, ensuring that the robot can respond to the recharging request first and suspend the current non-urgent operation. At the same time, the charging task priority preemption operator combines the charging pile reference location information pre-stored in the preset global point cloud map to generate a charging request signal pointing to the target charging coordinates. This charging request signal includes the robot's current pose information, target charging coordinate information, and charging task priority identifier, providing clear instruction guidance for the robot to perform pose locking and plan the recharging path in the preset global point cloud map.

[0029] Preferably, the generated charging request signal undergoes validity verification. Verification includes checking whether the target charging coordinates in the signal are within the effective range of a preset global point cloud map, whether the robot's current pose information is complete, and whether the charging task priority identifier is correct. Simultaneously, considering the robot's current operating status, it is determined whether there are any special circumstances preventing immediate recharging. If the verification passes, the charging request signal is confirmed as valid, and the signal is simultaneously sent to the subsequent pose locking module to initiate the recharging process. If the verification fails, energy consumption status data is re-extracted, the remaining range entropy value is calculated, and a new charging request signal is generated. This avoids abnormal recharging actions due to signal errors or missing information, further improving the accuracy and reliability of charging request triggering. This forms a complete processing flow from energy consumption data acquisition to charging request signal generation and verification, ensuring that the results of each step provide effective support for subsequent stages.

[0030] Optionally, robot pose locking in step 1 includes: The laser SLAM operator is used to capture the point cloud features of the current environment and then perform multi-scale feature matching with a pre-set global point cloud map to obtain the robot's initial pose in the global coordinate system. Inertial navigation data is retrieved to perform zero-bias drift compensation on the initial pose, thereby generating a robot pose locking result with coordinate reference attributes and performing robot pose locking accordingly.

[0031] Preferably, before capturing the current environmental point cloud features using the laser SLAM operator, the laser SLAM operator's parameters are pre-configured according to the application scenario of robot autonomous charging. This ensures that the captured point cloud features accurately adapt to the environmental characteristics around the charging station. During parameter pre-configuration, considering the scenario characteristics of charging stations typically being deployed in fixed areas and the potential presence of fixed obstacles such as equipment cabinets and signs, the scanning frequency, ranging range, and angular resolution of the laser SLAM operator are adjusted. The scanning frequency is adapted to the robot's speed when approaching the charging station, the ranging range covers a reasonable range around the charging station (e.g., 5-10 meters), and the angular resolution is set to a high level to accurately capture the subtle contour features of the charging station and its surrounding environment. Simultaneously, a linkage mechanism is established between the laser SLAM operator and the robot's instantaneous energy consumption status data. When the energy consumption status data shows low remaining power, the scanning frequency is appropriately increased to ensure rapid capture of clear environmental point cloud features, providing a high-quality data foundation for subsequent multi-scale feature matching.

[0032] Preferably, after parameter pre-configuration, the laser SLAM operator is activated to capture the point cloud features of the robot's current environment in real time, forming a current environment point cloud feature set. During the capture process, the laser SLAM operator emits laser beams through a lidar, receives laser echo signals reflected from objects in the environment, analyzes the propagation time and reflection intensity of each laser echo, calculates the three-dimensional spatial coordinates of each laser scanning point, and then integrates them to form point cloud data. Subsequently, preprocessing is performed on the original point cloud data. First, a statistical filtering algorithm is used to remove outliers and noise points caused by laser specular reflection and ambient light interference. Then, downsampling processing is used to simplify the amount of point cloud data while retaining core feature points such as charging piles, ground markings, and surrounding fixed equipment. Finally, a current environment point cloud feature set containing multi-dimensional information such as three-dimensional spatial coordinates, reflection intensity, and contour texture is generated. This feature set can clearly characterize the spatial structure of the robot's current environment, providing clear feature basis for matching with a pre-set global point cloud map.

[0033] Preferably, a pre-set global point cloud map is retrieved and preprocessed to adapt to multi-scale feature matching requirements. This pre-set global point cloud map is a global spatial map pre-constructed by the robot during its initial operations, containing reference location information of charging piles and covering the environmental point cloud features and global coordinate system definition of the entire work area. During preprocessing, the pre-set global point cloud map is first divided into multiple sub-maps according to spatial regions, each sub-map corresponding to a work zone. Simultaneously, a feature index is established for each sub-map, containing the core feature point types and spatial distribution patterns within that region. Then, a multi-scale feature pyramid is constructed, extracting feature descriptors from the pre-set global point cloud map at different scales. Different scales correspond to different feature details; low scales correspond to overall spatial structure features, while high scales correspond to local subtle features, ensuring that subsequent matching processes can balance overall spatial positioning with accurate local matching, thus improving the robustness of the matching.

[0034] Preferably, a designed multi-scale feature matching algorithm is used to perform hierarchical matching between the current environmental point cloud feature set and a pre-set global point cloud map to obtain the robot's initial pose in the global coordinate system. The matching process is divided into two stages: coarse matching and fine matching. In the coarse matching stage, based on the feature index of the pre-set global point cloud map, the sub-map region corresponding to the current environmental point cloud feature set is quickly located. By calculating the overall similarity between the two types of point cloud features, candidate regions with high matching degree are selected to initially determine the approximate position of the robot, avoiding computational redundancy caused by indiscriminate matching of the entire map and improving matching efficiency. In the fine matching stage, within the candidate regions determined by coarse matching, based on the multi-scale feature pyramid, the current environmental point cloud features are compared with the feature descriptors of the pre-set global point cloud map at each scale. The Euclidean distance and angular deviation between feature points are calculated. The matching residual is minimized through an iterative optimization algorithm, and finally the corresponding position of the current environmental point cloud feature set in the pre-set global point cloud map is determined. Then, the initial pose of the robot in the global coordinate system is calculated. This initial pose includes the robot's three-dimensional spatial coordinates and three-axis attitude angles, providing basic pose data for subsequent zero-bias drift compensation.

[0035] Preferably, the inertial navigation data output by the robot's inertial navigation unit is retrieved, preprocessed, and then used to perform zero-bias drift compensation on the initial pose. The inertial navigation data includes parameters such as the robot's angular velocity, acceleration, and attitude angle, which can reflect changes in the robot's motion state in real time. During preprocessing, zero-bias calibration is first performed on the inertial navigation data to eliminate the zero-bias error inherent in the inertial measurement unit itself. Then, smoothing filtering is used to reduce vibration noise generated during motion, ensuring the accuracy of the inertial navigation data. Subsequently, the preprocessed inertial navigation data and the initial pose data are timestamped and aligned to ensure consistency in the time dimension, avoiding errors in the compensation results due to timing deviations. The synchronized inertial navigation data accurately reflects the robot's motion offset within a short period after the initial pose is determined, providing a reliable motion state basis for zero-bias drift compensation.

[0036] Preferably, a fusion optimization algorithm is employed, combined with synchronized inertial navigation data, to perform zero-bias drift compensation on the initial pose, eliminating the cumulative drift error present in the initial pose. During the compensation process, the robot's position drift and attitude drift per unit time are first calculated based on the inertial navigation data. The position drift is calculated by integrating acceleration, and the attitude drift is calculated by integrating angular velocity. Then, the drift is fused with the initial pose, iteratively correcting the three-dimensional spatial coordinates and attitude angles of the initial pose to gradually reduce the cumulative drift generated during laser SLAM operator matching and the instantaneous drift generated during robot motion. Simultaneously, a compensation threshold is set. When the drift exceeds the threshold, the compensation weight is increased to ensure rapid correction of deviations; when the drift is below the threshold, the compensation weight is appropriately reduced to avoid over-compensation leading to pose fluctuations. Finally, corrected pose data is generated.

[0037] Preferably, the compensated pose data is structurally encapsulated and validated to generate a robot pose locking result with coordinate reference attributes. During encapsulation, the corrected 3D spatial coordinates, attitude angles, and global coordinate system information are bound together, and a pose confidence score is added. The confidence score is determined by the feature matching degree and drift compensation error; a higher score indicates higher reliability of pose locking. During validation, the compensated pose data is compared with the environmental features of the corresponding area in the preset global point cloud map to check the consistency between the pose data and environmental features, and the stability of the pose data is also verified. If the validation passes, the pose data is confirmed as the robot pose locking result, and the robot pose locking is completed accordingly. If the validation fails, the laser SLAM operator is restarted to capture environmental point cloud features, perform multi-scale feature matching and zero-bias drift compensation, until a pose locking result that meets the requirements is generated, ensuring the accuracy and reliability of pose locking and providing a stable coordinate reference for subsequent charging pile reference position mapping and recharge path planning.

[0038] Optionally, the pose fine-tuning process in step 2 includes: A visual docking operator is used to capture auxiliary positioning marks on the surface of the charging pile, and geometric center features in the image plane are extracted from the auxiliary positioning marks. The deviation angle and distance vector of the geometric center feature relative to the robot's visual center axis are calculated to generate pose correction commands that guide the moving mechanism to produce displacement for pose fine-tuning.

[0039] Preferably, before using the visual docking operator to capture the auxiliary positioning markers on the charging pile surface, the visual docking operator is first calibrated and adapted to the scenario of near-field docking during robot autonomous charging. This ensures that the operator can accurately identify the auxiliary positioning markers on the charging pile surface and adapt to changes in illumination and distance during near-field docking. During parameter calibration, the preset characteristics of the auxiliary positioning markers are first defined—the auxiliary positioning markers on the charging pile surface are preset high-contrast markers (such as black and white circular markers or rectangular positioning frames). Based on these marker characteristics, the image acquisition parameters of the visual docking operator are adjusted, including camera focal length, exposure time, and gain coefficient. The focal length is adapted to the near-field distance when the robot approaches the charging pile (e.g., 0.5-2 meters), and the exposure time is dynamically adjusted according to the ambient light intensity to avoid overexposure in strong light or underexposure in weak light, which could lead to blurred markers. Simultaneously, a marker recognition threshold is configured to filter out areas with a high degree of matching with the preset auxiliary positioning marker characteristics, eliminating interference objects in the environment similar to the markers (such as wall stains or markers on surrounding equipment), laying the foundation for accurate capture of the auxiliary positioning markers subsequently.

[0040] Preferably, after parameter calibration is completed, the visual docking operator is activated to acquire image data of the charging pile surface in real time using the depth vision camera mounted on the robot, thereby capturing the auxiliary positioning marker. During the acquisition process, the visual docking operator simultaneously receives the robot's instantaneous pose lock result and adjusts the camera's shooting angle according to the pose lock result to ensure that the auxiliary positioning marker on the charging pile surface falls completely within the camera's field of view, avoiding missed capture of the marker due to shooting angle deviation. Subsequently, preprocessing operations are performed on the acquired image data. First, a Gaussian filtering algorithm is used to eliminate high-frequency noise in the image. Then, image enhancement processing is used to improve the contrast between the auxiliary positioning marker and the background of the charging pile surface, strengthening the edge contour features of the marker. Next, a contour detection algorithm is used to extract regions that meet the preset size and shape features from the preprocessed image. A feature matching algorithm is used to compare these regions with the preset feature template of the auxiliary positioning marker, and the region with the highest matching degree is selected as the captured auxiliary positioning marker, ensuring the accuracy of the capture result and avoiding misjudging interference areas as auxiliary positioning markers.

[0041] Preferably, feature extraction is performed on the captured auxiliary positioning marker, focusing on extracting its geometric center feature within the image plane. The image plane is a two-dimensional plane imaged by the depth vision camera, and the geometric center feature serves as the core positioning reference for the auxiliary positioning marker, directly determining the accuracy of subsequent pose fine-tuning. During extraction, the outline of the auxiliary positioning marker is first fitted. A corresponding fitting algorithm is selected based on the marker's preset shape (circle, rectangle). If the auxiliary positioning marker is circular, a circle fitting algorithm is used, calculating the center coordinates of all pixels on the outline using the least squares method; these center coordinates represent the geometric center of the marker within the image plane. If the auxiliary positioning marker is rectangular, a rectangle fitting algorithm is used, calculating the two diagonals of the rectangular outline; the intersection of these diagonals represents the geometric center. After extraction, the two-dimensional coordinates (pixel coordinates) of the geometric center within the image plane are recorded, along with the corresponding image resolution parameters. This provides accurate feature data for subsequent calculations of the deviation angle and distance vector, ensuring that the geometric center feature accurately represents the core position of the auxiliary positioning marker.

[0042] Preferably, before calculating the deviation angle and distance vector of the geometric center feature relative to the robot's visual center axis, the robot's visual center axis is calibrated and standardized first. This clarifies the relative positional relationship between the visual center axis, the robot's moving mechanism, and the depth vision camera, ensuring the accuracy of the calculation reference. During calibration, the geometric center of the robot chassis is used as the reference to determine the installation position parameters of the depth vision camera, and then the visual center axis is calibrated—the visual center axis is the optical axis of the depth vision camera, perpendicular to the image plane and passing through the center of the image plane. This axis is aligned and calibrated with the forward direction of the robot's moving mechanism. The attitude parameters of the visual center axis in the robot's body coordinate system are recorded. Simultaneously, a mapping relationship between the image plane coordinates and the robot's body coordinate system is established, converting the image plane pixel coordinates of the geometric center feature into coordinates in the robot's body coordinate system. This provides a unified coordinate reference for the subsequent calculation of the deviation angle and distance vector, avoiding deviations in the calculation results due to inconsistent coordinate references.

[0043] Preferably, based on the calibrated coordinate data of the visual center axis and the geometric center feature, the deviation angle and distance vector of the geometric center feature relative to the visual center axis are calculated respectively. These two parameters jointly characterize the degree of deviation between the robot's current pose and the charging pile's assisted positioning marker, providing a quantitative basis for generating pose correction commands. The deviation angle is calculated as a horizontal deviation angle and a vertical deviation angle. The horizontal deviation angle is the angle of offset of the geometric center feature in the horizontal direction relative to the visual center axis, and the vertical deviation angle is the angle of offset in the vertical direction. By calculating the angle between the geometric center coordinates and the projected coordinates of the visual center axis in the image plane, combined with camera intrinsic parameters, the pixel angle is converted into an actual spatial angle. The distance vector is calculated as the spatial distance of the geometric center feature relative to the visual center axis. First, the geometric center pixel coordinates in the image plane are converted into three-dimensional coordinates in the robot's body coordinate system. Then, the shortest spatial distance between these three-dimensional coordinates and the visual center axis is calculated to obtain the distance vector. The direction of the distance vector is consistent with the direction of the deviation angle, jointly reflecting the specific location and magnitude of the deviation.

[0044] Preferably, based on the calculated deviation angle and distance vector, combined with the motion characteristics of the robot's mobile mechanism, a pose correction command is generated to guide the mobile mechanism to produce displacement. This command is used to control the robot's mobile mechanism to fine-tune its pose, aligning the geometric center feature with the visual center axis, thereby achieving accurate alignment between the robot and the charging pile. During the calculation process, the deviation angle and distance vector are first converted into motion parameters of the mobile mechanism. The horizontal deviation angle corresponds to the robot chassis's steering angle, the vertical deviation angle corresponds to the camera's pitch angle (if the camera is adjustable), and the distance vector corresponds to the robot's forward or backward distance. Then, based on the mobile mechanism's maximum steering angle, maximum moving speed, and other physical parameters, the motion parameters are limited to prevent the command parameters from exceeding the mobile mechanism's motion range, which could lead to damage or loss of control. Finally, the processed motion parameters are encapsulated into a standardized pose correction command. This command includes core parameters such as steering angle, moving distance, and motion execution speed, clearly defining the mobile mechanism's motion logic and ensuring that the mobile mechanism can accurately execute the pose fine-tuning actions according to the command.

[0045] Preferably, the pose correction command undergoes validity verification and dynamic adaptation to ensure that it can adapt to real-time deviation changes during near-field docking, further improving the accuracy of pose fine-tuning. During verification, the parameters of the pose correction command are first checked to ensure they conform to the motion constraints of the mobile mechanism, and to identify any abnormal parameters or logical contradictions. If the verification passes, the command is sent to the robot's mobile mechanism to initiate pose fine-tuning. If the verification fails, the deviation angle and distance vector are recalculated, and the pose correction command is regenerated. Simultaneously, during the fine-tuning process, the visual docking operator continuously captures the geometric center features of the auxiliary positioning marker, updates the deviation angle and distance vector in real time, and dynamically adjusts the parameters of the pose correction command. This ensures that the fine-tuning action can track deviation changes in real time, gradually reducing the deviation between the geometric center features and the visual center axis until they are aligned, completing the pose fine-tuning process. This provides an accurate pose basis for the subsequent generation of the autonomous alignment trajectory, ensuring the robot can achieve accurate near-field docking with the charging station.

[0046] Optionally, the actions for generating the autonomous alignment trajectory include: Based on the pose correction command, calculate the pose transformation matrix of the robot chassis physical space; The robot's motion path is smoothed by using the pose transformation matrix to construct an autonomous alignment trajectory that satisfies dynamic constraints and has spatial three-dimensional pose alignment properties.

[0047] Preferably, before calculating the pose transformation matrix of the robot chassis in physical space based on the pose correction instructions, the pose correction instructions generated by the upper-level process are preprocessed and parameter parsed to ensure that the instruction parameters can accurately adapt to the chassis kinematic model, providing a reliable input basis for the calculation of the pose transformation matrix. During preprocessing, abnormal parameters generated by real-time deviation detection in the pose correction instructions are first removed, and then smoothing filtering is used to weaken the instantaneous fluctuations of the instruction parameters, ensuring the stability of core parameters such as steering angle, movement distance, and motion execution speed. During parameter parsing, parameters such as steering angle (horizontal and vertical directions), movement distance, and motion execution sequence are accurately extracted from the instructions. The physical meaning of each parameter is also labeled, clarifying the impact of each parameter on the chassis pose change. For example, the steering angle determines the chassis's attitude deflection, and the movement distance determines the chassis's position offset. The parsed parameters will directly serve as the basis for calculating the pose transformation matrix, avoiding distortion of the matrix calculation results due to parameter parsing deviations.

[0048] Preferably, the physical parameters of the robot chassis are retrieved and associated with the parsed pose correction command parameters to clarify the motion constraints of the chassis's physical space, providing a physical reference for the construction of the pose transformation matrix. The chassis physical parameters include core parameters such as chassis wheelbase, track width, drive wheel radius, maximum steering servo angle, and chassis geometric center coordinates. These parameters directly determine the chassis's motion characteristics and pose transformation rules. During the association and binding process, the steering angle in the pose correction command is compared with the maximum steering servo angle of the chassis to ensure that the commanded steering angle is within the chassis's physical motion range. The movement distance is correlated with the chassis drive wheel radius and converted into the drive wheel rotation angle. Simultaneously, a chassis physical space coordinate system is established with the chassis geometric center as the origin, clarifying the coordinates of key points on the chassis (such as the drive wheel center and the vision camera mounting point) in this coordinate system, providing a spatial reference for the calculation of the elements of the pose transformation matrix.

[0049] Preferably, based on the parsed pose correction command parameters, associated chassis physical parameters, and chassis physical space coordinate system, a pose transformation matrix of the robot chassis physical space is constructed and calculated. This matrix characterizes the spatial transformation relationship of the robot chassis from its current pose to the target fine-tuned pose, including two types of transformation information: position translation and attitude rotation. The calculation process of the pose transformation matrix is ​​combined with the kinematic model of the robot chassis (such as the Ackerman steering model or differential drive model). The matrix elements consist of translational and rotational components. The translational components correspond to the displacement of the chassis in the x, y, and z directions in three-dimensional space, which are obtained by converting the movement distance in the pose correction command with the chassis physical parameters. The rotational components correspond to the changes in the pitch, yaw, and roll angles of the chassis in three-dimensional space, which are calculated from the steering angle in the pose correction command. Specifically, the rotation matrix is ​​first calculated to represent the rotational change of the chassis attitude, and then the translation vector is calculated to represent the translational change of the chassis position. The rotation matrix and the translation vector are integrated to form a complete 4×4 pose transformation matrix. This matrix can accurately describe the complete transformation process of the chassis pose and provide the core transformation basis for subsequent trajectory smoothing calculation.

[0050] Preferably, the calculated pose transformation matrix is ​​validated to ensure that it accurately reflects the actual pose transformation of the chassis and avoids deviations in subsequent trajectory construction due to matrix calculation errors. The validation process first verifies the mathematical validity of the pose transformation matrix by checking the existence of its determinant and inverse to ensure it conforms to the mathematical laws of spatial transformation and avoids singular matrices. Secondly, it verifies the physical rationality of the matrix by comparing the translation and rotation components with the chassis's physical motion constraints to check whether the displacement and rotation angles exceed the chassis's physical motion range and are consistent with the expected pose correction command. Finally, simulation is used to verify whether the chassis pose change corresponding to the pose transformation matrix can align the geometric center feature with the robot's visual center axis. If the validation passes, the pose transformation matrix is ​​confirmed as valid and used for subsequent trajectory smoothing calculations. If the validation fails, the pose correction command parameters are re-parsed, the chassis physical parameters are retrieved, and the pose transformation matrix is ​​recalculated until the validation passes.

[0051] Preferably, the motion path between the robot's current pose and the target charging coordinates, planned in step 2, is retrieved. Preprocessing is then performed on this motion path, focusing on extracting the end segment (i.e., the path segment from the robot's approach to the target charging coordinates to a preset distance), defining the core processing range for end-point trajectory smoothing calculation. During preprocessing, the entire motion path is first discretized into multiple continuous trajectory nodes, each bound to its corresponding three-dimensional spatial coordinates and attitude angle. Then, based on the robot's near-field docking requirements, trajectory nodes of a preset length at the end of the path (e.g., nodes within the range of 0.3-1 meter) are selected as the end-point trajectory segment. This segment directly determines the near-field docking accuracy between the robot and the charging pile and is the core object of trajectory smoothing calculation. Subsequently, redundant and abnormal nodes in the end-point trajectory segment are removed, and missing trajectory nodes are added to ensure the continuity and integrity of the end-point trajectory segment. Simultaneously, the coordinates of the end-point trajectory segment are transformed to the chassis physical space coordinate system, maintaining consistency with the coordinate reference of the pose transformation matrix, providing a unified coordinate basis for subsequent trajectory smoothing calculation.

[0052] Preferably, using the validated pose transformation matrix, a smoothing calculation is performed on the preprocessed path end trajectory segment. The smoothing calculation strictly adheres to the dynamic constraints of the robot chassis to ensure the smoothed trajectory can be stably executed by the chassis. The smoothing calculation employs a designed piecewise interpolation algorithm (such as B-spline interpolation), based on the discrete nodes of the end trajectory segment. Combined with the pose transformation rules corresponding to the pose transformation matrix, transition trajectory points are inserted between adjacent trajectory nodes, transforming the original polygonal end trajectory into a continuous, smooth curve trajectory. Simultaneously, the chassis's dynamic constraint parameters (such as maximum travel speed, maximum acceleration / deceleration, and maximum steering angular velocity) are incorporated into the interpolation calculation process. The speed, acceleration, and steering angle of the transition trajectory points are limited to prevent sudden stops, sharp turns, and abrupt acceleration changes, ensuring the trajectory conforms to the chassis's motion characteristics and can be smoothly executed by the drive mechanism. Furthermore, during the smoothing calculation, the trajectory's attitude parameters are adjusted synchronously to ensure that the robot's visual center axis remains aligned with the charging pile auxiliary positioning marker during its movement along the trajectory, laying the foundation for three-dimensional attitude alignment.

[0053] Preferably, after the end-point trajectory smoothing calculation is completed, a three-dimensional attitude alignment verification and dynamic constraint review are performed on the smoothed trajectory to construct an autonomous alignment trajectory that meets dynamic constraints and possesses spatial three-dimensional attitude alignment attributes. During the three-dimensional attitude alignment verification process, the robot attitude angle corresponding to each trajectory node on the smoothed trajectory is calculated in real time. The attitude angle is compared with the alignment requirements of the charging pile auxiliary positioning mark, and it is checked whether the robot's visual central axis and geometric central feature are always aligned. If there is an attitude deviation, the attitude parameters of the trajectory are adjusted until the three-dimensional attitude alignment requirements are met. During the dynamic constraint review process, the velocity, acceleration, and turning angle of each node on the smoothed trajectory are checked again to see if they conform to the physical motion constraints of the chassis. If there are any parameters that exceed the constraint range, the smoothing interpolation parameters are readjusted to correct the trajectory curve. Finally, the smoothed trajectory that has been verified and reviewed is seamlessly connected with the non-end-point segment of the path to form a complete autonomous alignment trajectory. This trajectory has both continuous and smooth motion characteristics, can adapt to the dynamic constraints of the chassis, and spatial three-dimensional attitude alignment attributes, which can guide the robot to accurately complete the near-field docking with the charging pile and provide reliable trajectory guidance for the execution of subsequent charging alignment actions.

[0054] Preferably, a final validity check is performed on the constructed autonomous alignment trajectory to ensure that the trajectory meets the accuracy and safety requirements of autonomous charging near-field docking. During the check, the entire process of the robot moving along the autonomous alignment trajectory is simulated, and parameters such as the robot's pose, velocity, and acceleration at each node of the trajectory are recorded. The smoothness of the trajectory and the accuracy of the posture alignment are checked to see if they meet the preset standards, and whether there is a risk of collision with the charging pile or surrounding obstacles. At the same time, combined with the robot's energy consumption data, the energy consumption during the trajectory execution is checked to see if it is within a reasonable range, avoiding excessive energy consumption due to unreasonable trajectory design, which would affect the smooth completion of charging docking. If the check passes, the autonomous alignment trajectory is confirmed to be valid and used to guide subsequent charging alignment actions. If the check fails, the steps of pose transformation matrix calculation and end-point trajectory smoothing calculation are repeated until a qualified autonomous alignment trajectory is generated, ensuring the accuracy and reliability of autonomous charging near-field docking.

[0055] Optionally, monitoring the sensor health status in step 3 includes: Configure a state machine snapshot extraction agent for the sensors involved in the alignment process to capture the signal-to-noise ratio and sampling latency of the sensor data stream in real time; Based on the signal-to-noise ratio and sampling delay of the sensor data stream, sensor health status characteristic data is generated to monitor the sensor health status.

[0056] Preferably, before configuring the state machine snapshot extraction agent for the sensors involved in the alignment process, the sensors involved in the charging alignment process are first classified, sorted, and their parameters are recorded. The functional positioning, data stream format, and operating parameters of each type of sensor are clarified, providing a targeted basis for the accurate configuration of the state machine snapshot extraction agent. The sensors involved in the alignment process mainly include depth vision cameras, LiDAR, inertial measurement units, and collision sensors. The depth vision camera is used to capture charging pile auxiliary positioning markers, the LiDAR is used to supplement near-field environmental features, the inertial measurement unit is used to assist in pose calibration, and the collision sensor is used to detect physical touch lock states. During the sorting process, the sampling frequency, data transmission protocol, and signal output format of each type of sensor are recorded. The extraction location and calculation benchmark of the signal-to-noise ratio and sampling delay in the data streams of each type of sensor are clarified, ensuring that the subsequent state machine snapshot extraction agent can accurately adapt to the operating characteristics of different sensors and avoid data capture failure or distortion due to differences in sensor types.

[0057] Preferably, based on the analyzed sensor parameters, a state machine snapshot extraction agent is configured separately for each type of sensor involved in the alignment action. This enables independent capture and accurate monitoring of data streams from various sensors, avoiding interference and misjudgment caused by mixed capture of multiple sensor data streams. The state machine snapshot extraction agent adopts a scenario-based configuration mode, adjusting the agent's snapshot capture cycle, data parsing rules, and caching strategy according to the characteristics of data streams from different types of sensors: For depth vision cameras, the snapshot capture cycle is consistent with the camera imaging cycle, focusing on parsing the signal strength and noise components in the image data stream for calculating the signal-to-noise ratio; for LiDAR and inertial measurement units, the snapshot capture cycle is adapted to the sensor's sampling frequency, focusing on recording the data acquisition time and the time when the data is transmitted to the processing unit for calculating the sampling delay; simultaneously, an independent cache space is configured for each agent to temporarily store the captured snapshot data, preventing data loss and providing complete data support for subsequent calculations of signal-to-noise ratio and sampling delay.

[0058] Preferably, after the state machine snapshot extraction agent is started, it captures the data stream of the corresponding sensor in real time. Using a preset data parsing algorithm, it accurately extracts the signal-to-noise ratio (SNR) parameter from the data stream, ensuring that the SNR data truly reflects the purity of the sensor signal. During the extraction process, the captured sensor data stream is first processed into frames, with each frame bound to a corresponding acquisition timestamp. Then, a signal separation operation is performed on each frame to distinguish between valid and noise signals—valid signals are the real environmental information or motion state information collected by the sensor, while noise signals are invalid signals generated by environmental interference, electromagnetic interference, and sensor wear. Subsequently, the intensity of the valid signal and the intensity of the noise signal in each frame are calculated, and the SNR of that frame is obtained by their ratio. The SNR of multiple consecutive frames is then smoothed to reduce the impact of instantaneous fluctuations, resulting in the real-time SNR of the sensor data stream. A higher SNR value indicates higher purity of the sensor signal and stronger data reliability.

[0059] Preferably, while extracting the signal-to-noise ratio, the state machine snapshot extraction agent synchronously captures the sampling delay of the sensor data stream, accurately quantifying the time interval from data acquisition to data availability, providing a quantitative basis for judging the sensor's response speed. The calculation of the sampling delay is centered on timestamp synchronization. The state machine snapshot extraction agent records two key time points: one is the moment the sensor completes data acquisition (acquisition timestamp), generated by the sensor's own timing module; the other is the moment the acquired data is transmitted to the charging alignment control module, completed parsing, and becomes usable (parsing completion timestamp), generated by the agent's own timing module. By calculating the difference between the two timestamps, the sampling delay of a single frame of data is obtained. Then, statistical analysis is performed on the sampling delay of multiple consecutive frames of data, eliminating abnormal delay values ​​caused by transmission congestion, to obtain the average sampling delay and instantaneous sampling delay of the sensor data stream. The average sampling delay is used to characterize the overall response performance of the sensor, while the instantaneous sampling delay is used to capture real-time response fluctuations of the sensor.

[0060] Preferably, the extracted real-time signal-to-noise ratio (SNR), average sampling delay, and instantaneous sampling delay are structurally integrated and quantified to generate sensor health status feature data. This data is used to comprehensively and accurately characterize the real-time health status of the sensor, providing a clear quantitative basis for subsequent sensor health status monitoring. During the integration process, the SNR and sampling delay are mapped to preset standardized intervals (such as the 0-1 interval). The standardized values ​​can intuitively reflect the sensor's health level; the higher the standardized SNR value and the lower the standardized sampling delay value, the better the sensor's health status. Simultaneously, sensor identifiers, data acquisition timestamps, and data validity identifiers are added to the sensor health status feature data to clarify the sensor type and acquisition time corresponding to each set of feature data, ensuring the traceability of the feature data. Furthermore, based on the sensor's rated operating parameters, corresponding reference thresholds are configured for each indicator in the feature data for subsequent health status determination, forming a complete sensor health status feature dataset.

[0061] Preferably, based on the generated sensor health status feature data, real-time monitoring of the sensor health status is performed. Through comprehensive analysis of multi-dimensional indicators, the system accurately determines whether the sensor is in normal working condition and promptly identifies potential health hazards. During the monitoring process, the signal-to-noise ratio (SNR) and sampling delay in the feature data are compared with preset reference thresholds in real time. Simultaneously, the system combines the changing trends of multiple consecutive frames of feature data for tiered judgment: if the SNR is higher than the preset threshold, the sampling delay is lower than the preset threshold, and the data change trend is stable, the sensor is determined to be in normal health, capable of continuously providing reliable data for charging alignment. If the SNR is close to the preset threshold, the sampling delay is close to the preset threshold, and the data shows slight fluctuations, the sensor is determined to be in a sub-healthy state, possibly with slight interference or slight loss, requiring continuous monitoring. If the SNR is lower than the preset threshold, the sampling delay is higher than the preset threshold, or the data shows drastic fluctuations, the sensor is determined to be in an abnormal state, exhibiting problems such as data acquisition distortion and response lag. Simultaneously, the monitoring results are fed back to the charging alignment control module in real time, providing accurate health status basis for subsequent identification of abnormal components in the link and determination of fault-tolerant remedial strategy vectors.

[0062] Preferably, the sensor health status monitoring process is logged and anomaly warnings are issued to ensure the traceability of the monitoring process and to promptly remind subsequent links to deal with sensor anomalies. During the logging process, the health status monitoring results, feature data changes, threshold comparisons, and monitoring time of each type of sensor are recorded in detail to form a complete sensor health monitoring log for subsequent fault diagnosis, sensor operation and maintenance, and algorithm optimization. During the anomaly warning process, when a sensor is detected to be in a sub-healthy or abnormal state, an alarm signal is immediately triggered. The alarm signal includes the sensor identifier, the type of anomaly (abnormal signal-to-noise ratio or abnormal sampling delay), the degree of anomaly, and the time of occurrence. The alarm signal is simultaneously integrated into the sensor health status feature data to ensure that when subsequent link anomaly component identification is performed, the abnormal sensor and the cause of the anomaly can be accurately located. This provides targeted support for the generation of fault-tolerant recovery strategy vectors and avoids charging alignment errors or loss of control due to sensor health problems.

[0063] Optionally, generating components to identify link anomalies includes: Kernel density estimation is performed on the sensor health status feature data to construct a probability density distribution function, and anomaly deviation calculation is performed based on the probability density distribution function to generate sensor health deviation features. The distribution of sensor health deviation characteristics over time is statistically analyzed to calculate the feature drift amount and determine the feature drift weight accordingly. Simultaneously, data loss rate statistics are performed to determine the data packet loss rate. Based on feature drift weights and packet loss rate, an identification link anomaly component is generated. This component is used to characterize the fault type, such as sensor hardware failure, communication link congestion, or data acquisition unit malfunction.

[0064] Preferably, before performing kernel density estimation on the sensor health status feature data, preprocessing is performed to remove data noise and abnormal interference, ensuring the integrity and reliability of the data input to the kernel density estimation algorithm, thus laying the foundation for the accurate construction of the subsequent probability density distribution function. During preprocessing, abnormal data points caused by instantaneous electromagnetic interference or temporary sensor malfunctions are first removed from the sensor health status feature data. Linear interpolation is then used to supplement missing data points, improving the continuity of the feature data. Subsequently, the feature data is standardized and normalized, mapping the signal-to-noise ratio, sampling delay, and data packet loss-related feature parameters to the same numerical range, eliminating estimation bias caused by differences in the magnitude of different parameters. Simultaneously, the feature data is classified and grouped according to the sensor type involved in the charging alignment action, with each group corresponding to the health status feature data of one type of sensor. This ensures that the kernel density estimation processing can be specifically adapted to the working characteristics of different sensors, avoiding estimation distortion caused by mixed processing of multi-sensor data.

[0065] Preferably, the designed kernel density estimation algorithm is used to process the preprocessed sensor health status feature data to construct a probability density distribution function that reflects the normal distribution pattern of the feature data. This function is used to characterize the normal fluctuation range and distribution characteristics of the sensor health status features, providing a benchmark for subsequent calculation of abnormal deviations. During the kernel density estimation process, a suitable Gaussian kernel function is selected based on the temporal characteristics and distribution features of the sensor health status feature data. By determining a reasonable kernel bandwidth (the kernel bandwidth is dynamically adjusted according to the dispersion of the feature data; the greater the dispersion, the larger the kernel bandwidth, and vice versa), probability density estimation is performed on the health status feature data of each type of sensor. Specifically, taking each sample point of the sensor health status feature data as the core, the kernel function is used to diffuse to surrounding sample points, calculating the probability contribution of each sample point in the overall dataset. Then, the probability contributions of all sample points are integrated to construct a complete probability density distribution function. The peak value of this function corresponds to the normal benchmark value of the sensor health status features, and the width of the function curve corresponds to the normal fluctuation range of the feature data, clearly reflecting the normal distribution pattern of the sensor health status.

[0066] Preferably, based on the constructed probability density distribution function, anomaly deviation calculation is performed to quantify the degree of deviation of each sensor health status feature data point from the normal distribution range, thereby generating sensor health deviation features. During the anomaly deviation calculation process, the normal distribution interval corresponding to the probability density distribution function is first determined. This interval, centered on the function peak, covers a preset proportion (e.g., 95%) of normal feature data points. Data within the interval is considered normal healthy feature data, while data outside the interval is considered potentially abnormal data. Then, the distance from each feature data point to the peak of the probability density distribution function is calculated. Combined with the slope of the function curve, the anomaly deviation of that data point is quantified. The larger the deviation value, the further the feature data point deviates from the normal range, and the more abnormal the sensor health status. Finally, the anomaly deviation, deviation direction (higher or lower than the normal baseline value), corresponding data acquisition timestamp, and sensor identifier of each feature data point are bound together to form the sensor health deviation feature. This feature can accurately characterize the abnormal details of each type of sensor health status, providing core data support for subsequent feature drift calculation.

[0067] Preferably, the generated sensor health deviation features are subjected to time-series statistical analysis to calculate the feature drift amount and determine the feature drift weight accordingly, thereby quantifying the development trend and impact of sensor health state abnormalities. During the time-series statistical analysis, a fixed statistical time window is set (the time window is dynamically adjusted according to the execution duration of the charging alignment action, such as 1-3 seconds). Within each time window, the time-series distribution of sensor health deviation features is statistically analyzed, and the difference between health deviation features at two adjacent moments is calculated. The differences at multiple consecutive moments are accumulated to obtain the feature drift amount. The feature drift amount is used to characterize the development speed and accumulation degree of sensor health state abnormalities; the larger the drift amount, the faster the sensor health state deteriorates. Subsequently, the feature drift weight is determined based on the functional importance of the sensor in the charging alignment action. The more critical the function of the sensor (such as a depth vision camera used to capture auxiliary positioning markers), the higher the drift weight is configured, meaning that its health state abnormality has a greater impact on the entire alignment link. Multiplying the feature drift weight by the feature drift amount yields the comprehensive impact value of the sensor health abnormality, providing a quantitative basis for subsequent identification of abnormal components in the link.

[0068] Preferably, while calculating the feature drift amount and feature drift weight, the sensor data stream is simultaneously subjected to loss rate statistical processing to accurately determine the data packet loss rate and quantify the data transmission status of the communication link. During the loss rate statistical processing, sensor data stream records captured by the agent are extracted based on state machine snapshots. The total number of data packets sent by the sensor per unit time is compared with the number of data packets actually successfully received and parsed. The data packet loss rate is calculated by the ratio of the number of successfully received data packets to the total number of data packets. During the statistical process, cases where data packets cannot be sent due to sensor malfunction are excluded; only data packets that are normally sent but not successfully received are counted to ensure that the loss rate accurately reflects the transmission status of the communication link. Simultaneously, the data packet loss rate is time-series tracked, recording instantaneous changes and average levels. If the loss rate suddenly increases and remains at a high level, it indicates that there may be congestion or a fault in the communication link. The data packet loss rate will serve as the core quantitative indicator for identifying communication link anomalies, working together with the feature drift weight to support the generation of anomaly identification components.

[0069] Preferably, based on the calculated feature drift weights and data packet loss rate, and combined with preset anomaly judgment rules, an identification link anomaly component is generated. This component can accurately characterize the specific fault type of sensor hardware failure, communication link congestion, or data acquisition unit anomaly. During the generation process, the correlation mapping relationship between feature drift weight, data packet loss rate, and fault type is first established: when the feature drift weight is high and the data packet loss rate is low, it indicates that the sensor health status is continuously deteriorating but data transmission is normal, and it is judged as a sensor hardware failure fault; when the data packet loss rate is high and the feature drift weight is low, it indicates that the sensor health status is normal but data transmission is abnormal, and it is judged as a communication link congestion fault; when both the feature drift weight and the data packet loss rate are high, it indicates that the sensor health status is abnormal and data transmission is abnormal, and it is judged as a data acquisition unit abnormal fault. Subsequently, the fault type, abnormality degree (quantified by feature drift weight and data packet loss rate), corresponding sensor identifier, abnormality occurrence timestamp, and other information are structurally integrated to generate a link abnormality identification component. This component can not only clarify the specific type of link abnormality, but also quantify the severity and impact range of the abnormality, providing accurate abnormality basis for subsequent determination of fault tolerance and remediation strategy vectors, ensuring that fault tolerance and remediation actions can specifically solve link abnormality problems.

[0070] Preferably, the generated abnormal link components are validated to ensure they accurately represent link anomalies and prevent misjudgments or omissions that could render subsequent fault-tolerant remediation strategies ineffective. During validation, the abnormal link components are first compared with sensor health monitoring logs and data stream transmission logs to check if the fault type and severity recorded in the abnormal components match the actual monitoring data. Secondly, considering the real-time status of the charging alignment action, it is determined whether the link anomaly will actually affect the alignment action. If the fault type corresponding to the abnormal component does not affect alignment accuracy and safety, it is deemed an invalid anomaly and removed. If the fault type corresponding to the abnormal component affects the alignment action and is consistent with the actual monitoring data, it is deemed a valid abnormal component. Finally, the valid abnormal link components are synchronized to the charging alignment control module to provide reliable support for determining subsequent fault-tolerant remediation strategy vectors. Simultaneously, the generation process and validation results of the abnormal components are recorded to form a complete link anomaly identification log for subsequent fault investigation and algorithm optimization.

[0071] Optionally, the fault tolerance mitigation strategy vector is determined, including: Based on the constructed robot fault-tolerant knowledge base and fault evolution model, the abnormal components of the identified link are subjected to inference mapping to extract fault evolution features, and the fault evolution features are subjected to feature clustering to generate fault propagation path descriptors. By performing correlation decision analysis on the fault propagation path descriptor to identify fault compensation needs, and based on this, performing optimal remedial strategy matching processing to generate alternative control compensation components. Calculate the compensation contribution of the alternative control compensation components to generate a fault-tolerant recovery strategy vector.

[0072] Preferably, before performing inference mapping processing on the identified link anomaly components, the construction of the robot fault-tolerant knowledge base and the scenario-based training of the fault evolution model are completed first to ensure that the two can accurately adapt to the robot's autonomous charging alignment scenario, providing reliable support for the inference mapping and fault evolution analysis of link anomalies. The construction of the robot fault-tolerant knowledge base revolves around the core components involved in the charging alignment action, such as sensors, communication links, and data acquisition units. It includes various fault cases, corresponding remedial strategies, fault impact ranges, and processing effect data. The fault cases cover various link anomaly types, including sensor hardware failure, communication link congestion, and data acquisition unit malfunctions. Each fault case is bound to corresponding identified link anomaly component characteristics, fault evolution patterns, and optimal remedial solutions. A dynamic update mechanism is established to promptly add new faults and remedial strategies that occur during each charging alignment process, ensuring the completeness and timeliness of the knowledge base. The fault evolution model adopts a time-series prediction model architecture, trained based on historical link anomaly data, sensor health status change data, and charging alignment action execution data. The model input is the identified link anomaly component, and the output is the fault evolution trend and impact range. During training, the model's time window and prediction parameters are adjusted according to the characteristics of the autonomous charging near-field docking scenario to ensure the model can accurately predict the evolution path and potential consequences of different link anomalies, providing model support for subsequent fault evolution feature extraction.

[0073] Preferably, based on the constructed robot fault-tolerant knowledge base and the trained fault evolution model, inference mapping processing is performed on the identified link anomaly components to accurately extract fault evolution features. These features characterize the development trend, impact range, and potential cascading faults of the link anomaly. During the inference mapping process, the identified link anomaly components are first input into the fault evolution model. The model, combined with its own trained fault evolution patterns, performs time-series deduction of the fault type corresponding to the anomaly component, predicting the fault development speed, affected link links, and potential secondary faults. For example, sensor hardware failure may lead to inaccurate capture of auxiliary positioning markers, thereby causing pose fine-tuning deviations. Simultaneously, fault cases matching the identified link anomaly component are retrieved from the robot fault-tolerant knowledge base. The fault evolution features in these cases are extracted as references. The model deduction results are fused and optimized with the knowledge base reference features, eliminating contradictory information and supplementing missing features. Finally, fault evolution features containing multi-dimensional information such as fault evolution speed, fault impact range, secondary fault risk, and fault duration are extracted. These features comprehensively reflect the dynamic evolution process of the link anomaly, providing core data for subsequent feature clustering processing.

[0074] Preferably, feature clustering processing is performed on the extracted fault evolution features, and similar fault evolution features are classified and integrated to generate a fault propagation path descriptor that can clearly characterize the fault propagation law. Feature clustering processing employs a designed hierarchical clustering algorithm. Combining multi-dimensional parameters of fault evolution characteristics, reasonable clustering thresholds and rules are set. The clustering process is divided into two levels: the first level uses fault type as the clustering basis, grouping fault evolution characteristics belonging to the same fault type (such as sensor hardware failure or communication link congestion) into one category to avoid feature confusion between different types of faults; the second level uses fault evolution speed and impact range as the clustering basis, further clustering fault evolution characteristics with similar evolution patterns and similar impact ranges within the same fault type category to form multiple clusters; subsequently, feature extraction is performed on each cluster to summarize the common patterns of fault evolution characteristics within the cluster, clarifying the fault propagation direction, propagation speed, core link links affected, and possible chain reactions. These common patterns are bound to the fault type and anomaly degree corresponding to the cluster, integrating them to generate fault propagation path descriptors. Each descriptor corresponds to a type of link anomaly with similar propagation patterns, accurately characterizing the propagation process and impact characteristics of this type of anomaly, providing a clear basis for subsequent correlation decision analysis.

[0075] Preferably, by performing correlation decision analysis on the fault propagation path descriptor and combining it with the real-time execution status of the charging alignment action, the fault compensation requirements can be accurately identified, clarifying what remedial measures are needed to mitigate or eliminate the impact of link anomalies. During the correlation decision analysis, the fault propagation path descriptor is first correlated with the core requirements of the charging alignment action. The impact of fault propagation on charging alignment accuracy, action execution stability, and charging docking safety is analyzed. For example, data packet loss due to communication link congestion may affect the real-time issuance of pose correction commands, leading to near-field docking deviation. In this case, the compensation requirement is to improve data transmission stability and reduce data packet loss. Simultaneously, by combining corresponding fault cases in the robot's fault-tolerant knowledge base and referring to historical remedial experience, the feasibility and adaptability of different compensation methods are analyzed. Compensation methods that are not suitable for the current scenario or cannot solve the actual fault are eliminated. The core objectives, priorities, and specific requirements of compensation are clarified. For example, when sensor hardware fails, the priority of the compensation requirement is to quickly switch to redundant sensors to ensure that auxiliary positioning markers can be captured normally. The second priority is to adjust the pose fine-tuning parameters to adapt to the acquisition characteristics of redundant sensors. Finally, clear and specific fault compensation requirements are formed, providing clear guidance for subsequent optimal remedial strategy matching.

[0076] Preferably, based on the identified fault compensation requirements, the set of remedial strategies in the robot's fault-tolerant knowledge base is retrieved, the optimal remedial strategy matching process is executed, and alternative control compensation components that can meet the compensation requirements and adapt to the current abnormal link scenario are generated. During the matching process, a mapping rule is established between fault compensation requirements and remedial strategies. Based on the priority of the compensation requirement, the fault type, and the degree of anomaly, all remedial strategies that can meet the compensation requirement are selected from the knowledge base. For example, for the compensation requirement of sensor hardware failure, multiple remedial strategies such as redundant sensor switching, emergency adjustment of sensor parameters, and adaptive trajectory correction are selected. Subsequently, the selected remedial strategies are evaluated for their adaptability. Evaluation indicators include the execution difficulty of the remedial strategy, the compensation effect, the degree of impact on the charging alignment action, and energy consumption cost. Combined with the current energy consumption status of the robot and the real-time progress of the charging alignment action, each remedial strategy is comprehensively scored, and the top-scoring remedial strategies are selected. These remedial strategies are converted into standardized control parameters to form candidate control compensation components. Each candidate component contains core information such as the type of remedial action, execution parameters, action sequence, and applicable scope, which can be directly used for the fault-tolerant control of subsequent charging alignment actions, ensuring the practicality and relevance of the candidate components.

[0077] Preferably, the compensation contribution of the generated alternative control compensation components is calculated. By quantifying the compensation effect and influence weight of each alternative component, the optimal compensation component is selected and integrated to generate a fault-tolerant remedial strategy vector. In the process of calculating the compensation contribution, multi-dimensional evaluation indicators are set, including compensation accuracy (the degree to which the impact of the fault can be mitigated), execution efficiency (the response speed of the remedial action), stability (the persistence of the effect after the remedial action is executed), and compatibility (the adaptability to the current charging alignment trajectory and pose correction instructions). Each evaluation indicator is assigned a corresponding weight, which is dynamically adjusted according to the priority of the fault compensation requirement. For example, compensation accuracy has a higher weight in sensor hardware failure scenarios, and execution efficiency has a higher weight in communication link congestion scenarios. Subsequently, the score of each candidate control compensation component under each evaluation indicator is calculated, and the comprehensive compensation contribution of each candidate component is obtained by weighted summation. The higher the contribution value, the better the compensation effect and the stronger the adaptability of the candidate component. Finally, the candidate control compensation component with the highest comprehensive compensation contribution is selected and structurally integrated with the corresponding fault type, compensation requirement, and execution sequence to generate a fault-tolerant remedial strategy vector. This vector contains complete fault-tolerant control parameters, which can be directly injected into the autonomous alignment trajectory to achieve accurate remediation of link anomalies and ensure that the charging alignment action can be executed normally.

[0078] Preferably, the generated fault-tolerant remedial strategy vector undergoes validity and adaptability verification to ensure that the vector accurately adapts to the current link anomaly scenario, meets fault compensation requirements, and conforms to the dynamic constraints of the robot chassis and the real-time state of the charging alignment action. During the verification process, firstly, it checks whether the control parameters in the fault-tolerant remedial strategy vector conform to the working constraints of the robot's mobile mechanism, sensors, and communication links, and whether there are any parameter anomalies or logical contradictions. Secondly, through simulation, the execution process of the strategy vector is simulated to verify whether it can effectively alleviate or eliminate the impact of link anomalies, whether it can ensure charging alignment accuracy and action stability, and whether it will cause new faults or anomalies. Finally, combined with the robot's real-time energy consumption status and the progress of the charging alignment action, the execution cost and feasibility of the strategy vector are verified. If the verification passes, the fault-tolerant remedial strategy vector is confirmed as effective and used for subsequent injection of autonomous alignment trajectories. If the verification fails, the optimal remedial strategy matching and compensation contribution calculation are re-executed, and the control parameters of the strategy vector are adjusted until a fault-tolerant remedial strategy vector that meets the requirements is generated. Simultaneously, the verification results and adjustment process are recorded in the fault-tolerant knowledge base to provide a reference for handling similar faults in the future.

[0079] The fault evolution model described in this application comprises an input adaptation layer, a time-series deduction core layer, a knowledge base fusion and correction layer, a feature output layer, and a closed-loop feedback layer.

[0080] The input adaptation layer is the sole data entry point for the fault evolution model. It unifies and standardizes the multi-source heterogeneous input data, eliminating data timing deviations, format differences, and noise interference. This provides qualified input data with unified format, time alignment, and no redundant noise for subsequent time series extrapolation, ensuring the consistency and reliability of the model input. The input data of this layer is divided into two categories. The first category is the identification link anomaly component output from the preceding sensor health monitoring stage, which includes three core parameters: feature drift weights, data packet loss rate, and fault type identifiers. The second category is supporting data from the robot's local storage and the robot's fault-tolerant knowledge base, including historical link anomaly data, sensor health status change data, and charging alignment action execution data. The specific processing flow of this layer is as follows: First, perform time-series alignment processing, unifying multi-source data with different acquisition frequencies to the same time-series granularity through linear interpolation, ensuring that each set of data has a unique corresponding timestamp and eliminating inference bias caused by time-series misalignment; Second, perform data cleaning processing, using the 3σ criterion to remove abnormal data points caused by electromagnetic interference and instantaneous acquisition failures, and using linear interpolation to complete missing values ​​lost during data transmission, ensuring data continuity and integrity; Third, perform standardization mapping processing, uniformly mapping parameters with different dimensions and numerical ranges to a standardized interval of 0-1, eliminating the impact of parameter magnitude differences on subsequent inference; Fourth, perform scenario-based filtering processing, retaining only feature data related to the core links of the charging alignment link (sensor acquisition, data transmission, pose fine-tuning, trajectory execution), removing irrelevant and redundant data, and compressing data dimensionality. This layer ultimately outputs a standardized time-series feature set, which serves as the sole input to the core layer of time-series inference, and is simultaneously backed up to the robot's fault-tolerant knowledge base for subsequent model iteration training.

[0081] The temporal inference core layer is the core computational unit of the fault evolution model. Based on the standardized temporal feature set output by the input adaptation layer and combined with the fault evolution rules learned during model training, it performs temporal inference on the development trend, impact range, and secondary fault risk of the current link anomaly, generating initial fault evolution inference results. This provides basic data for subsequent fusion correction and is the core link in realizing the transformation from "identifying link anomaly components to fault evolution trends". The input data of this layer consists of the standardized temporal feature set output by the input adaptation layer and the fault evolution rule library completed by model pre-training. The fault evolution rule library comes from the model's scenario-based training phase and is obtained by training with historical link anomaly data, sensor health status change data, and charging alignment action execution data. The specific processing flow of this layer is as follows: First, perform evolution pattern matching, matching the current standardized temporal feature set with the historical fault evolution patterns in the fault evolution pattern library to identify the historical evolution pattern with the highest matching degree with the current anomaly, thus clarifying the basic evolution pattern of the current fault; Second, set a scenario-based temporal simulation window, dynamically adjusting the simulation window length based on the characteristics of the autonomous charging near-field docking scenario and the current distance between the robot and the charging pile, as well as the execution progress of the charging alignment action, to ensure that the simulation accuracy matches the scenario requirements; Third, perform temporal simulation calculation, based on the matched evolution pattern, within the set simulation window, quantitatively calculating the development speed and duration of the fault, clarifying the charging alignment link affected by the fault, and predicting the secondary faults that the fault may cause; Fourth, integrate and generate initial simulation results, integrating the four dimensions of fault evolution speed, fault impact range, secondary fault risk, and fault duration into the initial simulation results of fault evolution. This layer ultimately outputs the initial inference result of the fault evolution, which is directly used as the input of the knowledge base fusion correction layer and simultaneously synchronized to the robot fault-tolerant knowledge base for matching and retrieval of fault cases.

[0082] The knowledge base fusion and correction layer is the optimization unit for the inference results of the fault evolution model. It fuses and optimizes the initial inference results output by the time-series inference core layer with matching fault cases in the robot fault-tolerant knowledge base, eliminating contradictory information in the initial inference results, supplementing missing evolutionary features, correcting inference biases, and improving the accuracy and scenario adaptability of the fault evolution results. It is the core link in combining model inference results with actual engineering experience. The input data for this layer are the initial fault evolution inference results output by the time-series inference core layer and the corresponding fault cases matched in the robot fault-tolerant knowledge base. The cases contain the evolutionary features, evolutionary rules, and actual processing effect data of the corresponding faults. The specific processing flow of this layer is as follows: First, perform matching case retrieval. Based on the fault type and anomaly degree in the initial simulation results, retrieve the 3-5 sets of historical fault cases with the highest matching degree from the robot fault-tolerant knowledge base, and extract the fault evolution features in the cases as reference benchmarks. Second, perform consistency comparison. Compare the initial simulation results with the case reference features dimension by dimension to identify contradictions and missing points. Contradictions refer to the content of the initial simulation prediction that does not match the actual occurrence of historical cases. Missing points refer to the fault impact links that are not covered by the initial simulation but are clearly recorded in historical cases. Third, perform fusion optimization processing. For contradictory information, make a reasonable judgment based on the real-time status of the current charging alignment, retain the content that is suitable for the current scenario, and remove the simulation results that do not conform to the actual working conditions. For missing information, directly supplement the corresponding evolution features from historical cases to improve the dimensions of the initial simulation results. Fourth, perform consistency verification. Perform logical verification on the optimized simulation results to ensure that the fault evolution law, impact range, and secondary faults are logically consistent and without contradictions. The final output of this layer is the optimized full feature set of fault evolution. This result is directly used as the input of the feature output layer and is simultaneously synchronized to the robot fault-tolerant knowledge base for subsequent fault clustering analysis.

[0083] The feature output layer is the result output unit of the fault evolution model. It performs structured encapsulation and format adaptation of the optimized full set of fault evolution features, generating standardized fault evolution features that are directly output to subsequent feature clustering and association decision-making stages. Simultaneously, it ensures that the output features fully match the processing requirements of subsequent stages, serving as a bridge connecting the model to the entire fault-tolerant control process. The input data for this layer is the optimized full set of fault evolution features output by the knowledge base fusion correction layer. The specific processing flow of this layer is as follows: First, perform structured encapsulation, binding the four core dimensions of the full set of fault evolution features—fault evolution speed, fault impact range, secondary fault risk, and fault duration—with corresponding fault type identifiers, anomaly degree parameters, and timestamps to form structured fault evolution features. Each feature has a unique identifier and complete dimensional information. Second, perform format adaptation, converting the structured fault evolution features into a standardized format that can be directly called by the corresponding steps according to the format requirements of subsequent feature clustering processing and association decision analysis, ensuring that subsequent steps can use them directly without secondary processing. Third, perform validity verification, checking whether the output fault evolution features have missing dimensions, abnormal parameters, or logical contradictions. Output is only allowed after the verification passes. This layer finally outputs standardized fault evolution features, which are transmitted in two paths: one path is sent directly to the subsequent feature clustering processing stage as the core data basis for clustering, and the other path is backed up to the robot fault-tolerant knowledge base for the construction and storage of fault propagation path descriptors.

[0084] The closed-loop feedback layer is the closed-loop optimization unit of the fault evolution model. It receives the execution results and validity verification data of subsequent fault-tolerant remediation strategies, performs dynamic optimization of model parameters and updates the robot's fault-tolerant knowledge base, realizes continuous model iteration, improves the model's inference accuracy in autonomous charging and alignment scenarios, and implements the dynamic update mechanism of the robot's fault-tolerant knowledge base. The input data of this layer includes the validity verification results of the fault-tolerant remediation strategy vector, strategy execution effect data, the final completion status of the charging and alignment actions, and the actual implementation matching degree data of the fault evolution features. The specific processing flow of this layer is as follows: First, perform an execution effect review, comparing the actual evolution process, final impact range, and secondary fault occurrence of this fault with the fault evolution characteristics output by the model, and calculating the accuracy of the model inference; Second, perform a knowledge base update, supplementing the robot's fault-tolerant knowledge base with the complete case of this fault, including the characteristics of the identified abnormal components of the link, the actual fault evolution law, the fault-tolerant remediation strategy, and the processing effect data, thereby improving the case coverage of the knowledge base; Third, perform model parameter optimization, based on the accuracy of this inference and the deviation between the actual fault and the inference results, fine-tuning and optimizing the model's time window, prediction parameters, and similarity matching threshold, completing the scenario-based iterative training of the model, and improving the accuracy of subsequent inferences; Fourth, perform closed-loop verification, checking whether the updated knowledge base and optimized model parameters meet the requirements of the autonomous charging alignment scenario, ensuring no logical errors and data redundancy. This layer ultimately outputs optimized model parameters and an updated robot fault-tolerant knowledge base. The optimized model parameters directly affect the next fault evolution time-series deduction stage, and the updated knowledge base provides more comprehensive case support for the next fault matching and fusion correction.

[0085] Optionally, in step 3, a fault-tolerant recovery strategy vector is injected into the autonomous alignment trajectory to trigger a charging circuit control signal after physical touch locking is completed, including: The autonomous alignment trajectory is spatiotemporally discretized to determine a set of discrete trajectory points, and action command control point matching is performed based on the set of discrete trajectory points to generate a reference action command sequence. Based on the fault-tolerant recovery strategy vector and the baseline action instruction sequence, fault-tolerant guidance control instructions are generated; Based on the fault-tolerant guidance control command, the robot is guided to connect with the charging pile at the target charging coordinate and the charging circuit control signal is triggered after the physical contact lock is completed.

[0086] To improve the accuracy of fault-tolerant strategy injection and the compliance of trajectory execution, before trajectory discretization and instruction generation, it is necessary to complete the compliance pre-verification of the autonomous alignment trajectory and the structured analysis of the fault-tolerant remediation strategy vector, establishing unified constraint boundaries and execution benchmarks for subsequent processing. Specifically, the autonomous alignment trajectory compliance pre-verification uses the global point cloud map coordinate system, the robot chassis kinematic model, and near-field obstacle avoidance safety rules as constraints to verify the spatial continuity, attitude smoothness, and dynamic compliance of the trajectory. Abnormal segments in the trajectory that exhibit pose jumps, exceed driving capability boundaries, or breach safety distances are eliminated, ensuring that the trajectory to be processed is the optimal continuous trajectory that meets the charging docking accuracy requirements under fault-free operating conditions. The structured parsing of fault-tolerant recovery strategy vectors involves decomposing the strategy vectors generated in the preceding stages into dimensions and extracting five core pieces of information bound to the vectors: fault type identifier, fault compensation dimension, parameter correction weight, safety constraint boundary, and action execution timing requirements. This clarifies the link anomaly scenarios (sensor hardware failure, communication link congestion, data acquisition unit anomaly) and core compensation targets corresponding to the vectors, ensuring that subsequent strategy injections can accurately match fault compensation needs and avoid trajectory inaccuracies caused by indiscriminate corrections.

[0087] This study employs a spatiotemporal dual-dimensional discretization method to decouple the pre-verified autonomous alignment trajectory, generating a discrete trajectory point set with both temporal and spatial constraints. The mathematical essence of this discretization process is to decouple the continuous three-dimensional spatial trajectory curve from the continuous temporal motion process into a discrete spatiotemporal node set strictly synchronized with the robot chassis control cycle. This process consists of two core components: spatial adaptive discretization and temporal synchronization binding. The spatial adaptive discretization stage, based on the curvature change rate and docking accuracy requirements of the autonomous alignment trajectory, adopts a variable step-size discretization strategy: for near-field alignment segments and attitude adjustment segments with large curvature, millimeter-level small step sizes are used for discretization to ensure no loss of trajectory details and alignment accuracy; for long-distance approach segments with small curvature, decimeter-level large step sizes are used for discretization, improving trajectory continuity while reducing computational load. Each discrete spatial node is bound to three-dimensional spatial coordinates in the global coordinate system, three-axis attitude angles, trajectory curvature, upper limits of linear and angular velocity, and obstacle avoidance safety boundary parameters. The time synchronization binding process maps the discrete spatial nodes one-to-one with the fixed control cycle of the robot chassis controller, assigns a unique timestamp to each spatial node, and ensures that the time interval between adjacent nodes is completely consistent with the chassis control cycle. This ultimately forms a set of discrete trajectory points that are coupled in time and space, providing an accurate time and space reference for subsequent motion command control point matching.

[0088] Based on a set of discrete trajectory points, a reference motion command sequence aligned with the discrete trajectory timeline and adapted to the parameters is constructed through motion command control point matching processing. The core logic of motion command control point matching is to establish a quantitative mapping relationship between the spatial pose parameters of discrete trajectory points and the executable control parameters of the robot chassis actuator, realizing the transformation from "trajectory pose to execution command". In the specific implementation process, firstly, based on the robot's pre-calibrated kinematic model (Ackerman steering model / differential drive model), for each spatiotemporal node in the discrete trajectory point set, the node's three-dimensional coordinates, attitude angles, and velocity constraint parameters are solved into underlying control parameters such as the target speed of the drive wheel, the target angle of the steering servo, and the acceleration / deceleration gradient within the corresponding control cycle. Then, based on the parameter threshold range specified by the drive chassis control protocol, the solved control parameters are subjected to amplitude limiting processing and standardization encapsulation to ensure that all parameters are within the rated working range of the actuator, avoiding problems such as command overshoot and actuator overload. Finally, the standardized control parameters within all control cycles are concatenated in chronological order according to timestamps to generate a reference motion command sequence. Each instruction in the sequence is bound to a corresponding trajectory node identifier, timestamp, and loop check code, serving as the baseline execution sequence for the robot to complete charging docking under fault-free operating conditions, while also providing a stable reference benchmark for subsequent fault-tolerant corrections.

[0089] By integrating fault-tolerant recovery strategy vectors and baseline action command sequences, fault-tolerant guidance and control commands adapted to the current link anomaly scenario are generated through multi-dimensional fault-tolerant correction processing. The core logic of fault-tolerant correction is: without exceeding safety constraints and docking accuracy requirements, the baseline action command sequence is adaptively adjusted based on fault compensation needs to achieve accurate injection of the fault-tolerant recovery strategy vector into the trajectory execution stage. The correction process consists of three core dimensions: First, parameter domain correction: for scenarios where observation accuracy is reduced due to sensor hardware failure, the approach velocity and acceleration / deceleration gradient parameters in the baseline action command sequence are lowered to increase the pose verification frequency; for scenarios where command transmission delay is caused by communication link congestion, the execution cycle of a single command is extended, and command retransmission mechanisms and timeout fault-tolerant parameters are added. Second, timing domain correction: for scenarios where sampling delay is increased due to data acquisition unit anomalies, the action execution timing of the baseline action command sequence is adjusted, and the timing nodes of pose observation, trajectory tracking, and action execution are delayed to eliminate the problem of asynchronous command execution caused by sampling delay. Thirdly, trajectory domain correction addresses the local obstacle avoidance requirements caused by secondary faults. This involves fine-tuning the poses of corresponding nodes in the discrete trajectory point set and simultaneously correcting the control parameters of the corresponding commands to ensure the adjusted trajectory still meets the attitude alignment requirements for charging pile docking. After correction, all adjusted commands undergo safety and compliance verification to ensure that command parameters do not exceed chassis dynamics constraints and obstacle avoidance safety boundaries. Finally, the verified commands are standardized and encapsulated to generate fault-tolerant guidance control commands with fault identifiers, correction weights, and verification information.

[0090] Based on fault-tolerant guidance control commands, a phased closed-loop control framework is designed to guide the robot to complete the entire charging docking process with the charging pile at the target charging coordinates. The docking process is divided into three progressive closed-loop control stages. Each stage uses the fault-tolerant guidance control commands as the execution benchmark and simultaneously integrates real-time perception data to perform deviation correction: The first stage is the near-field approach closed-loop stage. The robot approaches the charging pile along the corrected autonomous alignment trajectory according to the timing and parameter requirements of the fault-tolerant guidance control commands. During the process, the auxiliary positioning marks on the surface of the charging pile are captured in real time through the visual docking operator. The deviation between the actual pose and the target pose of the trajectory is continuously compared. Based on the deviation, the driving parameters are dynamically fine-tuned to ensure that the robot and the charging pile always maintain the preset posture alignment relationship during the approach process; The second stage is the accurate alignment closed-loop stage. When the robot and the charging pile are in close alignment, the robot approaches the charging pile in close alignment. Once the straight-line distance enters the preset near-field alignment threshold range, the system switches to a low-speed micro-motion control mode. Based on the micro-adjustment parameters in the fault-tolerant guidance control command, it performs millimeter-level fine-tuning of the lateral, longitudinal, and attitude angles to ensure that the coaxiality and parallelism of the robot's charging docking mechanism and the charging pile interface meet the mechanical docking tolerance requirements. The third stage is the physical contact closed-loop stage. After accurate alignment, the robot is driven to slowly approach the charging pile at an extremely low contact speed set by the fault-tolerant guidance control command. Simultaneously, contact signals are collected in real time through the front-end collision sensor and pressure sensor until stable physical contact is detected, completing the physical contact process between the charging docking mechanism and the charging pile. Throughout the docking process, the health status of the sensor link is continuously monitored. If a new link anomaly occurs, the fault-tolerant recovery strategy vector is immediately updated and the fault-tolerant guidance control command is dynamically adjusted to improve the stability of the docking process.

[0091] To improve the safe conduction of the charging circuit, a multi-dimensional progressive physical contact lock validity verification mechanism is designed, triggering the charging circuit control signal only after the verification is passed. Physical contact lock validity verification is a prerequisite for triggering the charging signal and consists of three progressive verification stages: The first stage is mechanical contact stability verification, which uses continuous sampling signals from collision and pressure sensors to confirm that the contact pressure between the charging docking mechanism and the charging pile is within a preset rated range, and that there are no pressure fluctuations or contact detachment within multiple consecutive sampling cycles, eliminating the risk of false or misaligned contact; the second stage is posture consistency verification, which compares the deviation between the robot's current real-time posture and the target charging coordinates to confirm that the three-dimensional spatial position deviation and attitude angle deviation are within the mechanical tolerance range of the charging docking, ensuring that the charging contacts can completely overlap; the third stage is electrical connection pre-verification, which uses the low-voltage detection circuit of the robot's charging management unit to confirm the electrical continuity of the positive and negative charging contacts, ensuring there are no abnormalities such as short circuits, open circuits, or excessive contact resistance. After the lock validity verification is completed, the robot main controller sends a charging enable command with encryption verification to the charging management unit. After receiving the command, the charging management unit closes the high-voltage contactor of the main charging circuit and sends a charging handshake protocol frame to the charging pile through the charging communication bus to complete the two-way communication handshake between the robot and the charging pile. Finally, the charging circuit control signal is triggered to start the constant current and constant voltage charging process.

[0092] Throughout the entire process of fault-tolerant guidance, charging docking, and charging circuit triggering, a full-cycle closed-loop verification and anomaly fallback mechanism is constructed. Simultaneously, the entire process data is archived and the knowledge base is updated, forming a complete fault-tolerant control closed loop. During the full-cycle closed-loop verification, the deviation between the execution effect of the fault-tolerant guidance control command and the expected target is continuously compared in fixed control cycles. If the deviation exceeds a preset threshold, the current action is immediately paused, and fault identification, fault-tolerant remediation strategy matching, and command correction are re-executed until the deviation returns to the allowable range. During the anomaly fallback process, multiple safety protection mechanisms are set up. In the event of emergency abnormal conditions such as mechanical contact misalignment, electrical short circuit, or complete link interruption, the drive system power output and the high-voltage power supply of the charging circuit are immediately cut off, triggering emergency braking and audible and visual alarms to prevent mechanical damage and electrical safety accidents. During the full-process data archiving process, all data, including the type of link anomaly, fault-tolerant remediation strategy vector, fault-tolerant guidance control instruction correction process, charging docking execution effect, and charging circuit triggering result, are fully archived into the robot fault-tolerant knowledge base. At the same time, the processing effect data of the corresponding fault cases are updated, providing data support for the iterative optimization of the subsequent fault evolution model and the improvement of the matching accuracy of the fault-tolerant remediation strategy.

[0093] like Figure 2 As shown in the figure, this application provides a robot, which includes: The energy consumption monitoring module is used to monitor the robot's energy consumption status data in real time. In response to the recognition of a charging request signal triggered by the energy consumption status data, it uses a relocation algorithm to lock the robot's pose in a preset global point cloud map and performs charging pile reference position mapping processing to determine the target charging coordinates. The pose fine-tuning module is used to plan the motion path between the robot's current pose and the target charging coordinates. When the robot approaches the target charging coordinates to a preset distance, it uses a visual docking operator to perform pose fine-tuning to generate an autonomous alignment trajectory with spatial vector compensation attributes. The fault-tolerant control module is used to monitor the health status of sensors in real time during the charging alignment action by using the obstacle-sensing unit to generate identification link abnormal components and determine the fault-tolerant recovery strategy vector. The fault-tolerant recovery strategy vector is then injected into the autonomous alignment trajectory to trigger the charging circuit control signal after physical touch locking is completed.

[0094] like Figure 3 The image shows an electronic device that includes a processor and a memory. The memory is used to store computer programs; When the processor executes the program stored in the memory, it implements the steps of the method described in any of the above embodiments.

[0095] A computer-readable storage medium storing a computer program, which, when executed by a processor, implements any of the aforementioned robot autonomous charging and fault-tolerant control methods.

[0096] Figures 2-3 For an exemplary explanation of the embodiments, please refer to the above. Figure 1 I will not go into details here.

Claims

1. A method for autonomous charging and fault-tolerant control of a robot, characterized in that, include: Step 1: Monitor the robot's energy consumption status data in real time. In response to the recognition of a charging request signal triggered by the energy consumption status data, use a relocation algorithm to lock the robot's pose in a preset global point cloud map and perform charging pile reference position mapping processing to determine the target charging coordinates. Step 2: Plan the motion path between the robot's current pose and the target charging coordinates, and use the visual docking operator to perform pose fine-tuning when the robot approaches the target charging coordinates to a preset distance, so as to generate an autonomous alignment trajectory with spatial vector compensation attributes. Step 3: Use the obstacle sensing unit to monitor the health status of the sensors during the charging alignment action in real time, so as to generate the identification link abnormal component and determine the fault tolerance and recovery strategy vector, and inject the fault tolerance and recovery strategy vector into the autonomous alignment trajectory to trigger the charging circuit control signal after the physical touch lock is completed.

2. The robot autonomous charging and fault-tolerant control method according to claim 1, characterized in that, The step 1 of identifying the charging request signal includes: Extract the instantaneous voltage value and load current component from the energy consumption status data, and generate the remaining range entropy value by calculating the correlation distribution of the instantaneous voltage value and the load current component; In response to the remaining range entropy value being lower than a preset safety threshold, a charging task priority preemption operator is triggered to generate the charging request signal pointing to the target charging coordinates.

3. The robot autonomous charging and fault-tolerant control method according to claim 1, characterized in that, The robot pose locking process in step 1 includes: The current environmental point cloud features are captured using a laser SLAM operator and then matched with the preset global point cloud map at multiple scales to obtain the robot’s initial pose in the global coordinate system. Inertial navigation data is retrieved to perform zero-bias drift compensation on the initial pose, thereby generating a robot pose locking result with coordinate reference attributes and performing robot pose locking accordingly.

4. The robot autonomous charging and fault-tolerant control method according to claim 1, characterized in that, The pose fine-tuning process in step 2 includes: The visual docking operator is used to capture auxiliary positioning marks on the surface of the charging pile, and the geometric center features in the image plane are extracted from the auxiliary positioning marks. The deviation angle and distance vector of the geometric center feature relative to the robot's visual center axis are calculated to generate pose correction commands that guide the moving mechanism to produce displacement for pose fine-tuning.

5. The robot autonomous charging and fault-tolerant control method according to claim 4, characterized in that, The action of generating the autonomous alignment trajectory in step 2 includes: Based on the pose correction command, calculate the pose transformation matrix of the robot chassis physical space; The pose transformation matrix is ​​used to perform end-point trajectory smoothing calculation on the robot's motion path to construct the autonomous alignment trajectory that satisfies dynamic constraints and has spatial three-dimensional pose alignment properties.

6. The robot autonomous charging and fault-tolerant control method according to claim 1, characterized in that, Step 3, monitoring the sensor health status, includes: Configure a state machine snapshot extraction agent for the sensors involved in the alignment action to capture the signal-to-noise ratio and sampling delay of the sensor data stream in real time; Based on the signal-to-noise ratio and sampling delay of the sensor data stream, sensor health status characteristic data is generated to monitor the sensor health status.

7. The robot autonomous charging and fault-tolerant control method according to claim 6, characterized in that, Step 3, which generates the component for identifying link anomalies, includes: Kernel density estimation is performed on the sensor health status feature data to construct a probability density distribution function, and anomaly deviation calculation is performed based on the probability density distribution function to generate sensor health deviation features. The distribution of the sensor health deviation characteristics over time is statistically analyzed to calculate the feature drift amount and determine the feature drift weight accordingly. Simultaneously, data packet loss rate statistical processing is performed to determine the data packet loss rate. Based on the feature drift weights and data packet loss rate, an identification link anomaly component is generated. The identification link anomaly component is used to characterize the fault type of sensor hardware failure, communication link congestion, or data acquisition unit abnormality.

8. The robot autonomous charging and fault-tolerant control method according to claim 7, characterized in that, Step 3, determining the fault tolerance and mitigation strategy vector, includes: Based on the constructed robot fault-tolerant knowledge base and fault evolution model, the abnormal components of the identified link are subjected to inference mapping processing to extract fault evolution features, and the fault evolution features are subjected to feature clustering processing to generate fault propagation path descriptors. By performing correlation decision analysis on the fault propagation path descriptor to identify fault compensation needs, and based on this, performing optimal remedial strategy matching processing to generate alternative control compensation components. Calculate the compensation contribution of the alternative control compensation components to generate a fault-tolerant recovery strategy vector.

9. The robot autonomous charging and fault-tolerant control method according to claim 1, characterized in that, Step 3 involves injecting the fault-tolerant recovery strategy vector into the autonomous alignment trajectory to trigger the charging circuit control signal after physical touch locking is completed, including: The autonomous alignment trajectory is spatiotemporally discretized to determine a set of discrete trajectory points, and action command control point matching is performed based on the set of discrete trajectory points to generate a reference action command sequence. Based on the fault-tolerant recovery strategy vector and the baseline action instruction sequence, fault-tolerant guidance control instructions are generated; Based on the fault-tolerant guidance control command, the robot is guided to dock with the charging pile at the target charging coordinates and the charging circuit control signal is triggered after the physical contact lock is completed.

10. A robot, characterized in that, Autonomous charging is performed using the method of any one of claims 1-9.

Images