Sensor system of a logistics robot and data processing method thereof

Abstract

The application provides a sensor system of a logistics robot and a data processing method thereof, and belongs to the technical field of intelligent logistics, robot perception and artificial intelligence signal processing. The method comprises the following steps: collecting multi-source data of the logistics robot in the moving process in real time through a multi-source sensor, and generating a multi-source sensor original data set; performing data fusion preprocessing on the multi-source sensor original data set, and generating a fused synchronous data stream; based on the synchronous data stream, the dynamic barycenter position data of the goods are obtained through a barycenter inversion algorithm combined with a three-dimensional model of the goods and pressure distribution characteristics, and corresponding timestamp information is recorded synchronously to generate a dynamic barycenter position-time sequence data of the goods; and through the fusion of the multi-source sensor data, the dynamic barycenter position of the goods can be inverted in real time and accurately, and the perception ability of the robot to the barycenter change is greatly improved.

Description

TECHNICAL FIELD

[0001] The application provides a sensor system of a logistics robot and a data processing method thereof, and belongs to the technical field of intelligent logistics, robot perception and artificial intelligence signal processing. BACKGROUND

[0002] With the improvement of the automation degree of logistics, logistics robots are increasingly widely applied in warehousing, carrying and other links.

[0003] However, in actual operation, logistics robots often face complex working conditions such as cargo tilting and center of gravity deviation, which leads to a decrease in stability and increases the risk of overturning. This not only may cause damage to the goods, but also threatens the safety of the surrounding environment and personnel.

[0004] Traditional stability control methods mostly rely on a single sensor or a preset strategy, and it is difficult to cope with the actual environment that changes dynamically. For example, only through the acceleration sensor to monitor the tilt and make simple adjustment, when the center of gravity of the goods changes rapidly, the effect is limited.

[0005] At the same time, although some robots are equipped with multiple sensors, they lack effective data fusion and processing mechanisms, and the information utilization rate is low, and the regulation and control are not accurate.

[0006] In recent years, new technologies such as multi-source sensor data fusion and dynamic stability prediction have provided new ideas for the stability control of logistics robots. However, how to apply these technologies to logistics robots to realize active and accurate regulation and control of their stability is still a problem to be solved. SUMMARY

[0007] The application provides a sensor system of a logistics robot and a data processing method thereof, which are used to solve the problems mentioned in the background art:

[0008] The application provides a sensor data processing method of a logistics robot, which comprises the following steps:

[0009] S1, collecting multi-source data of the logistics robot in the moving process in real time through a multi-source sensor to generate a multi-source sensor original data set;

[0010] S2, performing data fusion preprocessing on the multi-source sensor original data set to generate a fused synchronous data stream; based on the synchronous data stream, through a center of gravity inversion algorithm, in combination with a three-dimensional model of the goods and pressure distribution characteristics, the dynamic center of gravity position data of the goods is inversely obtained, and corresponding timestamp information is synchronously recorded to generate a dynamic center of gravity position-time sequence data of the goods;

[0011] S3, according to the dynamic center of gravity position-time sequence data of the goods, combined with the path planning information preset by the logistics robot, a dynamic stability prediction model is used for stability feedforward prediction; and a stability feedforward prediction result data is generated;

[0012] S4, based on the stability feedforward prediction result data, a collaborative adjustment instruction of the counterweight and the stability fin is generated through a collaborative adjustment strategy; the collaborative adjustment instruction is sent to the actuator of the logistics robot, and real-time collaborative adjustment of the counterweight and the stability fin is performed to generate execution feedback data;

[0013] S5, according to the execution feedback data, the actual stability state of the logistics robot is verified in a closed loop; through comparison of the prediction result and the actual execution effect, the parameters of the center of gravity inversion algorithm and the stability prediction model are dynamically optimized; at the same time, the closed loop verification result is fed back to the multi-source sensor data processing flow to form a perception-decision-execution closed loop control system; finally, real-time data dashboard of logistics robot stability regulation is generated.

[0014] The sensor system of the logistics robot provided by the application, the system comprises:

[0015] One or more processors; a memory for storing one or more programs; wherein when the one or more programs are executed by the one or more processors, the one or more processors implement the data processing method of any one of the above.

[0016] The application has the beneficial effects that: through fusion of multi-source sensor data, the dynamic center of gravity position of the goods can be inversely calculated in real time and accurately, and the perception ability of the robot to the center of gravity change is greatly improved. Combined with path planning, stability feedforward prediction is carried out, so that the robot can predict and adjust the posture in advance, and the overturning risk is effectively reduced. The generated collaborative adjustment instruction of the counterweight and the stability fin realizes active and accurate regulation of the stability of the robot, and the working safety of the robot in complex environment is significantly improved. The method forms a perception-decision-execution closed loop, breaks through the limitation of traditional isolated monitoring, and makes the robot have stronger environmental adaptability and autonomous regulation ability. Through the application of the technical scheme of the application, the working efficiency and reliability of the logistics robot are greatly improved, and the maintenance cost is reduced. BRIEF DESCRIPTION OF DRAWINGS

[0017] Figure 1 The flowchart of the sensor data processing method of the logistics robot of the application. DETAILED DESCRIPTION

[0018] The preferred embodiments of the application are described below in conjunction with the drawings, and it should be understood that the preferred embodiments described herein are only used to illustrate and explain the application, and are not used to limit the application.

[0019] One embodiment of the present application, as shown in Figure 1 A sensor data processing method of a logistics robot, the method comprising:

[0020] S1, a multi-source sensor is deployed on the logistics robot, the multi-source sensor including an acceleration sensor, a gyroscope sensor, a pressure sensor, and a vision sensor; multi-source data of the logistics robot during movement is collected in real time by the multi-source sensor, the multi-source data including acceleration data, angular velocity data, cargo pressure distribution data, and environmental vision image data, and a multi-source sensor original data set is generated;

[0021] S2, the multi-source sensor original data set is preprocessed by data fusion, a Kalman filter algorithm is used to eliminate noise interference and time synchronization error between sensor data, a synchronized data stream after fusion is generated; based on the synchronized data stream, a gravity inversion algorithm is used to obtain cargo dynamic gravity center position data by combining a cargo three-dimensional model and pressure distribution characteristics, and corresponding timestamp information is recorded synchronously, and cargo dynamic gravity center position-time sequence data is generated;

[0022] S3, based on the cargo dynamic gravity center position-time sequence data, a dynamic stability prediction model is used to perform stability feedforward prediction in combination with preset path planning information of the logistics robot; the model analyzes the correlation between the gravity center position change trend and environmental parameters such as path curvature and slope, evaluates the overturn risk probability of the logistics robot in a future time window, and generates stability feedforward prediction result data;

[0023] S4, based on the stability feedforward prediction result data, a collaborative adjustment strategy is used to generate collaborative adjustment instructions for counterweights and stability fins; the counterweight adjustment instruction dynamically adjusts the position of the internal counterweight block of the robot according to the gravity center offset direction and amplitude; the stability fin adjustment instruction controls the fin deployment angle and speed according to the predicted overturn risk level; the collaborative adjustment instructions are sent to the actuator of the logistics robot to perform real-time collaborative adjustment of the counterweights and stability fins, and execution feedback data is generated;

[0024] S5, the actual stability state of the logistics robot is verified in a closed loop according to the execution feedback data; the parameters of the gravity inversion algorithm and the stability prediction model are dynamically optimized by comparing the prediction result with the actual execution effect; at the same time, the closed loop verification result is fed back to the multi-source sensor data processing flow to form a perception-decision-execution closed loop control system; finally, real-time data dashboards for stability regulation and control of the logistics robot are generated.

[0025] The working principle and effect of the above technical solution are as follows:

[0026] The robot motion and goods data are comprehensively collected through multiple source sensors, the completeness of data coverage is improved, unlike a single sensor which is prone to lack of key information, the balance state misjudgment caused by one-sided data is reduced, and a reliable foundation is laid for subsequent analysis;

[0027] The Kalman filter is used to eliminate sensor noise and time synchronization error, and the gravity inversion is performed in combination with the three-dimensional model of the goods and the pressure distribution, so that the calculation accuracy of the dynamic gravity center position of the goods is improved (can be accurate to millimeter level), the influence of interference factors on stability analysis is reduced, and improper adjustment caused by inaccurate gravity calculation is avoided, so that the balance judgment is more reliable;

[0028] The stability feedforward prediction is performed in combination with the path planning information, the overturn risk of the future time window can be evaluated in advance, the goods damage and robot failure caused by sudden imbalance are reduced, and the safety of logistics transportation is improved, so that problems do not need to be handled passively;

[0029] The position of the counterweight block and the action of the stabilizing fin are controlled through a cooperative strategy, the response speed of the balance adjustment is improved, the robot can be quickly stabilized in the face of complex scenes such as sharp bends and steep slopes, the stability in complex working conditions is enhanced, and the imbalance probability is reduced;

[0030] The closed-loop verification optimizes the algorithm parameters and generates real-time data dashboard, which on the one hand enhances the adaptability of the system to different goods weights and transportation scenes, reduces the workload of manual repeated calibration, and on the other hand improves the monitoring efficiency of the operation and maintenance personnel, reduces the time cost of fault troubleshooting, and greatly improves the running stability and transportation reliability of the robot, which is very suitable for actual needs such as warehouse and dynamic logistics.

[0031] In an embodiment of the present application, the S1 comprises:

[0032] S11, based on the chassis structure of the logistics robot, the goods bearing area and the dynamic operation scene (such as warehouse transportation and goods shelf carrying), in combination with the core needs of the sensor system to monitor the goods and the robot state in real time, the specific installation position of the acceleration sensor (chassis center, monitoring three-dimensional motion), the gyroscope sensor (coaxial with the acceleration sensor, capturing angular velocity), the pressure sensor array (full coverage of the bearing platform, collecting pressure distribution), and the visual sensor (top + chassis side, obtaining the environment and the goods posture) is planned, the monitoring range of each sensor is determined, and a multi-source sensor deployment scheme is formed;

[0033] S12, according to the deployment scheme, the sensor type meeting the accurate measurement requirement (such as acceleration sensor accuracy ±0.01g, pressure sensor resolution 0.1kPa, visual sensor frame rate ≥20fps) is selected, and the communication compatibility of the sensor and the robot control system (such as supporting CAN bus protocol) is verified, to obtain a multi-source sensor kit suitable for the logistics robot;

[0034] S13, after fixing the sensors in place according to the deployment plan, calibrate the sensors, including: the acceleration sensor calibrates zero drift and sensitivity through the standard value of gravitational acceleration (9.81 m / s²); the pressure sensor loads different weights through standard weights to establish a linear correspondence between pressure value and output signal; the visual sensor corrects lens distortion through a standard calibration board to ensure that the measurement error of each sensor is controlled within the allowed range, generating a calibrated multi-source sensor system;

[0035] S14, according to the real-time requirements of robot dynamic operation (such as millisecond level response requirements), set the collection frequency of each sensor (acceleration / gyroscope 100 Hz, pressure sensor 50 Hz, visual sensor 20 frames / second), configure the data collection trigger conditions (such as automatically starting when the robot moves, reducing the frequency when it is stationary), after starting the collection system, real-time acquisition of robot motion data (acceleration, angular velocity), cargo pressure distribution data, environment and cargo visual image data, get real-time collected multi-source raw data;

[0036] S15, structure and package the real-time collected multi-source raw data according to the format of sensor type-collection timestamp (accurate to milliseconds)-data value, and label the preliminary operation state of the robot at the time of data collection (such as path turning in warehouse handling), generate a multi-source sensor raw data set.

[0037] The working principle and effect of the above technical scheme are:

[0038] By combining the robot structure, operation scene and monitoring requirements to plan the sensor position and monitoring range, the pertinence and rationality of sensor deployment are improved, unlike blind installation, which is easy to have monitoring blind area, reducing the rework amount due to improper position, so that the sensor can accurately cover the key monitoring points of robot motion and cargo state;

[0039] According to the scheme, high-precision sensors are selected and communication compatibility is verified, which improves the adaptability of sensors and robot control system, avoids the problem of insufficient sensor precision or failure to connect to the system, reduces the cost of equipment replacement due to wrong model selection and poor compatibility, and ensures the basic quality of collected data;

[0040] Through special calibration of the sensor, the measurement accuracy of each sensor is greatly improved, the error is controlled within the allowed range, and the subsequent data distortion caused by large sensor error is reduced, providing reliable raw data for subsequent center of gravity calculation and stability analysis;

[0041] According to the real-time requirement, the collection frequency is set, and the trigger condition is configured, so that the timeliness of data collection in dynamic operation is ensured, unnecessary high-frequency collection in static state is avoided, redundant data generation is reduced, and the storage pressure and the workload of subsequent data processing are reduced.

[0042] According to the type-time stamp-data value encapsulation data and the job state is marked, the structuring degree and the readability of the original data are improved, the time for combing the chaotic data format is saved, the corresponding job scene of the data can be quickly associated, the time cost of data arrangement and understanding is reduced, and the overall data processing efficiency is improved.

[0043] An embodiment of the present application, the S2, comprises:

[0044] S21, the generated multi-source sensor original data set is subjected to data cleaning, and a cleaned multi-source data set is obtained; due to the difference in collection frequency of different sensors, data time dislocation (affecting the accuracy of the center of gravity calculation) may be caused, a unified clock of a robot control system is used to align the time stamp of the cleaned multi-source data set, the collection delay of a visual image is corrected through an interpolation algorithm (for example, the image time difference is compensated according to the moving speed of the robot), the time deviation of the data of each sensor is controlled within 10 ms, and a time-synchronized multi-source data set is generated;

[0045] S22, the time-synchronized multi-source data set is input into a Kalman filtering algorithm, an acceleration and angular velocity data is used to construct a robot motion state equation, sensor noise (for example, high-frequency interference of an acceleration sensor and drift error of a gyroscope) is eliminated through a filtering prediction-update link, and feature information of pressure data and visual data is fused, so that a fused synchronous data stream is obtained;

[0046] S23, a pre-constructed three-dimensional model (containing parameters such as the density, geometric size and preset center of gravity reference position of the goods) of the current transported goods is called from a robot database, the three-dimensional model parameters are associated with the pressure and visual data in the fused synchronous data stream, the actual placement posture of the goods on the bearing platform is determined, and synchronous data associated with the goods model is obtained;

[0047] S24, based on the synchronous data associated with the goods model, a center of gravity inversion algorithm is used to calculate the pressure center coordinates of the goods , the geometric center coordinates are extracted by identifying the goods contour through a visual image , the pressure center and the geometric center are weighted and fused (for example, the weight of the pressure center of the metal goods is 0.6, and the weight of the geometric center is 0.4) according to the density distribution of the three-dimensional model of the goods, and the dynamic center of gravity position data of the goods is obtained.

[0048] S25, the dynamic center of gravity position data of the goods calculated each time is input into a robot control system Bind with the corresponding collection time stamp (millisecond level), arrange in time sequence, and associate the motion state data (acceleration, angular velocity) of the robot at this moment to form the dynamic center of gravity position-time sequence data of the goods.

[0049] The working principle and effects of the above technical solutions are:

[0050] The original data is cleaned first, and then time synchronization is performed based on a unified clock, and the delay of the visual image is also corrected by interpolation, so that the time deviation of each sensor is controlled within 10ms, which not only improves the time consistency of multi-source data and avoids the problem of asynchronous data caused by frequency difference, but also reduces the error caused by time misalignment during subsequent center of gravity calculation, laying a foundation for accurate analysis.

[0051] The synchronized data is processed by Kalman filtering, which not only eliminates the high-frequency interference of the acceleration sensor and the drift error of the gyroscope, but also fuses the characteristics of pressure and visual data, improves the reliability and integrity of the data, reduces the influence of single sensor noise on subsequent calculation, and makes the obtained synchronized data stream more consistent with the actual running state of the robot.

[0052] The pre-constructed three-dimensional model of the goods is retrieved and associated with the fused data to clearly determine the actual placement posture of the goods, enhance the association between the data and the actual state of the goods, avoid the calculation of the center of gravity deviating from the actual situation due to unclear placement of the goods, and reduce the analysis deviation based on the wrong posture.

[0053] Through the center of gravity inversion algorithm combined with the weighted fusion of the pressure center and the geometric center (such as the weight setting of metal goods), different goods characteristics can be targeted and adapted, the calculation accuracy of the dynamic center of gravity position of the goods is improved, the error caused by relying on single center data (such as only using the pressure center) is reduced, and the accuracy of the center of gravity data is ensured.

[0054] The center of gravity position is bound with the time stamp and motion data to form time sequence data, which improves the time sequence continuity and scene association of the data, and reduces the data arrangement cost when making stability prediction without additional time matching of the motion state at the corresponding moment, and makes the subsequent analysis more smoothly combined with the time dimension.

[0055] In an embodiment of the present application, the S22 comprises:

[0056] From the generated time-synchronized multi-source data set, the core data types required by Kalman filtering are selected, the core data types include robot three-dimensional acceleration data, three-dimensional angular velocity data, goods pressure distribution data and goods visual feature data (such as contour coordinates, posture angle), and the data type-time stamp is classified and arranged to obtain a multi-source data subset to be filtered;

[0057] The acceleration and angular velocity data in the multi-source data subset to be filtered are taken as the core, the state vector (including position, velocity, angular velocity, and attitude angle) is defined in combination with the kinematics model of the logistics robot rigid body (considering the chassis motion damping and the influence of cargo inertia), and the robot motion state equation describing the motion law of the robot is established;

[0058] Based on the sensor calibration data (such as the zero drift error of the acceleration sensor and the drift coefficient of the gyroscope), the process noise covariance matrix Q (quantifying the uncertainty of the motion model) and the measurement noise covariance matrix R (quantifying the sensor measurement error) of the Kalman filter are set, and the initial value of the state vector and the state covariance matrix are initialized , to obtain the initialized Kalman filter parameter set;

[0059] The initialized Kalman filter parameter set is combined with the robot motion state equation, and based on the filtering result of the previous moment, the motion state (position, velocity, and attitude angle) of the robot at the current moment is predicted, and the predicted motion state vector and the predicted state covariance matrix are calculated to preliminarily estimate the current motion state of the robot;

[0060] The pressure data (reflecting the force distribution of the goods on the robot) and the visual data (reflecting the actual attitude of the goods) in the time-synchronized multi-source data subset are taken as the measurement values, and the deviation is calculated with the predicted motion state vector, and the Kalman gain is solved through the Kalman gain formula (K = P , wherein K is the Kalman gain, is the predicted state covariance matrix, is the observation matrix, R is the measurement noise covariance matrix, and H is the observation matrix), and then the state vector and the state covariance matrix are updated to eliminate the interference of the acceleration sensor high-frequency noise, the gyroscope drift error, etc., and the updated robot motion state data is obtained;

[0061] The updated robot motion state data is associated and mapped with the pressure distribution features (such as the pressure center coordinates) and the visual features (such as the cargo contour center point coordinates) in the time-synchronized multi-source data subset to ensure the time consistency and logical correlation of the motion state, cargo pressure, and cargo attitude data, and finally the fused synchronous data stream is generated.

[0062] The working principle and effects of the above technical solution are as follows:

[0063] First, the core data needed for filtering is selected from the time-synchronized data and classified by type-timestamp, which can improve the pertinence of the data, avoid mixing irrelevant data into the filtering process, reduce the influence of redundant data on the calculation efficiency, and focus on key information in the subsequent filtering, without wasting effort in messy data;

[0064] When building the motion equation, the chassis motion damping, the cargo inertia and other practical factors are also considered in combination with the robot rigid body kinematics model, not only relying on the theoretical formula, the model is enhanced to be more consistent with the real motion state of the robot, the estimation deviation caused by the disconnection between the pure theoretical model and the actual operation is reduced, and the motion law description is more reliable;

[0065] The sensor calibration data is used to set the matrix Q and R when initializing the filter parameters, the accuracy of the filter parameters is improved, the problems of poor filter effect and data distortion are reduced, and an accurate foundation is laid for subsequent filter calculation;

[0066] Based on the previous time result, the current motion state is predicted, which can make the state estimation more coherent, not starting from zero every time, enhance the continuity of the motion state change, reduce the mutation error that may occur when calculating the state of the current time alone, and make the preliminary estimation result more stable;

[0067] Using pressure and visual data as measurement values to calculate deviation and update state with Kalman gain can accurately eliminate the high-frequency noise of the acceleration sensor and the drift error of the gyroscope, improve the reliability of the robot motion state data, reduce the influence of the sensor itself interference on the data, and make the final obtained motion state data more real;

[0068] Finally, the motion state data and the pressure and visual features are associated and mapped to ensure consistency in time and logic, improve the integrity of the fused data stream, and reduce the time required to align different types of data during subsequent processing, reduce the subsequent analysis lag caused by data mismatch, and make the entire data processing process smoother.

[0069] In an embodiment of the present application, the S24 comprises:

[0070] From the synchronized data of the associated cargo model generated by S23, the key data required for the center of gravity inversion is selected, including the real-time pressure value of the pressure sensor array and the corresponding sensor coordinates, the outline pixel data of the cargo visual image, the density distribution matrix and geometric size parameters (length, width and height) of the cargo three-dimensional model, which are classified and arranged according to the data type-computation dimension (X / Y axis) to obtain the center of gravity inversion core calculation data set;

[0071] Based on the pressure sensor array data in the center of gravity inversion core calculation data set, the physical coordinates of each pressure sensor on the robot bearing platform are first calibrated Then, based on the pressure value weighting method, the pressure distribution center of the cargo on the bearing platform is calculated to eliminate the random error of a single sensor and obtain the cargo pressure center coordinates ;

[0072] For the core calculation data set of the gravity inversion of the visual image data of the goods, image preprocessing (grayscale, edge enhancement, contour extraction) is performed first to identify the complete contour of the goods in the image; then, combined with the geometric features (such as the symmetry plane and the center point reference) of the three-dimensional model of the goods, the image pixel coordinates are converted into physical coordinates of the bearing platform by the contour minimum circumscribed rectangle method or the contour centroid algorithm to obtain the geometric center coordinates of the goods ;

[0073] According to the density distribution matrix of the three-dimensional model of the goods in the core calculation data set of the gravity inversion, the high-density area (such as metal parts) and the low-density area (such as plastic packaging) of the goods are divided, and the contribution weight of different areas to the overall gravity center is calculated; the density weighting rule (such as the weight of the pressure center of the high-density area is increased, and the weight of the geometric center of the low-density area is increased) is set to obtain the weighted fusion coefficient of the pressure center and the geometric center (for example, metal goods =0.6, =0.4);

[0074] The pressure center coordinates , the geometric center coordinates , and the weighted fusion coefficient are substituted into the fusion formula: , , and the preliminary position of the dynamic gravity center of the goods is calculated ;

[0075] Combined with the robot motion state (acceleration, angular velocity) in the synchronous data related to the goods model, the physical rationality of the preliminary gravity center position (such as the gravity center should be offset along the motion direction when the robot accelerates, which conforms to the law of inertia) is verified; if there is a deviation (such as exceeding the reasonable range of ±3mm of the preliminary gravity center), the weighted coefficient is corrected based on the motion inertia compensation formula, and the dynamic gravity center position data of the goods is calculated again .

[0076] The working principle and effect of the above technical solution are:

[0077] First, the key data for gravity inversion is selected from the related data, and it is also sorted according to the data type-computation dimension, which can focus on the core information such as pressure, vision, and model parameters in subsequent calculation, improve the pertinence of data use, avoid irrelevant data interference in the calculation process, and reduce the problems of low calculation efficiency and easy errors caused by disordered data;

[0078] When calculating the pressure center, the physical coordinates of the sensor are first calibrated, and then the pressure value weighting method is used for calculation, which can effectively offset the random error of a single sensor, unlike using only single sensor data which is easily affected by accidental deviation, improving the accuracy of the pressure center coordinates of the goods and laying a reliable foundation for subsequent gravity center calculation;

[0079] When processing visual data, pre-processing is done before converting to physical coordinates, and grayscale and edge enhancement can make the cargo profile clearer, avoiding environmental light interference that can cause inaccurate profile recognition. In combination with the geometric features of the three-dimensional model, the conversion to coordinates reduces the error caused by the disconnection between pixel coordinates and actual carrying platform coordinates, and improves the accuracy of the geometric center coordinates.

[0080] According to the division of cargo density distribution and the setting of weighted rules, such as increasing the pressure center weight of metal cargo, the actual gravity center characteristics of different materials can be fitted, the adaptability of gravity center calculation to different cargos is enhanced, and the deviation of gravity center calculation caused by a single weight is reduced.

[0081] Finally, the preliminary gravity center is verified in combination with the motion state of the robot, and the deviation is corrected, such as the gravity center shifting in the direction of motion when accelerating, and the weight is adjusted if it does not conform, which can exclude incorrect results that do not conform to physical laws, improve the rationality of the final dynamic gravity center position data, reduce the situation of calculating an out-of-range gravity center (such as exceeding the carrying platform range), and ensure that the gravity center data can truly reflect the state of the cargo.

[0082] In one embodiment of the present application, the S3 comprises:

[0083] S31, obtain the preset path planning data from the robot control system, extract the key path parameters, the key path parameters include the curvature (R) of the turning section, the slope angle (a) of the uphill and downhill section, the road roughness level (such as flat / bumpy), and the expected driving time of each path section, and arrange them into a path planning parameter table;

[0084] S32, match the dynamic gravity center position-time sequence data generated by S2 with the path planning parameter table according to the time dimension (such as when the robot is expected to enter a sharp turn with a curvature R = 3m, the gravity center data of the time period is retrieved), analyze the change rule of the gravity center position under different path parameters (such as the lateral displacement amplitude of the gravity center when turning sharply, and the longitudinal displacement rate of the gravity center when climbing a steep slope), and obtain the gravity-path correlation feature data set;

[0085] S33, the logic based on the judgment of the overturning moment wherein, represents the total weight; the risk features: gravity center displacement , wherein, is the ideal lateral reference position of the gravity center of the cargo under the stable state of the logistics robot, which is the reference origin for judging whether the gravity center is laterally displaced, is the ideal longitudinal reference position of the gravity center of the cargo under the stable state of the logistics robot, and Corresponding, is the reference origin of judging whether the center of gravity is shifted in the front-back direction or longitudinal direction, the center of gravity shift rate , the coupling coefficient of path parameters and the center of gravity shift (for example, the greater the curvature, the higher the lateral shift coupling coefficient), forming a set of stability risk feature indexes;

[0086] S34, based on the stability prediction model trained by the dynamic adjustment algorithm (using LSTM neural network, inputting historical center of gravity shift, path parameters and overturning case data), the model input layer is adapted to the set of stability risk feature indexes, and the future 1 to 3 seconds are set as the prediction time window (to meet the demand for rapid response), and the loading and initialization of the dynamic stability prediction model are completed;

[0087] S35, input the stability risk feature index set into the initialized prediction model, the model analyzes the correlation between the center of gravity shift trend and the path parameters, calculates the overturning moment (Mt) and the theoretical restoring moment (Mr) at each prediction time node, and quantitatively outputs the overturning risk probability at each time in the future (for example, the risk probability is 8% at t1=1s, and the risk probability is 15% at t2=2s) based on the criterion that Mr≥Mt is stable, and obtains the overturning risk probability time series data;

[0088] S36, grading the overturning risk probability time series data (low risk <10%, medium risk 10%-30%, high risk >30%), marking the key influencing factors corresponding to each risk level (for example, medium risk: sharp turn causes lateral center of gravity shift ), encapsulating in the format of prediction time point-risk probability-risk level-impact factor, and generating stability feedforward prediction result data.

[0089] The working principle and effect of the above technical solution are:

[0090] By extracting path key parameters and arranging them into a table, the core information such as curvature, slope and driving time is well organized, the readability and retrieval efficiency of path data are improved, and the time cost of parameter query is reduced;

[0091] Matching the center of gravity time series data and the path parameters according to the time dimension, and analyzing the center of gravity change law under different paths (such as the lateral shift amplitude when turning sharply), enhances the correlation between the center of gravity shift and the path characteristics, avoids one-sided analysis caused by the disconnection between the two, and reduces the error of stability judgment based on isolated data;

[0092] Based on the overturning moment logic, the risk features are extracted, the shift amount, shift rate and coupling coefficient are closely related to the stability judgment, the pertinence of the risk index is improved, irrelevant features are avoided, the invalid data interference on the prediction result is reduced;

[0093] The trained LSTM model is used to adapt the features, and a 1 to 3 second prediction window is also set, which improves the adaptation of the model to the current data, meets the rapid response demand, reduces the problem of not timely adjustment caused by prediction lag, and makes the prediction keep up with the dynamic operation rhythm of the robot;

[0094] The risk probability is obtained by calculating the overturning moment and the restoring moment, the risk is changed from fuzzy judgment to specific value, the accuracy of risk evaluation is improved, and subjective deviation caused by experience in determining the risk is avoided;

[0095] The risk is also marked with influencing factors (such as deviation caused by sharp bending), the prediction result is clear and easy to understand, the guidance of the generated subsequent adjustment instruction is improved, the adjustment direction caused by the fuzzy result is reduced, and a foundation for accurate adjustment is laid.

[0096] In an embodiment of the present application, the S32 comprises:

[0097] From the cargo dynamic barycenter position-time sequence data generated by S2, the timestamp (accurate to milliseconds)-cargo barycenter coordinates are extracted - the robot motion state (acceleration ax / ay, angular velocity ) core field, forming a standardized barycenter time sequence data set; meanwhile, from the path planning parameter table generated by S31, the path segment number-path parameter (curvature R / slope α / surface level)-predicted driving start time t start-predicted driving end time t end field is extracted, forming a standardized path parameter data set;

[0098] Based on the unified clock of the logistics robot control system, the time alignment accuracy (±5ms) is determined, the unified time identifier (such as Unix timestamp) is added to each timestamp of the standardized barycenter time sequence data set and each t start-t end time interval of the standardized path parameter data set, and the barycenter time sequence data with the unified time identifier and the path parameter data with the unified time identifier are generated;

[0099] Through a time interval matching algorithm, the barycenter time sequence data with the unified time identifier and the path parameter data are associated, if the timestamp of the barycenter data falls within the [t start, t end] interval of a path segment, the barycenter data is bound with the parameters (such as R=3m, α=8°) of the corresponding path segment, and a preliminary association record of timestamp-barycenter coordinates-motion state-path parameters is formed; for the cross-path segment timestamp (such as t=7ms in the sharp bending segment of t start=5ms-t end=10ms), the linear interpolation method is used to confirm the path parameter attribution, and the time-matched barycenter-path preliminary association data set is generated;

[0100] The time-matched barycenter-path preliminary association data set is subjected to effectiveness screening, and abnormal data (such as Data that exceeds the chassis width range, null values ​​with missing path parameters, and data that contradicts the motion state and path parameters (such as high angular velocity data corresponding to a straight path) are retained. Complete and logically consistent association records are preserved to obtain an effective center of gravity-path association dataset.

[0101] For the effective center-of-gravity-path association dataset, a classification analysis is performed according to the path parameter type. The classification analysis includes: ① Turning sections: with a fixed slope α=0°, the lateral center-of-gravity offset is compared under different curvatures R (e.g., R=2m / R=5m / R=8m). Calculate the same R value ① Mean and fluctuation range; ② Steep slope section: with a fixed curvature R→∞ (straight road), compare the longitudinal center of gravity offset rate under different slopes α (e.g., α=5° / α=10° / α=15°). Record α and The changing trend; ③ Bumpy road surface: With R and α fixed, analyze the influence of road surface grade on the stability of center of gravity shift (e.g., under bumpy road surface). The fluctuation amplitude is twice that of a smooth road surface, generating a path parameter-center of gravity change pattern analysis report;

[0102] The qualitative patterns in the path parameter-center of gravity variation analysis report are transformed into quantitative features. Feature weights are assigned to different path parameters (e.g., the weight of curvature R is ωR=0.6, and the weight of slope α is ωα=0.4). The characteristic values ​​of the path parameters are then calculated. ( For maximum design curvature, (for the maximum design slope), while also considering the center of gravity change index The feature values, standardized to the [0,1] interval, are integrated according to the format of path segment-time interval-path parameter feature value-centroid change feature value-association confidence, and finally generate centroid-path association feature dataset.

[0103] The working principle and effects of the above technical solution are as follows:

[0104] First, extract the core fields of the center of gravity time series and path parameters, organize them into a standardized dataset, and clarify the key information to improve the data's organization. This way, when performing time matching later, you don't have to search through scattered data, reducing the time cost of data sorting.

[0105] Using a unified clock and time marker, and setting an alignment accuracy of ±5ms, the time base of the center of gravity data and the path data are completely consistent, reducing the matching deviation caused by inconsistent time bases.

[0106] The cross-section data is processed through time interval matching and linear interpolation, the gravity data and corresponding path parameters are accurately bound, the correlation of the two is enhanced, the situation that the timestamp of the cross-path section cannot be found is reduced, and the preliminary associated data is more complete;

[0107] The abnormal, null and contradictory data are screened out, and the complete and logically consistent records are left, the quality of the associated data is improved, the invalid data is avoided to interfere with subsequent regular analysis, and the analysis deviation caused by data problems is reduced;

[0108] The gravity change is analyzed according to the path type (such as lateral offset for sharp bend and longitudinal speed for steep slope), the influence law of different path parameters on the gravity is clearly found out, the intuitiveness of the law is enhanced, and the trouble of not finding the key when extracting the risk feature is reduced;

[0109] The qualitative law is converted into quantitative features and integrated, the associated information of the path and the gravity is changed into usable feature data, the usability of the data is improved, the problem that the qualitative information is not used in subsequent model input is avoided, and a better adaptive foundation is laid for stability prediction.

[0110] One embodiment of the application, the S34 comprises:

[0111] The LSTM neural network model file (including model structure, training weight and historical training log) trained based on the dynamic adjustment algorithm is called from the local database or cloud model warehouse of the logistics robot, the model has obtained a pre-training dynamic stability prediction model file through a large number of gravity offset-path parameter-rollover case samples training (such as stability data under different goods weight and path type);

[0112] The configuration information of the pre-training model file is read through a model analysis tool, the feature dimension of the model input layer (such as 3 types of features including gravity offset, offset rate and path coupling coefficient, a total of 6 dimensions), the data format (such as float32 type and normalized range [0, 1]) and the prediction result type of the output layer (such as single output as rollover risk probability value) are determined, and a model input and output specification document is generated;

[0113] According to the model input and output specification document, the stability risk feature index set generated by S33 is adjusted in format and dimension, and the stability risk feature index set is adjusted in format and dimension. 、 、 、 , the path lateral coupling coefficient and the path longitudinal coupling coefficient are arranged in the required order of the model, the feature values are mapped to the [0, 1] interval through the Min-Max normalization algorithm, the feature dimension is matched with the input layer, the data format meets the model requirements, and the adaptive feature input set is obtained.

[0114] According to the stability adjustment requirement of millisecond-level response, set the future 1 to 3 seconds as the core prediction time window, divide the prediction time step (such as 1 prediction node per 100 milliseconds, a total of 30 prediction time points), and clearly define the path parameters associated with each time node (such as t=0.5 seconds corresponding to the emergency curve section, t=2 seconds corresponding to the steep slope section), and generate a prediction time window parameter table;

[0115] Based on the actual load of the current robot (such as the mass of the goods , the mass of the counterweight ), adjust the model hyperparameters, such as optimizing the learning rate from the default 0.001 to 0.0005 (to avoid model prediction oscillation caused by load changes), and setting the iteration threshold to 50 (to ensure fast convergence); At the same time, initialize the hidden layer state variables of the model (such as the cell state and hidden state of the LSTM unit) to zero vectors, and obtain the optimized model hyperparameters and initial state variables;

[0116] Input the first 5 time step data in the adapted feature input set into the model for single forward propagation calculation, check whether the output overturn risk probability value is within a reasonable range (0-100%) and whether the calculation process is error-free (such as dimension mismatch, data type error), and if the verification is passed, the model initialization is successful, and finally the dynamic stability prediction model is obtained.

[0117] The working principle and effect of the above technical solution are:

[0118] Retrieve the pre-trained model trained by a large number of center of gravity offset-path parameter-overturn case, without building and training from scratch, which improves the basic prediction accuracy of the model and reduces the time cost of model training, avoiding repeated trial and error and wasting resources;

[0119] Parse the model configuration to generate input and output specification documents, clearly list the key requirements such as feature dimension and data format, reduce the difficulty of not knowing the requirements during subsequent data adaptation, and reduce the problems of dimension mismatch and format error caused by unclear model requirements;

[0120] Adjust the feature index set according to the specifications and do Min-Max normalization, so that the data fully meets the model input requirements, enhances the adaptability of data and model, avoids the situation that the model cannot be calculated due to unordered features and out-of-range numerical values, and reduces invalid prediction attempts;

[0121] Set the 1 to 3 second prediction window and 100 millisecond step, which perfectly matches the millisecond-level response requirement, improves the timeliness of the prediction, and can keep up with the dynamic operation rhythm of the robot, reducing the situation that the adjustment is not timely due to prediction lag;

[0122] In combination with the actual load, the hyperparameters are also initialized, the hidden layer state is initialized, the model is adapted to the load bearing condition of the current robot, the adaptability of the model to the real-time working condition is enhanced, the prediction oscillation caused by the load change is reduced, and it is ensured that the model can quickly converge;

[0123] The first 5 time steps data are used for verification, the probability range and error are checked, the problems are checked in advance, the reliability of model initialization is improved, the risk of sudden error in subsequent formal prediction is reduced, and the subsequent stable prediction can proceed smoothly.

[0124] In an embodiment of the present application, the S35 comprises:

[0125] The stability risk feature index set generated by S33 is split according to the prediction time window of 1 to 3 seconds, 1 node per 100 milliseconds set by S34, and each prediction time node (such as t1=0.1s, t2=0.2s…t30=3.0s) corresponds to a group of feature data (including the node 、 、 、 , path lateral coupling coefficient, path longitudinal coupling coefficient) to form a feature input subset divided by the prediction time node;

[0126] The feature input subset divided by the time node is input into the dynamic stability prediction model initialized by S34 in sequence, the time sequence memory capability of the LSTM neural network is used to analyze the correlation between the gravity center offset change trend and the path parameter change between adjacent time nodes, for example, whether the growth rate of the lateral gravity center offset of t1=0.1s (straight path, R→∞) to t2=0.2s (entering a sharp bend, R=3m) matches the historical training rule, and the gravity-path correlation degree data of each time node is output.

[0127] Based on the overturning moment judgment logic, the total mass M (including the mass of the robot itself + the mass of the goods + the mass of the counterweight ) is extracted from the feature input subset of each prediction time node (the horizontal distance of the gravity center from the center line of the chassis is converted from 、 in combination with the path parameters (such as the road disturbance coefficient k needs to be superimposed on the bumpy road) of the node), and the formula is calculated to obtain the overturning moment time sequence data (Mt1, Mt2, …, Mt30) of each prediction time node.

[0128] Based on the restoring moment formula ( ), the mass of the counterweight , the reference moving distance of the counterweight The robot tilt angle θ (calculated from acceleration sensor data, θ = arcsin (ax / g) ) is substituted into the formula to calculate the theoretical restoring moment time series data (Mr1, Mr2, …, Mr30) of each predicted time node. If the predicted node has a stable fin deployment requirement (such as a high-risk path), the fin restoring moment Mrfin (Mrfin = kfin x θ, kfin is the fin angle coefficient) is superimposed to obtain the theoretical restoring moment time series data (Mr1, Mr2, …, Mr30) of each predicted time node.

[0129] The stability probability Pst of each predicted time node is calculated using Mr≥Mt as the stability criterion. If Mr≥Mt, Pst = 1 - (Mt / Mr) x 100% (e.g. Mr = 2Mt, Pst = 50%); if Mr < Mt, Pst = 0%; then the risk probability Prisk of each node is obtained through Prisk = 100% - Pst (e.g. t1 = 0.1s, Prisk = 8%, t2 = 0.2s, Prisk = 10%), and the overturning risk probability data of each predicted node is generated;

[0130] The timestamp (t1-t30) - overturning risk probability (Prisk1-Prisk30) - corresponding Mt / Mr value of each predicted time node is arranged in chronological order, and abnormal fluctuation values (such as nodes with Prisk exceeding the range of 0-100% need to be recalculated) are removed, and finally the overturning risk probability time series data is obtained.

[0131] The working principle and effect of the above technical solution are:

[0132] The characteristic index set is split according to the predicted time node, so that each node has exclusive center of gravity offset and path coupling coefficient data, improving the matching degree of data and predicted nodes, reducing the problem of data corresponding to wrong nodes in subsequent calculation, and providing accurate basis for analysis of each time point;

[0133] The LSTM time series memory analysis is used to analyze the correlation between adjacent nodes, such as whether the center of gravity offset and curvature change match historical rules when looking at straight roads and sharp turns, which can enhance the relevance judgment of the center of gravity and path change, and avoid one-sided analysis caused by looking at a single node in isolation;

[0134] When calculating the overturning moment, the total mass and road disturbance coefficient are combined, not just the basic formula, which is more consistent with the actual robot operation scene, improving the accuracy of moment calculation and reducing the situation where the calculated moment is too different from the actual situation due to ignoring road factors;

[0135] When calculating the restoring moment, the contribution of the fin is superimposed, which conforms to the double stability mechanism of counterweight + fin, enhances the completeness of the moment calculation, and does not miss the effect of the stabilizing fin, reducing the risk of misjudgment caused by missing the moment calculation;

[0136] Using formulas to quantify the probability of risk transforms the vague notion of stability into a concrete value, improving the accuracy of risk assessment and avoiding subjective biases based on experience.

[0137] Finally, the time series data was processed and outliers were removed to ensure data continuity and within a reasonable range. This improved the reliability of the final risk time series data, reduced the interference of outliers on subsequent S36 risk classification, and made the classification results more reliable.

[0138] In one embodiment of the present invention, step S4 includes:

[0139] S41. Decompose the generated stability feedforward prediction data and extract the core adjustment information: center of gravity offset direction (lateral / longitudinal), offset magnitude ( , Specific numerical values), overturning risk level (low / medium / high), and expected risk occurrence time are combined with the dual adjustment mechanism of counterweight and stabilizing fins to clarify the adjustment priority (e.g., high risk requires simultaneous adjustment of both counterweight and fins, while low risk only requires fine adjustment of the counterweight), thus obtaining the basis for adjustment decisions;

[0140] S42, Design based on adjusting the center of gravity by moving the counterweight along the slide rail ( Based on the direction and magnitude of the center of gravity offset in the adjustment decision criteria, calculate the distance the counterweight needs to move. , , , To adjust the gain factor, it is dynamically set according to the robot's total mass M, for example, when M = 500 kg. =1.2), moving direction (e.g., if the lateral offset is to the right, the counterweight moves to the left), moving speed (20mm / s for high risk, 10mm / s for medium risk), to obtain the counterweight adjustment parameter set;

[0141] S43. Based on the principle of generating restoring torque by deploying stabilizing fins, fin adjustment parameters are set according to the overturning risk level: at low risk, the fins retract (deployment angle 0°); at medium risk, the lateral fins in the corresponding offset direction deploy 30°-45° at a deployment speed of 5° / s; at high risk, the lateral fins deploy 45°-60° and the rear fins deploy 20°-30° at a deployment speed of 10° / s. The fin numbers (left / right / rear) and action sequences are clearly defined to obtain the stabilizing fin adjustment parameter set. The counterweight adjustment parameter set and the stabilizing fin adjustment parameter set are integrated and encapsulated into an instruction format according to the robot actuator's communication protocol (such as CAN bus), including instruction ID, adjustment parameters, and execution time limit (e.g., high-risk instructions require a response within 0.5 seconds). A checksum is added to avoid instruction transmission errors, generating a counterweight and stabilizing fin coordinated adjustment instruction.

[0142] S44, send the coordinated regulation instruction to the actuator of the robot (motorized lead screw: control the movement of the counterweight; servo motor: control the action of the stabilizing fin), monitor the instruction receiving state in real time (ensure that the actuator successfully receives), drive the actuator to act according to the instruction parameters, record the instruction sending time and the execution starting time, and obtain instruction execution starting confirmation information;

[0143] S45, through the position sensor (monitoring the actual movement position of the counterweight) installed on the counterweight slide rail and the angle sensor (monitoring the actual expansion angle of the fin) of the fin rotating shaft, collect the actual action data of the actuator, compare the deviation of the actual value and the instruction parameters (such as the difference between the actual movement distance of the counterweight and the theoretical value), record the execution completion time and the action stability (such as whether the fin expansion is smooth), and generate execution feedback data.

[0144] The working principle and effects of the above technical solutions are:

[0145] By disassembling the prediction result to extract core regulation information, and by clearly defining the regulation priority of high / medium / low risk (such as adjusting the counterweight and the fin at the same time for high risk), the regulation action can accurately correspond to the risk situation, the pertinence of regulation is improved, and resource waste or insufficient regulation is avoided.

[0146] When calculating the counterweight parameters according to the center of gravity offset, the dynamic gain factors kx and ky are used to adapt to the total mass M of the robot, which are not fixed parameters, so that the movement distance of the counterweight can exactly match the offset demand, the accuracy of the counterweight regulation is improved, and the deviation caused by parameter rigidity is reduced.

[0147] By setting fin parameters according to risk levels (such as multiple fin coordinated expansion for high risk), and by integrating instructions with check codes and packaging according to protocols, the adaptability of fin regulation to risk is improved, the reliability of instruction transmission is improved, the execution error caused by instruction loss and missing parameters is reduced, and the problem that the fin action cannot keep up with high risk is avoided.

[0148] Monitoring the instruction receiving state and recording the starting time can ensure that the actuator actually receives the instruction before acting, reduce the regulation delay caused by instruction loss, avoid the situation of waiting for regulation without receiving the instruction, and ensure that the regulation action starts in time.

[0149] Collecting actual action data and comparing deviations can help to understand the regulation execution effect in real time, and enhance the controllability of the execution process.

[0150] In one embodiment of the present application, the S5 comprises:

[0151] S51, after the completion of the S4 execution mechanism action, start the multi-source sensor system (the same as the calibrated sensor in S1), re-collect the stability state data of the robot: acceleration / gyroscope sensor monitors angular velocity fluctuation (stability threshold <0.5° / s), pressure / visual sensor recalculates the dynamic center of gravity position of the goods (verify whether it returns to the reference range), while recording the current path parameters and working state, get the actual stability state data of the robot;

[0152] S52, compare the stability feedforward prediction result data in S3 with the actual stability state data collected in S5, calculate the core index deviation (such as predicted center of gravity deviation 15mm, actual deviation 5mm, deviation 10mm; predicted risk probability 15%, actual risk probability 5%, deviation 10%), and analyze the deviation reason (such as unreasonable adjustment gain factor, prediction model not adapted to current goods characteristics), form a prediction and actual deviation analysis report;

[0153] S53, based on the deviation analysis report, optimize the key parameters, including: for the center of gravity inversion algorithm, adjust the weighting coefficient of the pressure center and the geometric center (such as reducing the weight of the pressure center when the pressure data fluctuation is large); for Kalman filter algorithm, optimize the process noise and measurement noise covariance matrix (improve data fusion accuracy); for dynamic stability prediction model, add this prediction-actual data pair to update the model weight, get the optimized algorithm and model parameters;

[0154] S54, feedback the optimized algorithm and model parameters to the initial link of the multi-source sensor data processing flow (such as sensor collection parameters, data fusion algorithm), adjust the sensor collection frequency (such as increase the acceleration sensor frequency to 120Hz when the deviation is large) or recalibrate the sensor, perfect the closed-loop logic of sensor perception-data processing-stability prediction-regulation decision-execution feedback-optimization perception, get the updated perception-decision-execution closed-loop control system;

[0155] S55, collect the whole process data of S1 to S5 (multi-source original data, fusion data flow, center of gravity time series data, prediction result, regulation instruction, execution feedback, deviation analysis, optimization parameter), structure storage according to time axis+data type, label key node data (such as whole process data corresponding to high-risk regulation instruction), form the whole process data set of logistics robot sensor data processing;

[0156] S56, based on the full process data set, a visual real-time board is constructed, and the visual real-time board comprises four modules: ① a state monitoring area (a dashboard displays a current gravity center position, angular velocity, risk level, etc.); ② an adjustment instruction area (a list displays historical and current adjustment instructions and execution deviations); ③ a stability trend area (a line chart displays gravity center offset and risk probability changes in the last 10 minutes); and ④ a system optimization area (algorithm parameter optimization records and deviation rate downward trend are displayed), and a stability regulation real-time data board of the logistics robot is formed.

[0157] The working principle and effects of the above technical solutions are as follows:

[0158] The sensor calibrated by S1 reacquires stability data, which can ensure the accuracy of actual state data, maintain consistent standards with previous data, and reduce comparison deviations caused by different sensor accuracies.

[0159] Comparing the prediction results with the actual data and finding reasons by calculating deviations can accurately locate the problem source, avoid blind parameter optimization, reduce trial-and-error costs, and make the optimization direction more clear.

[0160] Targeted optimization of algorithm and model parameters (such as gravity center inversion weight adjustment and Kalman filter covariance matrix) can directly improve the accuracy of the core algorithm, reduce previous deviations (such as gravity center offset deviation from 10 mm to 3 mm), and make subsequent data processing and prediction more accurate.

[0161] Feedback of optimized parameters back to the initial link and improvement of closed-loop logic can enable the system to iteratively optimize itself, enhance adaptability to different goods and scenes, reduce the workload of manual repeated calibration, and eliminate the need for manual parameter adjustment every time the goods are changed.

[0162] Storing full process data and labeling key nodes improves data traceability, and when high-risk adjustment problems are checked later, corresponding full process data can be quickly found.

[0163] The real-time board displays state, instruction and trend in a modular manner, which is intuitive and clear, and operation and maintenance personnel can quickly grasp the robot situation, improve monitoring and troubleshooting efficiency, and reduce problem processing delays caused by information confusion.

[0164] An embodiment of the present application is a sensor system of a logistics robot, the system comprising:

[0165] One or more processors; a memory for storing one or more programs; wherein when the one or more programs are executed by the one or more processors, the one or more processors implement the data processing method of any one of the above.

[0166] Obviously, many modifications and variations of the present application are possible in light of the above teachings. It is, therefore, to be understood that within the scope of the appended claims and their equivalents, the application can be practiced otherwise than as specifically described.

Claims

1. A method for processing sensor data of a logistics robot, characterized in that, The method comprises: S1, collecting multi-source data of the logistics robot in the moving process in real time through a multi-source sensor, and generating a multi-source sensor original data set; S2, data fusion preprocessing is performed on the multi-source sensor original data set to generate a fused synchronous data stream; based on the synchronous data stream, the gravity inversion algorithm is used to combine the three-dimensional model of the goods and the pressure distribution characteristics to perform weighted fusion on the goods pressure center and the geometric center, and the dynamic gravity center position data of the goods is obtained by inversion, and the corresponding timestamp information is recorded synchronously to generate the dynamic gravity center position and time sequence data of the goods; S3, according to the dynamic gravity center position and time sequence data of the goods, combining the preset path planning information of the logistics robot, a dynamic stability prediction model is used for stability feedforward prediction; generate stability feedforward prediction result data; including: S31, obtain the preset path planning data from the robot control system, extract the path key parameters, the path key parameters include the curvature of the turning section, the slope angle of the uphill and downhill section, the road roughness grade and the expected driving time of each path section, and arrange them into a path planning parameter table; S32, match the dynamic gravity center position and time sequence data generated in S2 with the path planning parameter table according to the time dimension, analyze the change rule of the gravity center position under different path parameters, and obtain the gravity center and path correlation feature data set; S33, based on the logic of overturning moment judgment, extract the risk features from the gravity center and path correlation feature data set, the risk features include the gravity center offset, the gravity center offset rate and the coupling coefficient of the path parameters and the gravity center offset, and form a stability risk feature index set; S4, based on the stability feedforward prediction result data, generate the collaborative adjustment instructions of the counterweight and the stability fin through the collaborative adjustment strategy; send the collaborative adjustment instructions to the actuator of the logistics robot to realize the real-time collaborative adjustment of the counterweight and the stability fin, and generate the execution feedback data; S5, according to the execution feedback data, the actual stability state of the logistics robot is verified in a closed loop; by comparing the prediction result with the actual execution effect, the parameters of the gravity inversion algorithm and the stability prediction model are dynamically optimized; at the same time, the closed loop verification result is fed back to the multi-source sensor data processing flow to form a perception, decision and execution closed loop control system; finally, the real-time data board of the stability regulation of the logistics robot is generated. 2.The method of claim 1, wherein, The S1 comprises: S11, based on the chassis structure of the logistics robot, the cargo carrying area and the dynamic operation scene, combining the core demand of the sensor system, planning the specific installation position of the multi-source sensor, determining the monitoring range of each sensor, and forming a multi-source sensor deployment scheme; S12, according to the deployment scheme, select the sensor model, and at the same time verify the communication compatibility of the sensor and the robot control system, obtain the multi-source sensor kit suitable for the logistics robot; S13, after fixing the sensor in place according to the deployment scheme, calibrate the sensor to generate a calibrated multi-source sensor system; S14, set the acquisition frequency of each sensor according to the real-time requirement of the dynamic operation of the robot, configure the data acquisition trigger condition, and obtain the multi-source raw data collected in real time after starting the acquisition system; S15, structure and package the multi-source raw data collected in real time, label the preliminary operation state of the robot during data acquisition, and generate a multi-source sensor raw data set. 3.The method of claim 1, wherein, The S2 comprises: S21, data cleaning is performed on the generated multi-source sensor raw data set to obtain a cleaned multi-source data set, time stamp alignment is performed on the cleaned multi-source data set based on a unified clock of a robot control system, and a time-synchronized multi-source data set is generated; S22, the time-synchronized multi-source data set is input into a Kalman filtering algorithm, and the feature information of the pressure data and the visual data is fused to obtain a fused synchronous data stream; S23, a pre-constructed three-dimensional model of the current transported goods is called from a robot database, the three-dimensional model parameters are associated with the pressure and visual data in the fused synchronous data stream, and synchronous data of the associated goods model is obtained; S24, based on the synchronous data of the associated goods model, the gravity inversion algorithm is used to calculate the pressure center coordinates of the goods through the pressure sensor array data, the geometric center coordinates are extracted by identifying the goods contour through the visual image, and the pressure center and the geometric center are weighted and fused by combining the density distribution of the goods three-dimensional model, to obtain the dynamic gravity center position data of the goods; S25, the dynamic gravity center position of the goods calculated each time is bound with the corresponding acquisition time stamp, arranged in chronological order, and associated with the motion state data of the robot at this moment to form the dynamic gravity center position and time sequence data of the goods. 4.The method of claim 3, wherein, The S22 comprises: From the generated time-synchronized multi-source data set, the core data types required by the Kalman filtering are screened out, classified and arranged according to the data types and time stamps, and a multi-source data subset to be filtered is obtained; Taking the acceleration and angular velocity data in the multi-source data subset to be filtered as the core, combining the rigid body kinematics model of the logistics robot, defining the state vector, and establishing the robot motion state equation describing the motion law of the robot; Based on the sensor calibration data, the process noise covariance matrix and the measurement noise covariance matrix of the Kalman filtering are set, and the initial value of the state vector and the state covariance matrix are initialized, to obtain the initialized Kalman filtering parameter set; Combining the initialized Kalman filtering parameter set with the robot motion state equation, predicting the motion state of the robot at the current moment based on the filtering result at the previous moment, and calculating the predicted motion state vector and the predicted state covariance matrix to preliminarily estimate the current motion state of the robot; Taking the pressure data and the visual data in the time-synchronized multi-source data subset as the measurement values, calculating the deviation from the predicted motion state vector, solving the Kalman gain through the Kalman gain formula, and obtaining the updated robot motion state data; The updated robot motion state data is associated and mapped with the pressure distribution features and the visual features in the time-synchronized multi-source data subset, and finally the fused synchronous data stream is generated. 5.The method of claim 3, wherein, The S24 comprises: From the synchronization data of the associated cargo model generated from S23, the key data required for the barycentric inversion is screened out and classified to obtain the barycentric inversion core calculation data set; Based on the pressure sensor array data in the barycentric inversion core calculation data set, first calibrate the physical coordinates of each pressure sensor on the robot bearing platform, and then based on the pressure value weighting method, calculate the pressure distribution center of the cargo on the bearing platform to obtain the cargo pressure center coordinates; For the cargo visual image data in the barycentric inversion core calculation data set, first perform image preprocessing to identify the complete contour of the cargo in the image; Then, combined with the geometric features of the cargo three-dimensional model, the image pixel coordinates are converted into bearing platform physical coordinates by the minimum circumscribed rectangle method or the contour centroid algorithm to obtain the cargo geometric center coordinates; According to the cargo three-dimensional model density distribution matrix in the barycentric inversion core calculation data set, divide the high-density and low-density regions of the cargo, and calculate the contribution weight of different regions to the overall barycenter; Obtain the weighted fusion coefficient of the pressure center and the geometric center; Substitute the pressure center coordinates, geometric center coordinates and weighted fusion coefficient into the fusion formula to calculate the preliminary position of the dynamic barycenter of the cargo; Combine the synchronization data of the associated cargo model to verify the physical rationality of the preliminary barycenter position; If there is a deviation, then based on the motion inertia compensation formula to correct the weighting coefficient, and then calculate the dynamic barycenter position data of the cargo. 6.The method of claim 1, wherein, The S3 further comprises: S34, based on the stability prediction model trained by the dynamic adjustment algorithm, adapt the model input layer to the stability risk feature index set, set the future 1 to 3 seconds as the prediction time window, complete the loading and initialization of the dynamic stability prediction model; S35, input the stability risk feature index set into the initialized prediction model, calculate the overturning moment Mt and the theoretical restoring moment Mr of each prediction time node by analyzing the correlation between the barycenter deviation trend and the path parameters, based on the criterion Mr≥Mt, quantitatively output the overturning risk probability at each time in the future, and obtain the overturning risk probability time series data; S36, grade the overturning risk probability time series data, label the key influencing factors corresponding to each risk level, and generate stability feedforward prediction result data. 7.The method of claim 6, wherein, The S32 comprises: From the cargo dynamic barycenter position and time series data generated from S2, extract the timestamp, cargo barycenter coordinates and corresponding time robot motion state core field to form a standardized barycenter time series data set; At the same time, from the path planning parameter table generated from S31, extract the path segment number, path parameter and predicted driving start time tstart to predicted driving end time t end field to form a standardized path parameter data set; Based on the unified clock of the logistics robot control system, determine the time alignment accuracy, add a unified time identifier, and generate barycenter time series data with a unified time identifier and path parameter data with a unified time identifier; The barycenter time series data with uniform time identifier is associated with the path parameter data through a time interval matching algorithm. If the timestamp of the barycenter data falls within the [tstart, tlast] interval of a path segment, the barycenter data is bound with the parameters of the corresponding path segment to form a preliminary associated record. For the timestamp across the path segment, the linear interpolation method is used to confirm the path parameter attribution, and the preliminary associated data set of the barycenter and path with time matching is generated; The preliminary associated data set of the barycenter and path with time matching is subjected to effectiveness screening to obtain an effective barycenter and path associated data set. According to the effective barycenter and path associated data set, classified analysis is performed according to the path parameter type to generate a path parameter and barycenter change rule analysis report; The qualitative rule in the path parameter and barycenter change rule analysis report is converted into a quantitative feature, the feature weight is assigned to different path parameters, the path parameter feature value is calculated, the barycenter change index is standardized to a feature value in the [0, 1] interval, and integration is performed to finally generate a barycenter and path associated feature data set. 8.The method of claim 1, wherein, The S4 comprises: S41, the generated stability feedforward prediction result data is disassembled, core regulation information is extracted, the mechanism of double regulation of the counterweight block and the stability fin is combined, the regulation priority is determined, and the regulation decision basis is obtained; S42, based on the design of adjusting the barycenter by moving the counterweight block along the slide rail, the distance, direction and speed of movement of the counterweight block are calculated according to the barycenter offset direction and amplitude in the regulation decision basis, and a counterweight regulation parameter set is obtained; S43, the fin regulation parameter is set according to the capsizing risk level, a stability fin regulation parameter set is obtained, the counterweight regulation parameter set and the stability fin regulation parameter set are integrated, and the integrated parameter set is packaged into an instruction format according to the communication protocol of the robot execution mechanism to generate a counterweight and stability fin cooperative regulation instruction; S44, the cooperative regulation instruction is sent to the execution mechanism of the robot, the instruction receiving state is monitored in real time, the execution mechanism is driven to act according to the instruction parameters, the instruction sending time and execution starting time are recorded, and instruction execution starting confirmation information is obtained; S45, the actual action data of the execution mechanism is collected through the position sensor installed on the counterweight block slide rail and the angle sensor of the fin rotating shaft, the deviation between the actual value and the instruction parameter is compared, the execution completion time and action stability are recorded, and execution feedback data is generated. 9.The method of claim 1, wherein, The S5 comprises: S51, after the execution mechanism of S4 is completed, the multi-source sensor system is started, the stability state data of the robot is re-collected, the current path parameters and working state are recorded, and the actual stability state data of the robot is obtained; S52, the stability feedforward prediction result data of S3 and the actual stability state data collected in S5 are compared, the core index deviation is calculated, and the deviation reason is analyzed to form a prediction and actual deviation analysis report; S53, based on the deviation analysis report, the key parameters are optimized, and the optimized algorithm and model parameters are obtained; S54, the optimized algorithm and model parameters are fed back to the initial link of the multi-source sensor data processing flow, the sensor collection frequency is adjusted or the sensor is recalibrated, and an updated perception, decision and execution closed-loop control system is obtained. S55, collect the whole process data of S1 to S5, store the structured data according to the time axis and the data type, mark the key node data, and form the whole process data set of the logistics robot sensor data processing; S56, based on the whole process data set, a visual real-time board is constructed, and a logistics robot stability regulation real-time data board is generated.

10. A sensor system for a logistics robot, characterized in that The system comprises: one or more processors; a memory for storing one or more programs; wherein when the one or more programs are executed by the one or more processors, the one or more processors implement the data processing method of any one of claims 1 to 9.

Images