Collector data monitoring management system and method based on artificial intelligence

Abstract

The invention discloses a collector data monitoring management system and method based on artificial intelligence, relates to the technical field of artificial intelligence, and is applied to collecting body state and article parameter data for preprocessing after a robot receives a grabbing instruction; multi-modal features are extracted based on the preprocessed data, key features are screened through mutual information and a random forest, a random forest regression model is trained, and mapping between vibration features and the maximum safe grabbing weight of a grabbing mode is established; the maximum safe grabbing weight of each mode is dynamically adjusted in real time; reconstructing a three-dimensional model through the depth map, and analyzing article parameter data to obtain potential grabbing points; selecting a grabbing mode and a grabbing point in combination with the weight of the object; a grabbing strategy is adjusted according to the article parameter data, and grabbing is carried out; during placement, a target area is scanned to generate candidate positions, the space utilization rate and the stacking safety are evaluated, weighted scores are comprehensively weighted, and finally the optimal placement position and posture are selected.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence technology, specifically to an artificial intelligence-based data acquisition monitoring and management system and method. Background Technology

[0002] Industrial robots are widely used in fields such as automated production, logistics sorting, and warehouse management to grasp and transport various items. They need to be able to adapt to items of different shapes, sizes, weights, and materials, and complete the grasping and releasing tasks safely and reliably.

[0003] However, existing robotic grasping methods mainly rely on fixed programs or simple sensors, which have significant limitations. Grasping strategies are simplistic, often limited to single-handed operation, and ineffective for heavy, irregular, or unstable objects. Multimodal data is not fully integrated and utilized, and its relationship with grasping weight and safety parameters lacks effective modeling, resulting in an inability to dynamically and accurately adjust strategies. Grasping point selection is often based on simple geometry, ignoring the center of gravity, friction, avoidance of fragile areas, and force closure conditions, while also lacking consideration of grasping force, easily leading to unstable grasping or damage to objects. Robots also face challenges in object placement. Existing technologies often lack the ability to finely plan the placement environment, making it difficult to scan and analyze the target area in real time, calculate the optimal candidate placement point, and calculate and decide on space utilization and stacking safety. Summary of the Invention

[0004] The purpose of this invention is to provide an artificial intelligence-based data acquisition device monitoring and management system and method to solve the problems raised in the prior art.

[0005] To achieve the above objectives, the present invention provides the following technical solution: In a first aspect, the present invention provides a data monitoring and management method for data collectors based on artificial intelligence, comprising: The robot receives instructions to grab an item and place it in a designated location, collects robot body status data and item parameter data, and preprocesses the raw data. Based on the preprocessed data, multimodal feature extraction is performed; mutual information and random forest feature importance are used to select the most sensitive and discriminative key feature subset that distinguishes different vibration levels and associated grasping weights; using the selected key feature subset, a random forest regression model is selected for training to establish a mapping relationship between vibration features and the maximum safe grasping weight of the current robot grasping mode. The trained model is applied to the robotic arm to perform grasping and releasing tasks. Features are extracted in real time and input into the model. The maximum safe grasping weight of each grasping mode in the current state is calculated and dynamically adjusted. A 3D model is reconstructed using a depth map, and the item's parameter data is analyzed to obtain potential gripping points. Based on the obtained potential gripping points and the item's weight, a gripping mode and gripping points are selected. After calculating the gripping mode and gripping point, adjust the gripping mode, gripping point and gripping force according to the item parameter data to grip the item; Real-time scanning and analysis of the target area at a specified location to generate candidate placement locations; evaluation of candidate placement locations and calculation of the overall space utilization rate of the area after placement at each candidate location; and performance of stacking security analysis, including protection of items below and stacking stability. Considering both space utilization and stacking safety, a weighted scoring algorithm is used to select the optimal placement location and orientation.

[0006] In conjunction with the first aspect, in the first implementation of the first aspect of this application, the robot receives an instruction to grasp an item and place it at a designated location, collects robot body state data and item parameter data, and preprocesses the raw data, including: Multimodal sensors are deployed at the end effector and joints of the robot arm to collect real-time robot body status data, including three-axis acceleration, joint angles, motor current and joint temperature; object parameter data, including object weight, three-dimensional dimensions, shape features, fragility and surface material information, are collected through vision sensors, weight sensors and tactile sensors. The raw data is preprocessed; noise, outliers and missing values ​​generated during sensor acquisition are removed or corrected; and data from different sensors and timestamps are time-aligned and spatially registered.

[0007] In conjunction with the first aspect, in the second implementation of the first aspect of this application, the step of performing multimodal feature extraction based on the preprocessed data; using mutual information and random forest feature importance, selecting the most sensitive and discriminative key feature subset that distinguishes different vibration levels and associated grasping weights; and using the selected key feature subset, selecting a random forest regression model for training to establish a mapping relationship between vibration features and the maximum safe grasping weight of the current robot grasping mode, including: Based on the preprocessed robot body state data and item parameter data, multimodal feature extraction is performed; the vibration level is divided into four levels, the mutual information between each feature and the four vibration levels and the grasping weight is calculated, the importance of the feature is evaluated using a random forest model, and all features are ranked by combining the scores of the two methods. According to the set screening criteria, the feature with the highest score is selected from the ranked list to form a subset of key features that can distinguish different vibration levels and associate with the grasping weight. The selected key feature subset is divided into feature vectors for model input. The dataset used to train the random forest regression model consists of multiple sets of samples, including key feature vectors and the corresponding maximum safe gripping weight values ​​for single-handed gripping mode and two-handed gripping mode. Random forests improve the accuracy and robustness of a model by constructing multiple decision trees and combining prediction results. Using a prepared dataset, by optimizing the split nodes and the number of trees in the decision trees, the model learns a non-linear mapping relationship between the key features of the input and the maximum safe grab weight.

[0008] In conjunction with the first aspect, in the third implementation of the first aspect of this application, the step of applying the trained model to the robotic arm to perform grasping and releasing tasks, extracting features in real time and inputting them into the model, and calculating and dynamically adjusting the maximum safe grasping weight for each grasping mode in the current state includes: Collect robot body state data and object parameter data during the robot arm grasping and releasing task, extract key feature subsets in real time, and input them into a trained random forest regression model. Based on the input features, the model predicts the maximum safe grasping weight corresponding to the single-hand grasping mode and the two-hand grasping mode, and grasps the object. Vibration data is continuously monitored during the grasping and releasing process, and the actual vibration level is compared with the expected level. When the actual vibration level is detected to be moderate or severe, a dynamic adjustment mechanism is activated. The maximum safe grasping weight predicted by the grasping mode is multiplied by a preset discount factor as a correction to obtain the actual maximum safe grasping weight. The information recorded in this capture and release task is used as a new data sample to trigger the retraining or incremental update of the random forest regression model.

[0009] In conjunction with the first aspect, in the fourth implementation of the first aspect of this application, the step of reconstructing a 3D model through a depth map, analyzing item parameter data, and obtaining potential grab points includes: The robot uses a structured light camera to acquire multi-frame depth map data of the object from different angles and preprocesses it. Then, using the point cloud processing algorithm in the PCL library, the multi-frame preprocessed depth map data is fused to reconstruct a 3D model and the model is unified to the robot coordinate system. Using model fitting algorithms, the pose of the object in the robot's base coordinate system is calculated; the overall size of the object is calculated, the overall shape category of the object is analyzed, and the geometric features of the object's surface are detected; based on the geometric shape and mass distribution of the 3D model, the position of the object's center of gravity is calculated; the texture features of the object's surface are analyzed to determine smoothness and identify special surface areas; through visual features and pre-loaded object database information, marked or known fragile parts on the object are identified; and weak parts are inferred based on the object's shape and material. Based on the identified geometric features, a series of potential grab points are generated.

[0010] In conjunction with the first aspect, in the fifth implementation of the first aspect of this application, the step of selecting the gripping mode and gripping point based on the obtained potential gripping points and the weight of the item includes: The weight of the item is compared with the maximum safe gripping weight of the robot in the current state of single-hand gripping mode and double-hand gripping mode, which are obtained through dynamic adjustment. When the weight of the item is lower than the maximum safe gripping weight of the single-hand gripping mode in the current state, select the single-hand gripping mode for gripping; from the identified potential gripping points, prioritize areas near the center of gravity of the item, with flat surfaces and not easy to slip as gripping points, while avoiding marked fragile areas and sharp edges, and plan the position and posture that the end effector needs to reach; When the weight of the object is between the maximum safe gripping weight of the single-hand gripping mode and the double-hand gripping mode in the current state, the double-hand gripping mode is selected for gripping; based on the force closure condition and the center of gravity position, the optimal gripping point is selected for each of the two robotic arms from the identified potential gripping points, and the position and posture that the two end effectors need to reach are planned. When the weight of an item exceeds the maximum safe gripping weight in the current two-hand gripping mode, it is determined to be an overload anomaly. An anomaly message is sent and the robot is instructed to skip the item. At the same time, the person in charge is notified to handle the situation manually.

[0011] In conjunction with the first aspect, in the sixth implementation of the first aspect of this application, after calculating the grasping mode and grasping point, adjusting the grasping mode, grasping point, and grasping force according to the item parameter data to grasp the item includes: When the gripping mode is confirmed to be single-handed gripping mode, the minimum gripping force required for the gripping point is determined based on the size, weight, and center of gravity of the item, combined with the surface characteristics of the gripping point. The friction conditions of the gripping point are taken into account, and a gripping force with safety redundancy is set. The smoothness of the item's surface and the fragility of the item are evaluated to determine whether it is necessary to switch to two-handed gripping mode. When the gripping mode is confirmed to be a two-hand gripping mode, based on the size, weight, center of gravity position of the object, and the specific surface features and geometric relationships of the two gripping points, force closure analysis is used to calculate the minimum total gripping force required to maintain stable gripping and the force distribution between the two gripping points; considering the smoothness of the object, gripping force command values ​​with safety redundancy are set for the gripping points; considering fragility, the force applied to each gripping point and the balance between the two forces are adjusted through force closure analysis and force distribution.

[0012] In conjunction with the first aspect, in the seventh implementation of the first aspect of this application, the real-time scanning and analysis of the target area at the specified location generates candidate placement locations; the candidate placement locations are evaluated, and the space utilization rate of the overall area after each candidate location is placed is calculated; a stacking security analysis is performed, including protection of items below and stacking stability, including: The target area specified in the scanning command is used to construct a 3D model of the current item distribution in the area; the position, shape and space occupied by existing items in the area are obtained through real-time scanning, mapped onto a 3D mesh, and the occupied voxels are marked; combined with item parameter data, multiple candidate locations are generated. The candidate locations are evaluated. For each candidate location, the current item is simulated to be placed. The proportion of space utilized in the entire specified area and the number of voxels occupied on the 3D mesh are calculated after placement. The total number of voxels in the space is calculated. The number of voxels occupied before the new item is placed is added to the number of voxels occupied after the new item is placed to obtain the total number of voxels occupied after placement. This is compared with the total number of voxels in the entire specified area to obtain the space utilization rate. When the candidate location is on top of an existing object, the system performs an object protection analysis based on the object parameter data analyzed during the previous grabbing and the current scanned state of the object below. It calculates the impact force generated within a preset time frame during placement and the static pressure on the supporting surface of the object below after placement, based on the weight and drop height of the object to be placed. The system calculates the contact area by simulating the placement posture and object shape. The calculated impact force and static pressure are compared with the known vulnerability thresholds of the object below. Based on the comparison results and the state of the object below, a damage risk index for the object below is generated. The current item to be placed and all the items supporting it below it are considered as a temporary stacking structure, including the mass, center of gravity, shape and contact surface information of each item, and stacking stability analysis is performed. Analyze the new stacking structure state after the item is placed in this position; calculate the total center of gravity of the entire stacking structure, determine the base of the stacking structure on the support surface, and check whether the vertical projection point of the total center of gravity falls within the base range; if it falls within the range, it is statically stable; otherwise, it is statically unstable, and set the corresponding static score respectively. In a simulated environment, different preset thrusts are applied to the stacked structure, and the state of the items and the amplitude of shaking are observed. Score levels are set according to the shaking amplitude threshold to obtain a dynamic score. Different weights are set for the static score and the dynamic score, and the stacking stability score is obtained by adding them together. The stack safety is calculated by weighting the damage risk index of the items below and the stack stability score.

[0013] In conjunction with the first aspect, in the eighth implementation of the first aspect of this application, the comprehensive consideration of space utilization and stacking security, and the selection of the optimal placement position and orientation using a weighted scoring algorithm, includes: Based on the obtained space utilization rate, the damage risk index of the items below, and the stacking stability score, these are regarded as different input signals. According to the priority requirements of the current operation scenario, different weights are assigned to the signals, namely w1, w2, and w3. The comprehensive score W is calculated using a linear weighted scoring algorithm. The scores for each candidate position are multiplied by their corresponding weights and then summed to obtain the overall score W. The formula is as follows: ; Wherein, SU represents space utilization, DRI represents the risk index of damage to the items below, and SSS represents the stacking stability score; compare the comprehensive scores of all candidate positions, select the one with the highest score as the best placement position, determine the placement posture at that position, and place the item.

[0014] Secondly, the present invention provides an artificial intelligence-based data acquisition device monitoring and management system, comprising: Command receiving and data acquisition module: includes command receiving unit, data acquisition unit and data preprocessing unit; wherein, the command receiving unit receives the command to grab the item and place it in the designated position; the data acquisition unit acquires the robot body status data and item parameter data; the data preprocessing unit performs preprocessing operations on the raw data; Model building and dynamic adjustment module: includes a mapping model building unit and a maximum safe gripping weight adjustment unit; wherein, the mapping model building unit extracts key features, uses mutual information and random forest to select feature subsets, trains a random forest regression model, and establishes a mapping relationship between vibration features and the maximum safe gripping weight of each gripping mode; the maximum safe gripping weight adjustment unit dynamically adjusts the actual maximum safe gripping weight under the current state and updates and trains the model based on the recorded data; The 3D perception and gripping point analysis and decision-making module includes a 3D reconstruction unit, a gripping point identification and evaluation unit, a gripping mode decision-making unit, a gripping point selection unit, and a coordinate transformation and planning unit. Specifically, the 3D reconstruction unit reconstructs the 3D model of the object using a depth map; the gripping point identification and evaluation unit analyzes the object's parameter data and the 3D model to identify and evaluate potential gripping points; the gripping mode decision-making unit compares the object's weight with the dynamically adjusted maximum safe gripping weight, selecting either a single-handed or double-handed gripping mode, or determining an overload anomaly; the gripping point selection unit selects the optimal gripping point based on the identified potential gripping points, combined with the object's parameter data and 3D model analysis; and the coordinate transformation and planning unit transforms the gripping point coordinates to the robotic arm coordinate system and plans the position and orientation of the end effector. The grasping strategy optimization and execution module includes a single-hand grasping force calculation and safety assessment unit, a two-hand grasping force calculation and allocation unit, and a grasping execution unit. Specifically, the single-hand grasping force calculation and safety assessment unit calculates the minimum single-hand force, sets safety redundancy, and determines whether to switch modes when the grasping mode is confirmed to be single-handed. The two-hand grasping force calculation and allocation unit calculates the total force and allocated force, sets safety redundancy, and adjusts the force balance when the grasping mode is confirmed to be two-handed. The grasping execution unit controls the robot to perform the grasping action according to the final grasping strategy. Space utilization and stacking safety calculation module: including space utilization unit and stacking safety unit; wherein, the space utilization unit calculates the space utilization of the entire specified target area after the current item is placed in a specific candidate position; the stacking safety unit calculates the damage risk index of the items below and the stacking stability score, and obtains the stacking safety by weighted summation; The placement location selection and placement module includes a priority weight allocation unit and a comprehensive score calculation and placement unit. The priority weight allocation unit assigns different weights to the signal according to the priority requirements of the current operation scenario. The comprehensive score calculation and placement unit calculates and sums the space utilization and stacking security of each candidate location to obtain a comprehensive score, and selects the one with the highest score as the best placement location for placement.

[0015] Compared with the prior art, the beneficial effects of the present invention are: 1. This invention establishes a dynamic mapping relationship between vibration characteristics and the maximum safe gripping weight of the current robot gripping mode.

[0016] 2. Based on the three-dimensional reconstruction technology of depth map, the analysis of item parameters and the dynamic calculation of safe weight threshold, the present invention realizes the intelligent selection of gripping mode and gripping point, and further adjusts and optimizes the gripping mode, gripping point and gripping force according to the item parameter data.

[0017] 3. This invention analyzes the placement area, generates and evaluates candidate placement locations, and selects the optimal placement location and orientation based on a combination of space utilization and stacking safety. Attached Figure Description

[0018] Figure 1 This is a schematic diagram illustrating the steps of a data acquisition monitoring and management method based on artificial intelligence according to the present invention; Figure 2 This is a system architecture diagram of an artificial intelligence-based data acquisition monitoring and management system according to the present invention. Detailed Implementation

[0019] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0020] Example: Figures 1-2 As shown, the present invention provides a technical solution. like Figure 1 A schematic diagram illustrating the steps of an AI-based data acquisition monitoring and management method is provided. This invention offers an AI-based data acquisition monitoring and management method, comprising: Step S100: The robot receives the instruction to grab the item and place it in the designated location, collects the robot's own status data and the item's parameter data, and preprocesses the raw data; Specifically, after the robot receives the instruction to grasp an object and place it in a designated location, it collects data through various types of sensors deployed on the end effector of the robot arm and the various moving joints of the arm; it measures the acceleration values ​​along the X, Y, and Z axes through a three-axis accelerometer; it measures the rotation angle of each joint through a joint angle sensor; it measures the current flowing through the motor through a motor current sensor; and it measures the temperature of the joint area through a joint temperature sensor. Using a visual sensor and a depth camera, images and depth information of an object are acquired. Through image processing and 3D reconstruction algorithms, the object's length, width, height, outline, surface curvature, and whether there are protrusions or depressions are extracted. By analyzing the color, texture, and reflective properties of the images, the material type and its properties are preliminarily identified. A weight sensor measures the weight of the object. A tactile sensor measures the pressure distribution, normal force, and tangential force at the contact point. By analyzing the contact pressure and deformation, the fragility of the object is assessed. Different materials result in different pressure distributions and force feedback, helping to determine which areas on the object are suitable for gripping and which areas should be avoided. The collected raw data undergoes preprocessing, including noise, outlier, and missing value handling. High-frequency noise is often present in sensor signals, which is addressed using digital filtering techniques. Low-pass filtering removes high-frequency noise, while moving average filtering preserves the main signal trend. Band-pass filtering retains components within a known signal frequency range. Wavelet transform is used for simultaneous time-frequency domain analysis to separate noise at different frequencies. Outliers may significantly deviate from the normal range due to sensor malfunctions, strong interference, or accidental collisions. These outliers are detected and processed using statistical methods, discarding data exceeding a preset threshold based on the mean and standard deviation, or correcting with nearby normal values ​​using linear interpolation. Missing data may occur due to temporary sensor failures or communication interruptions. A model prediction method is used, training the model with other relevant sensor data or historical data to predict and fill in missing values. Different types of sensors have different sampling frequencies and inconsistent sampling timestamps. Time synchronization and resampling techniques are employed, using the PTP protocol to record precise timestamps to keep the clocks of all sensors synchronized; the data from all sensors are unified onto a common time baseline; the lowest sampling frequency is selected as the reference, and high-frequency data is downsampled or interpolated to the reference frequency; for low-frequency data, the reference frequency is matched through repetition or interpolation; the goal is to ensure that at any given time point or time window, the corresponding data from all sensors at that moment can be obtained. Sensors are deployed in different physical locations and coordinate systems. In order to conduct comprehensive analysis, coordinate transformation technology is required to transform the data into the same coordinate system. A calibration board is used to accurately calibrate the position and orientation of each sensor relative to the robot's base coordinate system. Using coordinate transformation technology, the transformation relationship obtained from the calibration is used to transform the raw data of each sensor from its own coordinate system to a unified coordinate system.

[0021] In one specific embodiment, on an automated assembly line, a robotic arm is instructed to grasp a specific aluminum alloy part. Immediately after the grasping command is issued, the robot begins collecting data. The IMU sensor on its end effector records the triaxial accelerations at timestamp t=0.003s as X: 0.12 m / s², Y: -0.05 m / s², Z: 9.78 m / s²; the encoders of its six joints record the joint angles at t=0.005s as 15.2°, -8.5°, 30.1°, -20.3°, 45.0°, and -15.7°, with corresponding motor currents of 1.2A, 0.8A, 1.5A, 0.9A, 1.1A, and 0.7A. Meanwhile, a depth camera mounted on the robot's base captured an image of the part at t=0.002s. Using a point cloud processing algorithm, the geometric parameters of the part were identified: volume 120cm³, center of gravity offset relative to the gripping point (0.5cm, -0.3cm, 0.1cm), and maximum surface curvature change rate 0.08cm. -1 The raw data is first preprocessed to unify the IMU, joint, and camera data to a sampling rate of 10Hz. Then, the IMU data is low-pass filtered, and the joint angle and current data are filtered by moving average.

[0022] Step S200: Based on the preprocessed data, perform multimodal feature extraction; use mutual information and random forest feature importance to select the most sensitive and discriminative key feature subset that distinguishes different vibration levels and associated grasping weights; use the selected key feature subset to train a random forest regression model to establish a mapping relationship between vibration features and the maximum safe grasping weight of the current robot grasping mode. Specifically, for robot body state data, multiple analysis methods are used to extract features. In time-domain analysis, the statistical characteristics of the triaxial acceleration signal over time are calculated to reflect the overall intensity, fluctuation, and impact characteristics of the vibration. Frequency-domain analysis uses Fast Fourier Transform to convert the time-domain signal to the frequency domain, analyzes the energy distribution of different frequency components, identifies the main vibration frequencies and their amplitudes, and determines whether there is resonance or abnormal high-frequency vibration at a specific frequency. Time-frequency domain analysis uses Short-Time Fourier Transform to reveal how the frequency components of the vibration signal change over time. The dynamic evolution of vibration frequency and intensity is observed, and dynamic change analysis focuses on the rate of change or trend of the signal over time to capture changes in the vibration state. For object parameter data, including weight, three-dimensional dimensions, shape features, fragility rating, and surface material information, these are directly used as features. The robot's working state is evaluated using extracted vibration-related features, and the vibration level is divided into four preset categories: normal state, slight vibration, moderate vibration, and severe vibration. To achieve this classification, a threshold corresponding to these four levels needs to be set for each vibration-related feature. The feature value distribution under different vibration states is collected through a large number of experiments under typical operating conditions and determined based on quantile statistics. The currently extracted vibration feature values ​​are compared with these preset thresholds to evaluate the overall or local vibration state of the robot. The mutual information between each extracted feature and the four vibration levels and gripping weight is calculated. For all possible combinations of x and y, the probability of their co-occurrence is calculated and multiplied by an association factor. This factor measures the difference between the observed joint probability and the joint probability assumed that x and y are independent. If x and y are independent, the mutual information is 0; if x and y are strongly correlated, the mutual information is larger. The larger the value, the stronger the association. A high mutual information value means that the feature carries more information about the vibration level or gripping weight. Feature importance is evaluated using a random forest model. A preliminary random forest model is constructed, with all candidate features as input and the target variable as output. The random forest evaluates the contribution of each feature to the model's predictive ability by calculating the Gini importance of the features. When constructing each tree, the decrease in the Gini index brought by each feature at the split node is recorded. The decrease in the Gini index brought by a feature in all trees is averaged, and the average value is the Gini importance of the feature. A high importance score indicates that the feature contributes greatly to the model's decision. The scores from both methods are combined to rank all features; a weighted average algorithm is then used to select the highest-scoring features from the ranked list based on pre-defined screening criteria, forming a subset of key features; this subset is designed to distinguish different vibration levels and reflect their correlation with the weight being grasped. The selected key features are combined to form feature vectors for model input. Each feature vector contains information that can characterize the robot's current state and the item's parameters. The dataset used to train the random forest regression model consists of multiple sets of samples. Each set of samples contains a feature vector and a corresponding maximum safe grasp weight value. The maximum safe grasp weight value is determined through actual experiments and taking into account the physical limitations of the robot itself. It represents the maximum allowable grasp weight that can guarantee the safety of the grasping process under the current robot state and item parameter data. The dataset contains samples under different grasping modes. The random forest regression model is implemented by constructing multiple decision trees. Each decision tree is trained on a random subset of the data. For an input feature vector, each tree predicts a maximum safe grab weight value. The final prediction result is the average of the predictions of all decision trees, which reduces the risk of overfitting of a single decision tree and improves the accuracy and robustness of the model. Using the prepared dataset, we optimize hyperparameters such as the optimal features and thresholds for splitting data when the decision tree splits at nodes, and the number of decision trees in the forest. We use cross-validation to adjust the model so that it learns a nonlinear mapping relationship between the key input features and the maximum safe grasping weight. To increase the system's safety redundancy, we set a dynamic adjustment mechanism. When the monitored vibration-related features exceed the preset normal range, indicating a state of slight vibration, moderate vibration, or severe vibration, or when certain key item parameters exceed the preset safe operating range, we automatically reduce the predicted maximum safe grasping weight value.

[0023] In one specific embodiment, the robot's end effector collects three-axis acceleration signals X: 0.12 m / s², Y: -0.05 m / s², and Z: 0.98 m / s² before grasping. Time-domain analysis yields a root mean square (RMS) of 0.12 m / s² and a kurtosis of 3.2. Frequency-domain analysis shows a dominant frequency of 12 Hz (amplitude 0.08 m / s²) and a dominant frequency of 58 Hz (amplitude 0.03 m / s²), with the 58 Hz frequency close to the robotic arm's natural frequency. Time-frequency analysis reveals that the amplitude of the 12 Hz vibration increases by 15% within 0.2 seconds before grasping. The item parameters are: weight 1.5 kg, dimensions (120 × 80 × 50) mm, shape complexity score 0.7, fragility score 0.3, surface material aluminum alloy, and coefficient of friction 0.45. The mutual information between vibration features and vibration levels were 0.12, 0.18, 0.25, 0.31 and 0.08, 0.15, 0.22, 0.27, respectively. The Gini importance ranking of the random forest was 58Hz amplitude 0.35, fragility 0.28, RMS 0.22, and complexity 0.18. 58Hz amplitude, fragility, RMS, and complexity were selected as key features, and a feature vector [0.03, 0.3, 0.12, 0.7] was constructed. A random forest model trained on historical data (100 trees) predicted a maximum safe grasping weight of 2.8 kg. However, it detected that the 58Hz amplitude of 0.03 exceeded the preset threshold of 0.025, and automatically reduced the safe weight by 25% to 2.1 kg.

[0024] Step S300: Apply the trained model to the robotic arm to perform grasping and releasing tasks, extract features in real time and input them into the model, calculate and dynamically adjust the maximum safe grasping weight of each grasping mode in the current state; Specifically, during the robotic arm's grasping and releasing task, step S100 collects time-synchronized, pre-processed, and calibrated robot body state data and object parameter data; step S200 obtains a feature vector containing N key feature values, which is input into a trained random forest regression model to predict the maximum safe grasping weight for each mode; and during the grasping and releasing process, vibration data is continuously monitored, and the actual vibration level is compared with the expected level to determine whether the actual vibration exceeds the expected or normal range. The vibration level determination results are continuously monitored. Once the vibration level is detected to change to moderate or severe, the preset dynamic adjustment mechanism is activated. The maximum safe gripping weight predicted by the gripping mode is multiplied by a preset discount factor as a correction to obtain the actual maximum safe gripping weight. The value of the factor can be set according to factors such as safety requirements and the severity of vibration. After the capture and release task is completed, the key feature subset recorded in the capture and release task, the actual capture weight and the actual vibration level determined are used as a new data sample and appended to the existing original dataset file or database table used to train the random forest model. Ensure that the data format is consistent with the original training set to obtain the updated random forest regression model trained on the extended dataset.

[0025] In one specific embodiment, when the robot performs the task of grasping a blue packaging box, it collects real-time triaxial acceleration signals: X: 0.15 m / s, Y: -0.07 m / s, Z: 1.02 m / s. The calculated RMS is 0.14 m / s, kurtosis is 3.5, and FFT analysis shows that the 58Hz vibration amplitude rises to 0.035 m / s. The item parameters are: weight 1.8 kg, dimensions (150 × 100 × 60) mm, and fragility 0.4. The feature vector [0.035, 0.4, 0.14, 0.6] is input into a random forest model, predicting a maximum safe weight of 2.5 kg for single-handed grasping. If the 58Hz amplitude exceeds the threshold of 0.025, the vibration level is determined to be "moderate," and a dynamic adjustment mechanism is activated, multiplying the weight by a discount factor of 0.8 to correct it to 2.0 kg. After the task is completed, the feature vector [0.035, 0.4, 0.14, 0.6], the actual grasped weight of 1.8kg, and the actual vibration level of "moderate" are recorded, appended to the original dataset, and the training is updated to obtain a new model.

[0026] Step S400: Reconstruct the 3D model using the depth map, analyze the item parameter data, and obtain potential gripping points; based on the obtained potential gripping points and the item weight, select the gripping mode and gripping points; after calculating the gripping mode and gripping points, adjust the gripping mode, gripping points, and gripping force according to the item parameter data to grip the item. Specifically, the structured light camera is precisely calibrated by photographing a calibration board, including intrinsic and extrinsic parameters; with the robot arm moving or the camera fixed, the viewing angle is adjusted around the object to be grasped, and the object is photographed from multiple different angles; each photograph generates a depth map, where each pixel represents the depth value of that point from the camera; the corresponding RGB image data is acquired; Gaussian filtering is used to remove isolated pixels or high-frequency noise caused by noise in the depth map; nearest neighbor interpolation is used to fill holes in the depth sensor where effective depth values ​​cannot be obtained at edges, occluded areas, or reflective areas; distortion coefficients obtained from camera calibration are applied to correct geometric distortion of the image and depth values; and a reasonable range of depth values ​​is set according to scene lighting and sensor characteristics, filtering out invalid or abnormal depth values. Each preprocessed depth map frame is converted into point cloud data; each pixel is used to calculate its 3D coordinates in the camera coordinate system based on its depth value and camera intrinsic parameters, forming a point cloud; the registration algorithm in the PCL library is used to align the point clouds obtained from different viewpoints to the same coordinate system; the pose information of the camera when capturing each depth map frame is used as the initial alignment basis; all aligned point clouds are fused, overlapping areas are merged, and a complete and dense 3D point cloud model is generated; the final fused 3D model is then converted to the robot's base coordinate system. Significant and easily fittable features, such as planes, cylinders, spheres, or multiple corners, are extracted from the reconstructed 3D model. These features are then fitted using the RANSAC geometric fitting algorithm. Based on the fitted geometric model, the overall pose of the object relative to the robot's base coordinate system is calculated. In a point cloud or mesh model, find the maximum and minimum coordinate values ​​of the model along the X, Y, and Z axes, and calculate the difference to obtain the length, width, and height of the object. Based on the distribution and geometric features of the point cloud, use the K-means clustering algorithm to roughly classify the objects into simple or complex shapes. Use the RANSAC plane fitting algorithm to search for large planar regions in the point cloud. Calculate the normal vector of the point cloud, find regions with drastic changes in normal vector, or use curvature changes to detect edges. Find points in the point cloud with maximum curvature or drastic changes in normal vector direction, which are usually corner points of objects. Calculate the normal vector for each point or facet on the point cloud model. Its local normal vector; within a local region, calculate the standard deviation or average angle between the normal vectors of all points within the region. The smaller the angle, the smoother the surface of the region. It can analyze the uniformity of the texture, calculate the color and texture gradient of the local region. The smaller the gradient, the smoother the surface. Combining the normal vector consistency and texture analysis results, a smoothness score or index value is calculated. Based on experience or experiments, a smoothness threshold is set, representing a critical point. If it exceeds the threshold, it is considered that it is easy to slip when gripping. The obtained smoothness score is compared with the preset threshold. If the score exceeds the threshold, the surface of the item is marked as too smooth. By analyzing changes in the normal vector, curvature, or inter-point distance of the point cloud, the locations of abrupt changes in geometric features are identified, which are the boundaries of special regions. Combining RGB image information, regions with significantly different colors or textures from their surroundings are sought within geometrically discontinuous areas. These are analyzed to determine if they are labels, trademarks, or text regions, and their location is determined through color and texture clustering or edge detection. Within geometrically discontinuous areas, locally concave regions are identified. By calculating the perpendicular distance from the center point of this region to the surrounding plane, the depth of the concave region is obtained, and its boundary shape is analyzed. Within geometrically discontinuous areas, locally convex regions are identified. By calculating the perpendicular distance from the highest point of this region to the surrounding plane, the height of the convex region is obtained, and its boundary shape is analyzed. Visual features of objects are extracted from RGB images and compared with a pre-built object database to identify known fragile marks or parts. By combining known material properties in the database with the overall shape analysis of the current object, potential weak points or fragile areas can be inferred. For example, slender protrusions and thin-walled areas are usually more fragile. Based on the geometric feature information obtained in the previous steps, suitable areas for grasping are selected, prioritizing large planar areas and flat curved surfaces, while avoiding edges, corners, pits, protrusions, labeled areas, and fragile areas. Within the selected candidate areas, specific grasping point coordinates are generated according to the grasping strategy. For example, points can be sampled evenly on a plane, or the center point or geometric center of the area can be selected. Based on the subsequent grasping mode, a corresponding number of potential grasping points are generated. The actual weight of the object is directly measured using a weight sensor. From previous model prediction and dynamic adjustment steps, the maximum safe gripping weight values ​​for the robot's single-handed and two-handed gripping modes in the current state are obtained. The actual measured weight of the object is compared with these two safe weight values. When the object weight is less than the safe gripping weight for single-handed gripping, the single-handed gripping mode is adopted. From the generated list of potential gripping points, priority is given to points closest to the object's center of gravity and points with flat surfaces and consistent normal vectors, avoiding marked fragile areas and detected sharp edges. The coordinates of the selected optimal gripping point are transformed into the coordinate system of the robotic arm performing the gripping. Based on the transformed target point coordinates, combined with the size and posture requirements of the end effector, the inverse kinematics algorithm is used to calculate the angles that each joint of the robotic arm needs to reach, generating a smooth motion trajectory from the current position to the target pose. When the weight of the item is greater than or equal to the safe gripping weight with one hand but less than the safe gripping weight with both hands, a two-hand gripping mode is adopted. Two gripping points are selected from the generated list of potential gripping points, and their positions relative to the center of gravity of the item provide sufficient friction and torque to prevent the item from rotating or slipping during gripping and movement. The two gripping points should be roughly symmetrically distributed on both sides of the center of gravity, or adjusted according to the position of the center of gravity, so that the resultant torque generated by the gripping force is close to zero, maintaining the balance of the item. Fragile areas and dangerous edges are strictly avoided. Ensure that both points are within the reachable workspace of the two robotic arms. The coordinates of the optimal gripping points selected for each robotic arm are transformed into the coordinate system of their respective robotic arms. The inverse kinematics algorithm is used independently for each robotic arm to calculate the joint angles that need to be reached, generating a coordinated motion trajectory to ensure that the target pose is reached synchronously or in a predetermined order. When the weight of an item exceeds the maximum safe gripping weight in the current two-hand gripping mode, it is determined to be an overload anomaly. An abnormal signal or error code is sent to the control system or user interface, instructing the robot to skip the current gripping task, perform an avoidance action, and perform the next gripping task. The robot also notifies the on-site operators or maintenance personnel through sound, light, or network messages, informing them that there is an overweight item that requires manual intervention.

[0027] When the gripping mode is confirmed to be single-handed gripping mode, preprocessed item parameter data is extracted and utilized, including the item's size, weight, and center of gravity position estimated by vision or force sensors; surface features of the gripping point are analyzed using RGB images or depth maps; based on physical principles, including statics and tribology, the minimum normal force required to prevent the item from slipping from the single-handed gripper under gravity is calculated; this involves calculating the projection of the item's gravitational components onto the gripping plane, and the coefficient of friction estimated based on the surface features of the gripping point, as shown in the formula: ; Among them, F minW is the minimum normal force, g is the weight of the object, θ is the gravitational acceleration, θ is the angle between the sliding direction and the horizontal plane, sin(θ) is the sine value corresponding to the angle, representing the proportion of gravity in the sliding direction, and μ is the coefficient of friction. Surface features affect the estimated value of the coefficient of friction. Based on the calculated minimum gripping force, a preset safety factor is multiplied to obtain a final set gripping force command value used to control the closure of the gripper. The evaluated surface smoothness and fragility values ​​are compared with the preset safety thresholds inside the system. If either condition exceeds the threshold, the system determines that single-handed gripping is high-risk, triggers mode switching, and starts the strategy planning process for two-handed gripping mode. When the gripping mode is confirmed to be a two-handed gripping mode, it is also necessary to clarify the specific surface features of the two gripping points and their geometric relationship in three-dimensional space; use force closure analysis to determine whether multiple contact points can provide sufficient constraint force to stably hold the object; establish a force-torque space model, considering the normal and tangential forces that may be generated at the two gripping points, as well as their positions relative to the object's center of gravity; find a set of forces such that their resultant force-torque can balance the object's gravity and other external forces and torques, and satisfy the friction cone constraint; solve for the minimum total normal force that satisfies the force closure condition through linear algebra and optimization calculations, and determine the force distribution between the two gripping points; set a force with safety redundancy for each gripping point, or set redundancy for the total force. Then, the force is allocated; the safety factor is selected based on the combined friction of the two contact points. If both points are relatively smooth, the safety factor will be higher; after obtaining the initial force allocation and the force after safety redundancy, further adjustments are made according to the fragility of the object. If the calculated force at a certain gripping point exceeds the threshold that the material at that point can withstand, the force at that point is reduced, and the force at the other point is increased accordingly, or the total force is reduced overall; through optimization algorithms, under the premise of satisfying force closure, total force requirements, and single-point force limitations, a more uniform or optimized force distribution ratio is found to make the stress distribution of the two gripping points more reasonable and avoid stress concentration; during the gripping process, the magnitude of the two forces needs to be fine-tuned to maintain the stability of the object's posture, while continuously monitoring the stress distribution.

[0028] In one specific embodiment, a calibrated structured light camera captures images of a blue packaging box to be grasped from five different angles, obtaining five frames of depth maps and RGB images. Each depth map frame is subjected to Gaussian filtering (σ=1.5) and interpolation padding to correct distortion and filter out invalid data with depth values ​​greater than 3 meters. The preprocessed depth maps are converted into point clouds, and the ICP algorithm is used to register the point clouds from different viewpoints to the camera coordinate system. After fusion, a complete model containing approximately 200,000 points is obtained and converted to the robot's base coordinate system. A box-shaped geometric feature with a length of 150mm, a width of 100mm, and a height of 60mm is extracted from the model and determined to be a simple shape. RANSAC plane fitting is used to find the top plane, and the standard deviation of the surface normal vector is calculated to be 0.12, which is less than the threshold of 0.2, indicating that the surface smoothness is acceptable. Point cloud analysis did not reveal any obvious pits or protrusions, and comparison with the RGB image database identified a fragile label area on the side of the box. The top plane is preferentially selected as the grasping area, and a potential grasping point is generated at its center with coordinates (0.5, 0.3, 0.8) meters. The weight sensor measured an actual weight of 1.8 kg. The model predicted the maximum safe gripping weight for one hand to be 2.0 kg, and for both hands to be 4.5 kg. Since 1.8 kg < 2.0 kg, the one-handed gripping mode was selected, and the top center point was chosen as the gripping point. After transforming the coordinates of this point to the robotic arm coordinate system, the joint angles that the robotic arm needs to achieve are calculated using inverse kinematics as [45°, -30°, 15°, 0°, 0°, 10°], generating a smooth motion trajectory to execute the gripping.

[0029] After confirming the single-handed grasping mode, the item parameters are: weight 1.8kg, dimensions (150×100×60)mm, and visually estimated center of gravity located at the geometric center (75, 50, 30)mm. The grasping point is located on the top plane, the standard deviation of the surface normal vector is 0.12, and the friction coefficient μ is estimated to be 0.35 based on RGB image analysis. Calculate the minimum normal force F required to prevent slippage. min ≈(1.8kg×9.8m / s²×sin(90°))÷0.35≈50.1N. Multiply by a safety factor of 1.5, and set the gripping force command value to 75N. Surface smoothness of 0.12 is below the threshold of 0.3, and fragility of 0.4 is below the threshold of 0.7, so the risk is considered acceptable. If changing to a two-hand gripping method, set two gripping points at the opposite edges of the top plane at (75, 50)mm and (75, 100)mm, with a spacing of 100mm. Force closure analysis is used to establish a force-torque model to solve for the minimum total normal force that satisfies the equilibrium and friction cone constraints. The friction coefficients at both points are 0.35, and the calculated minimum total normal force is approximately 60N. Initially, 30N is allocated to each point, and multiplied by a safety factor of 1.4, resulting in 42N / point. Considering fragility of 0.4, the material threshold at point 1 is 80N, and the threshold at point 2 is 70N, both of which are satisfied. The optimized force distribution is 40N at point 1 and 35N at point 2, for a total force of 75N. In actual control, the two forces are fine-tuned to 38N and 37N respectively to maintain the stability of the object.

[0030] Step S500: Scan and analyze the target area at the specified location in real time to generate candidate placement locations; evaluate the candidate placement locations and calculate the overall space utilization rate of the area after each candidate location is placed; perform stacking security analysis, including protection of items below and stacking stability; combine space utilization and stacking security, and use a weighted scoring algorithm to select the optimal placement location and posture.

[0031] Specifically, based on received instructions, the robot activates its onboard depth camera sensor to scan the designated placement area in real time, acquiring 3D data of the area. Using this scan data, a 3D model of the object distribution is constructed and mapped onto a pre-defined 3D voxel grid, where each voxel represents a tiny cube in space. The model is analyzed, and all voxels occupied by existing objects are marked as occupied. Based on the size and shape parameters of the object to be placed, a search is conducted in the voxel grid for a continuous area of ​​free voxels sufficient to accommodate the object. These potential placement areas are defined as candidate placement locations, and their key coordinate information is recorded. For each candidate location, the object is simulated being placed there, and the number of newly occupied voxels is calculated. This number is added to the number of occupied voxels in the area before placement to obtain the total number of occupied voxels after placing the object. This total number of voxels is then divided by the total number of voxels in the entire designated area to calculate the space utilization rate corresponding to the candidate location.

[0032] When performing stacking safety analysis, the process involves checking whether the candidate location is on top of existing items, obtaining the weight of the item to be placed and the estimated drop height, and using a physical model to calculate the impact force at the moment of placement. A simplified model assumes free fall, and the average impact force is estimated using the law of conservation of energy and the momentum theorem, combined with a preset contact time. The static pressure exerted by the placed item on the supporting surface of the item below is calculated, directly equal to the item's weight divided by the contact area. Using a 3D model of the item, based on the planned placement posture, the contact between the item and the item below is simulated, and the actual contact surface area is calculated. The calculated impact force is compared with the known impact force tolerance threshold of the item below, and the calculated static pressure is compared with the known static pressure tolerance threshold of the item below. Two impact risk indicators and a pressure risk indicator are generated. Combining these two risk indicators with the current state of the item below, a preset weighted summation algorithm is used to generate a comprehensive damage risk index for the item below, which is used for subsequent comprehensive stacking safety assessment. Simultaneously, the stability of the stack is evaluated, considering factors such as the shape of the items, center of gravity distribution, friction coefficient of the contact surface, and stacking height. A stacking stability score is output through a comprehensive judgment using static equilibrium analysis and dynamic simulation.

[0033] The calculated space utilization rate, the damage risk index of the items below, and the stacking stability score are used as three key inputs. Based on the priority requirements of the current operation scenario, different weights w1, w2, and w3 are assigned to these three inputs. Through a linear weighted scoring algorithm, the scores of each candidate position are multiplied by their corresponding weights and then summed to obtain a comprehensive score W. The comprehensive scores of all candidate positions are compared, and the position with the highest score is selected as the best placement position. The optimal placement posture at this position is also determined. The robot will then perform grasping and placement actions to place the item in this selected position and posture.

[0034] In one specific embodiment, the robot receives an instruction to place a rectangular cardboard box measuring 30cm × 20cm × 10cm and weighing 5kg into a designated shelf area. The robot activates its onboard depth camera to perform real-time scanning of the target area of ​​approximately 1m × 1m × 1m, acquiring high-precision 3D point cloud data. Using this data, a 3D model is constructed and mapped onto a voxel mesh with 1cm sides, generating a total of 100,000 voxels. After analyzing the model, it is found that 40,000 voxels are already occupied. Based on the box's dimensions, the system searches the mesh and finds five consecutive empty voxel areas as candidate placement locations. Candidate location 1 is located at the edge of the shelf, just large enough to fit the box; after placement, it occupies an additional 6,000 voxels, bringing the total occupancy to 46,000, resulting in a space utilization rate of 46%. Candidate location 2 is located on top of an 8kg item; after placement, it occupies a total of 47,000 voxels, resulting in a space utilization rate of 47%. When performing stacking safety analysis, for candidate position 2, the system calculates its fall height as 0.5m, contact time as 0.1s, and estimates the impact force as approximately 245N; the effective contact area is 0.3m², and the static pressure is approximately 1633Pa. The impact force tolerance threshold for the item below is 300N, and the static pressure tolerance threshold is 2000Pa. Therefore, the impact risk indicator is approximately 0.82, and the pressure risk indicator is approximately 0.82. Considering the current good condition of the item below, the impact weight is set to 0.6, and the pressure weight to 0.4. After weighted summation, the damage risk index of the item below is 0.82. Stacking stability analysis considers the friction coefficient of 0.4, center of gravity distribution, etc., and outputs a stability score of 0.7. Based on the current scenario's priority space utilization, the weights w1 are set to 0.5, w2 to 0.3, and w3 to 0.2. The comprehensive score W1 for candidate position 1 is 23.5, and the comprehensive score W2 for candidate position 2 is approximately 23.62. Although position 2 has a slightly higher space utilization rate, its overall score is slightly lower than that of position 1. Therefore, position 1, which has the highest score, is selected as the best placement position, and its optimal posture is determined to be placing the carton along the edge of the shelf with its long side. The robot then performs the grasping and placing actions to place the carton in that position and posture.

[0035] like Figure 2The system architecture diagram of an AI-based data acquisition monitoring and management system is shown. This invention provides an AI-based data acquisition monitoring and management system, comprising: Command receiving and data acquisition module: includes command receiving unit, data acquisition unit and data preprocessing unit; wherein, the command receiving unit receives the command to grab the item and place it in the designated position; the data acquisition unit acquires the robot body status data and item parameter data; the data preprocessing unit performs preprocessing operations on the raw data; Model building and dynamic adjustment module: includes a mapping model building unit and a maximum safe gripping weight adjustment unit; wherein, the mapping model building unit extracts key features, uses mutual information and random forest to select feature subsets, trains a random forest regression model, and establishes a mapping relationship between vibration features and the maximum safe gripping weight of each gripping mode; the maximum safe gripping weight adjustment unit dynamically adjusts the actual maximum safe gripping weight under the current state and updates and trains the model based on the recorded data; The 3D perception and gripping point analysis and decision-making module includes a 3D reconstruction unit, a gripping point identification and evaluation unit, a gripping mode decision-making unit, a gripping point selection unit, and a coordinate transformation and planning unit. Specifically, the 3D reconstruction unit reconstructs the 3D model of the object using a depth map; the gripping point identification and evaluation unit analyzes the object's parameter data and the 3D model to identify and evaluate potential gripping points; the gripping mode decision-making unit compares the object's weight with the dynamically adjusted maximum safe gripping weight, selecting either a single-handed or double-handed gripping mode, or determining an overload anomaly; the gripping point selection unit selects the optimal gripping point based on the identified potential gripping points, combined with the object's parameter data and 3D model analysis; and the coordinate transformation and planning unit transforms the gripping point coordinates to the robotic arm coordinate system and plans the position and orientation of the end effector. The grasping strategy optimization and execution module includes a single-hand grasping force calculation and safety assessment unit, a two-hand grasping force calculation and allocation unit, and a grasping execution unit. Specifically, the single-hand grasping force calculation and safety assessment unit calculates the minimum single-hand force, sets safety redundancy, and determines whether to switch modes when the grasping mode is confirmed to be single-handed. The two-hand grasping force calculation and allocation unit calculates the total force and allocated force, sets safety redundancy, and adjusts the force balance when the grasping mode is confirmed to be two-handed. The grasping execution unit controls the robot to perform the grasping action according to the final grasping strategy. Space utilization and stacking safety calculation module: including space utilization unit and stacking safety unit; wherein, the space utilization unit calculates the space utilization of the entire specified target area after the current item is placed in a specific candidate position; the stacking safety unit calculates the damage risk index of the items below and the stacking stability score, and obtains the stacking safety by weighted summation; The placement location selection and placement module includes a priority weight allocation unit and a comprehensive score calculation and placement unit. The priority weight allocation unit assigns different weights to the signal according to the priority requirements of the current operation scenario. The comprehensive score calculation and placement unit calculates and sums the space utilization and stacking security of each candidate location to obtain a comprehensive score, and selects the one with the highest score as the best placement location for placement.

[0036] It will be apparent to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above, and that the invention can be implemented in other specific forms without departing from its spirit or essential characteristics. Therefore, the embodiments should be considered in all respects as exemplary and non-limiting, and the scope of the invention is defined by the appended claims rather than the foregoing description. Thus, all variations falling within the meaning and scope of equivalents of the claims are intended to be included within the present invention. No reference numerals in the claims should be construed as limiting the scope of the claims.

Claims

1. An artificial intelligence-based collector data monitoring management method, characterized by, The robot receives instructions for picking up an object and placing it in a designated location, collects robot body state data and object parameter data, and pre-processes the raw data. Based on the pre-processed data, multi-modal feature extraction is performed. Mutual information and random forest feature importance are used to select a key feature subset that is most sensitive and most discriminative in distinguishing different vibration levels and associated picking weights. Using the selected key feature subset, a random forest regression model is selected for training to establish a mapping relationship between vibration features and the maximum safe picking weight of the current robot picking mode. The trained model is applied to the robot arm to perform the pick-and-place task, real-time feature extraction and model input, calculation and dynamic adjustment of the maximum safe picking weight of each picking mode in the current state. A three-dimensional model is reconstructed from the depth map, and the object parameter data is analyzed to obtain potential picking points. Based on the obtained potential picking points, combined with the object weight, the picking mode and picking point are selected. After calculating the picking mode and picking point, the picking mode, picking point and picking force are adjusted according to the object parameter data, and the object is picked up. The target area of the designated location is scanned and analyzed in real time to generate candidate placement positions. The candidate placement positions are evaluated, and the spatial utilization of the overall area after placing each candidate position is calculated. Stacking safety analysis, including protection of underlying objects and stacking stability. Based on the spatial utilization and stacking safety, the optimal placement position and posture are selected using a weighted scoring algorithm. The robot receives instructions for picking up an object and placing it in a designated location, collects robot body state data and object parameter data, and pre-processes the raw data, including:

2. The method of claim 1, wherein the method is based on artificial intelligence. Multi-modal sensors are deployed at the end effector of the robot arm and at each joint to collect real-time robot body state data, including three-axis acceleration, joint angle, motor current, and joint temperature. Object parameter data is collected through vision sensors, weight sensors, and tactile sensors, including object weight, three-dimensional size, shape features, fragility, and surface material information. The collected raw data is pre-processed. Remove or correct noise, outliers, and missing values generated during sensor collection. Time alignment and spatial registration are performed on data from different sensors and different timestamps. Based on the pre-processed data, multi-modal feature extraction is performed. Mutual information and random forest feature importance are used to select a key feature subset that is most sensitive and most discriminative in distinguishing different vibration levels and associated picking weights. Using the selected key feature subset, a random forest regression model is selected for training to establish a mapping relationship between vibration features and the maximum safe picking weight of the current robot picking mode, including:

3. The method of claim 1, wherein the method further comprises: ​ Based on the pre-processed robot body state data and the article parameter data, multi-modal feature extraction is performed; the vibration level is divided into four levels, the mutual information between each feature and the four vibration levels and the grabbing weight is calculated, the importance of the features is evaluated using a random forest model, the scores of the two methods are combined to sort all the features, the highest scoring features are selected from the sorted list according to the set screening standard, and a key feature subset is formed which can distinguish different vibration levels and is related to the grabbing weight; The selected key feature subset is divided into a feature vector for model input, and the data set for training the random forest regression model is composed of multiple samples, including the key feature vector and the corresponding maximum safe grabbing weight value in single-hand grabbing mode and double-hand grabbing mode; The random forest improves the accuracy and robustness of the model by constructing multiple decision trees and integrating the prediction results; using the prepared data set, the model learns the nonlinear mapping relationship between the input key features and the maximum safe grabbing weight by optimizing the splitting nodes of the decision trees and the number of trees.

4. The method of claim 1, wherein the method further comprises: The trained model is applied to the robot arm to perform the pick-and-place task, real-time feature extraction and model input are performed, and the maximum safe grabbing weight of each grabbing mode in the current state is calculated and dynamically adjusted, including: Collecting robot body state data and article parameter data in the robot arm pick-and-place task, real-time extraction of the key feature subset, input into the trained random forest regression model, the model predicts the corresponding maximum safe grabbing weight of each single-hand grabbing mode and double-hand grabbing mode according to the input features, and the article is grabbed; During the pick-and-place process, the vibration data is continuously monitored, and the actual vibration level is compared with the expected value; when moderate or severe vibration is detected, the dynamic adjustment mechanism is started; the maximum safe grabbing weight predicted by the grabbing mode is multiplied by a preset discount factor as a correction to obtain the actual maximum safe grabbing weight; The information recorded in this pick-and-place task is used as new data samples to trigger retraining or incremental updating of the random forest regression model.

5. The method of claim 1, wherein the method further comprises: The three-dimensional model is reconstructed by depth map, the article parameter data is analyzed, and the potential grabbing points are obtained, including: Through the structured light camera carried by the robot, multiple frames of depth map data of the article at different angles are obtained and preprocessed; using the point cloud processing algorithm in the PCL library, the multiple preprocessed depth map data are fused to reconstruct a three-dimensional model, and the model is unified to the robot coordinate system; Using model fitting algorithm, the pose of the article in the robot base coordinate system is calculated; the overall size of the article is calculated, the overall shape category of the article is analyzed, and the geometric features of the article surface are detected; based on the geometric shape and mass distribution of the three-dimensional model, the center of gravity of the article is calculated; the texture features of the article surface are analyzed, the smoothness is judged, the special surface area is identified, the visual features and the preloaded article database information are used to identify the marks or known fragile parts on the article; according to the shape and material of the article, the weak part is inferred; Based on the identified geometric features, a series of potential grabbing points are generated.

6. The method of claim 1, wherein the method further comprises: The obtained potential grabbing points are combined with the weight of the article to select a grabbing mode and a grabbing point, including: The weight of the article is compared with the maximum safe grabbing weight of the single-hand grabbing mode and the double-hand grabbing mode of the robot in the current state obtained by dynamic adjustment respectively; When the weight of the article is lower than the maximum safe grabbing weight of the single-hand grabbing mode in the current state, the single-hand grabbing mode is selected for grabbing; from the identified potential grabbing points, an area near the center of gravity of the article, a smooth surface and an area not easy to slip are selected as the grabbing point in priority, while avoiding the marked fragile area and sharp edge, and the position and posture that the end effector needs to reach are planned; When the weight of the article is between the maximum safe grabbing weight of the single-hand grabbing mode and the double-hand grabbing mode in the current state, the double-hand grabbing mode is selected for grabbing; according to the force closure condition and the position of the center of gravity, the optimal grabbing point for each of the two robot arms is selected from the identified potential grabbing points, and the position and posture that the two end effectors need to reach are planned; When the weight of the article is higher than the maximum safe grabbing weight of the double-hand grabbing mode in the current state, it is determined that the article is overloaded, abnormal information is fed back, and the robot is instructed to skip the article, and the responsible personnel are notified for manual processing.

7. The method of claim 1, wherein the method further comprises: After the grabbing mode and the grabbing point are calculated, the grabbing mode, the grabbing point and the grabbing force are adjusted according to the article parameter data, and the article is grabbed, including: When the grabbing mode is confirmed as the single-hand grabbing mode, the minimum grabbing force required for the grabbing point is determined based on the size, weight and center of gravity position of the article, combined with the surface features of the grabbing point, the grabbing force with safety redundancy is set considering the friction condition of the grabbing point; the smoothness of the surface of the article and the fragility of the article are evaluated to determine whether to switch to the double-hand grabbing mode; When the grabbing mode is confirmed as the double-hand grabbing mode, the minimum total grabbing force required for stable grabbing and the force distribution between the two grabbing points are calculated based on the size, weight, center of gravity position of the article, specific surface features and geometric relationship of the two grabbing points by force closure analysis; considering the smoothness of the article, the grabbing force instruction value with safety redundancy is set for the grabbing point; considering the fragility, the force applied by each grabbing point and the balance between the two forces are adjusted through force closure analysis and force distribution.

8. The method of claim 1, wherein, The target area of the specified position is scanned and analyzed in real time to generate candidate placement positions; The space utilization of the overall area after placing each candidate position is calculated by evaluating the candidate placement positions; Stacking safety analysis is performed, including protection of the lower article and stacking stability, including: The target area of the specified position in the instruction is scanned to construct a three-dimensional model of the current article distribution in the area; the position, shape and occupied space of the existing articles in the area are obtained by real-time scanning and mapped to a three-dimensional grid to mark the occupied voxels; combined with the article parameter data, multiple candidate positions are generated; The candidate positions are evaluated, and for each candidate placement position, the simulation of placing the current item is calculated, the proportion of the space utilized in the entire specified area after placement and the occupied voxels on the three-dimensional grid; calculate the total number of all voxels in the space range, add the number of occupied voxels before placing the new item to the number of newly occupied voxels after placing the new item, and get the total number of occupied voxels after placement, compared with the total number of voxels in the entire specified area, the space utilization rate is obtained; When the candidate position is above the existing item, the analysis of the protection of the underlying item is carried out through the analysis of the item parameter data obtained during the previous grabbing and the scanning of the status of the underlying item; according to the weight of the to-be-placed item and the falling height, the impact force generated within the preset time during placement and the static pressure generated on the supporting surface of the underlying item after placement are calculated; the contact area is calculated by simulating the placement posture and the shape of the item; the calculated impact force and static pressure are compared with the known fragility threshold of the underlying item; based on the comparison result and the state of the underlying item, the underlying item damage risk index is generated; The current to-be-placed item and all the items supporting it are regarded as a temporary stacking structure, including the mass, center of gravity position, shape and contact surface information of each item, and the stacking stability analysis is carried out; The state of the new stacking structure after the item is placed in this position is analyzed; the total center of gravity of the entire stacking structure is calculated, the base of the stacking structure on the supporting surface is determined, and it is checked whether the vertical projection point of the total center of gravity falls within the base range; if it falls within the range, it is statically stable, otherwise, it is statically unstable, and the corresponding static score is set respectively; In the simulation environment, different preset thrusts are applied to the stacking structure, the item state and the shaking amplitude are observed, the score range is set according to the shaking amplitude threshold, and the dynamic score is obtained; different weights are set for the static score and the dynamic score, and the stacking stability score is calculated by adding them together; The stacking safety is calculated by weighting the underlying item damage risk index and the stacking stability score.

9. The method of claim 1, wherein the method further comprises: The comprehensive space utilization rate and the stacking safety use the weighted scoring algorithm to select the optimal placement position and posture, including: According to the obtained space utilization rate, underlying item damage risk index and stacking stability score, they are regarded as different input signals, different weights are assigned to the signals according to the priority requirements of the current operation scene, and the weights are w1, w2 and w3 respectively; Through linear weighted scoring algorithm, the scores of each candidate position are multiplied by their corresponding weights and then added together to obtain the comprehensive score W, the formula is: ； Wherein, SU is the space utilization rate, DRI is the underlying item damage risk index, and SSS is the stacking stability score; compare the comprehensive scores of all candidate positions, select the one with the highest score as the best placement position, and determine the placement posture on the position, and place it.

10. An artificial intelligence-based collector data monitoring management system using the artificial intelligence-based collector data monitoring management method of any one of claims 1-9. It includes: The instruction receiving and data collecting module includes an instruction receiving unit, a data collecting unit and a data preprocessing unit; the instruction receiving unit receives the instruction of placing the grabbed item to the specified position; the data collecting unit collects the robot body state data and the item parameter data; the data preprocessing unit performs preprocessing operation on the original data; The model establishment and dynamic adjustment module comprises a mapping model establishment unit and a maximum safe grabbing weight adjustment unit; the mapping model establishment unit extracts key features, screens feature subsets by using mutual information and a random forest, trains a random forest regression model, and establishes a mapping relationship between vibration features and maximum safe grabbing weights of each grabbing mode; the maximum safe grabbing weight adjustment unit dynamically adjusts the actual maximum safe grabbing weight under the current state, and updates and trains the model according to recorded data; The three-dimensional perception and grabbing point analysis and decision module comprises a three-dimensional reconstruction unit, a grabbing point identification and evaluation unit, a grabbing mode decision unit, a grabbing point selection unit and a coordinate conversion and planning unit; the three-dimensional reconstruction unit reconstructs a three-dimensional model of the article through a depth map; the grabbing point identification and evaluation unit analyzes article parameter data and the three-dimensional model, identifies and evaluates potential grabbing points; the grabbing mode decision unit compares the article weight with the dynamically adjusted maximum safe grabbing weight, selects a single-hand or double-hand grabbing mode or judges overload abnormality; the grabbing point selection unit selects the optimal grabbing point according to the identified potential grabbing points, in combination with article parameter data and three-dimensional model analysis; and the coordinate conversion and planning unit converts the grabbing point coordinates to the mechanical arm coordinate system and plans the end effector position and posture; The grabbing strategy optimization and execution module comprises a single-hand grabbing force calculation and safety evaluation unit, a double-hand grabbing force calculation and distribution unit and a grabbing execution unit; when the single-hand grabbing force calculation and safety evaluation unit confirms that the grabbing mode is single-hand, it calculates the single-hand minimum force, sets a safety redundancy and judges whether to switch the mode; when the double-hand grabbing force calculation and distribution unit confirms that the grabbing mode is double-hand, it calculates the total force and distributes the force, sets a safety redundancy and adjusts the force balance; and the grabbing execution unit controls the robot to perform the grabbing action according to the final grabbing strategy; The space utilization rate and stacking safety calculation module comprises a space utilization rate unit and a stacking safety unit; the space utilization rate unit calculates the space utilization rate of the entire designated target region after the current article is placed at a specific candidate position; and the stacking safety unit calculates an underlying article damage risk index and a stacking stability score, and obtains the stacking safety by weighted summation; The placement position selection and placement module comprises a priority weight allocation unit and a comprehensive score calculation and placement unit; the priority weight allocation unit allocates different weights to signals according to the priority requirements of the current operation scene; and the comprehensive score calculation and placement unit performs weighted summation on the space utilization rate and the stacking safety of each candidate position, obtains a comprehensive score, selects the highest score as the best placement position, and performs placement.

Images