Building equipment operation data intelligent analysis method and system

Abstract

The invention provides a building equipment operation data intelligent analysis method and system, and relates to the technical field of data processing, and the method comprises the steps: extracting historical positioning data in a space calibration interval, and calculating a position offset vector set between a predicted trajectory point and an actual trajectory point according to the trajectory point on a predicted motion trajectory; taking the weighted sum of squares of minimizing the position offset vector as an optimization target to obtain a three-dimensional space transformation parameter optimization model; according to the optimization model, a quasi-Newton optimization algorithm is adopted to solve the optimization model, and a final calibration parameter set is generated; and based on the final calibration parameter set, correcting the predicted motion trajectory, performing spatial relationship analysis on the corrected trajectory and a spatial risk field, calculating a collision risk probability and an accident severity level, generating a comprehensive risk score, and triggering a response action according to a preset threshold value. The construction safety level and the management efficiency are improved.

Description

Technical Field

[0001] The present invention relates to the field of data processing technology, and in particular to a method and system for intelligent analysis of construction equipment operation data. Background Art

[0002] With the intelligent transformation of the construction industry, the efficient operation and safe management of construction equipment have become critical to improving construction efficiency and ensuring personnel safety. Traditional equipment monitoring, while relying primarily on real-time sensor data collection, suffers from significant deficiencies in risk prediction and precise control in dynamic environments. For example, existing technologies typically assess equipment status based on single sensor data or simple thresholds, making them incapable of addressing the multi-source data integration and dynamic risk assessment requirements of complex construction scenarios.

[0003] In addition, the motion trajectory prediction and environmental risk modeling of construction equipment often rely on independent physical models or data-driven methods, resulting in insufficient prediction accuracy or poor adaptability. Some traditional methods fail to achieve closed-loop control of risk assessment and trajectory correction, and it may be difficult to dynamically adjust equipment operation strategies to reduce risks. Summary of the Invention

[0004] The technical problem to be solved by the present invention is to provide a method and system for intelligent analysis of construction equipment operation data, thereby improving construction safety level and management efficiency.

[0005] In order to solve the above technical problems, the technical solutions of the present invention are as follows: In a first aspect, a method for intelligent analysis of construction equipment operation data is provided, the method comprising: Step 1: Obtain device status data and environmental perception data; Step 2: Obtain the device's predicted motion trajectory based on the device status data and environmental perception data; Step 3: Fusion the environmental perception data with the preset building information model and real-time obstacle scanning data to generate a spatial risk field; Step 4: Based on the spatial risk field and the predicted motion trajectory, two calibration reference points with fixed spatial positions are selected within the target analysis domain. A reference direction axis is generated based on the spatial relationship between the two calibration reference points. A continuous spatial calibration interval is divided along the reference direction axis. Historical positioning data within the spatial calibration interval is extracted. Based on the trajectory points on the predicted motion trajectory, a set of position offset vectors between the predicted trajectory points and the actual trajectory points is calculated. The optimization goal is to minimize the weighted sum of the squares of the position offset vectors to obtain a three-dimensional spatial transformation parameter optimization model. Based on the optimization model, a quasi-Newton optimization algorithm is used to solve the optimization model and generate the final calibration parameter set. Step 5: Based on the final calibration parameter set, the predicted motion trajectory is corrected, and the spatial relationship between the corrected trajectory and the spatial risk field is analyzed to calculate the collision risk probability and accident severity level, generate a comprehensive risk score, and trigger a response action according to the preset threshold.

[0006] Furthermore, step 1, obtaining device status data and environmental perception data, includes: Step 11: Collect raw voltage signals through the strain gauge array and perform analog-to-digital conversion on the raw voltage signals to generate discrete stress data; input the discrete stress data and the number of joint angle pulses collected by the encoder into the kinematic forward solution process to generate real-time spatial pose coordinates; process the real-time spatial pose coordinates, hydraulic pressure sensor readings, motor current sampling values, and control operation instructions to generate a structured device status data matrix; Step 12: Collect original point cloud data in a polar coordinate system using a lidar, and perform coordinate transformation on the original point cloud data to generate a three-dimensional point set in a global coordinate system. Based on the three-dimensional point set, filter out noise points based on the spatial distribution characteristics of adjacent frames to generate de-noised point cloud data. Collect Doppler frequency shift signals using a millimeter-wave radar, perform moving target detection on the frequency shift signals, and calculate obstacle radial velocity vectors. Fusion the de-noised point cloud data, obstacle radial velocity vectors, and the RGB-D depth map collected by the binocular vision sensor generates an environmental perception data packet. Step 13: Time-align the structured device status data matrix with the environment perception data packet, perform interpolation alignment based on the hardware clock synchronization signal, and generate time-synchronized device status data frames and environment perception data frames.

[0007] Furthermore, step 2 is to obtain the device's predicted motion trajectory based on the device status data and the environmental perception data, including: Step 21: Receive the time-synchronized device status data frame and extract the real-time spatial pose coordinates and control operation instruction codes from it; input the real-time spatial pose coordinates into the device dynamics equation library and generate initial motion trajectory segments based on the rigid body kinematics principle; input the initial motion trajectory segments and the control operation instruction codes into the motion constraint filter and generate physically compliant trajectory segments in combination with the device mechanical structure parameters; Step 22: Input the physically compliant trajectory segment into a pre-trained trajectory feature extractor to decompose it into a three-dimensional feature vector representing the trajectory characteristics. The three-dimensional feature vector is matched with a library of actual trajectory patterns under similar working conditions to generate a set of trajectory offsets. The set of trajectory offsets is input into an adaptive weighting unit, and based on the environmental features contained in the time-synchronized environmental perception data frame, weights are dynamically assigned to generate a comprehensive correction factor. In step 23, the physically compliant trajectory segments and the comprehensive correction factors are input into the trajectory synthesizer, and the pose correction is performed frame by frame through time axis alignment to obtain the corrected trajectory segments. The corrected trajectory segments are then smoothed and processed to eliminate trajectory jitter, generating an optimized continuous motion trajectory as the predicted motion trajectory of the device.

[0008] Furthermore, in step 3, the environmental perception data is integrated with the preset building information model and real-time obstacle scanning data to generate a spatial risk field, including: Step 31: Receive the time-synchronized environmental perception data frame and extract the de-noised point cloud data from it; input the de-noised point cloud data into the spatial registration process, align it with the coordinate system of the preset building information model, and generate a registered point cloud; input the registered point cloud into the voxelization processor, segment it into equal-sized 3D grid cells, and mark the static obstacle occupied area; Step 32: Extract the obstacle radial velocity vector and RGB-D depth map from the time-synchronized environmental perception data frame; use the obstacle radial velocity vector, RGB-D depth map, and de-noised point cloud data to generate a future time series position set of the dynamic obstacle; and calculate the spatiotemporal distribution probability of the dynamic obstacle in the three-dimensional grid cells based on the future time series position set to generate a spatiotemporal probability distribution of the dynamic obstacle. Step 33: Generate an initial risk field based on the static obstacle occupancy area markers and the spatiotemporal probability distribution of dynamic obstacles; and perform normalization processing on the initial risk field to generate a spatial risk field.

[0009] Furthermore, in step 4, based on the spatial risk field and the predicted motion trajectory, two calibration reference points with fixed spatial positions are selected within the target analysis domain, and a reference direction axis is generated based on the spatial relationship between the two calibration reference points; a continuous spatial calibration interval is divided along the reference direction axis; historical positioning data within the spatial calibration interval is extracted, and a set of position offset vectors between the predicted trajectory points and the actual trajectory points is calculated based on the trajectory points on the predicted motion trajectory; the weighted sum of the position offset vectors is minimized as the optimization goal to obtain a three-dimensional spatial transformation parameter optimization model; based on the optimization model, a quasi-Newton optimization algorithm is used to solve the optimization model to generate a final calibration parameter set, including: Step 41: Based on the three-dimensional spatial distribution characteristics of the spatial risk field, two calibration reference points with fixed spatial positions are selected within the target analysis domain; a reference direction axis is generated based on the spatial coordinates of the two calibration reference points, and direction cosines are calculated; and equal-spaced segmentation is performed along the reference direction axis, with the segmentation points as boundaries to generate continuous spatial calibration intervals. Step 42: extract the historical positioning data of the device in each spatial calibration interval to obtain an actual trajectory point set; extract a predicted trajectory point set that matches the timestamp of the historical positioning data based on the predicted motion trajectory; calculate the spatial offset between the predicted trajectory point and the actual trajectory point point by point to generate a position offset vector set; Step 43: performing normalized weighted processing on the position offset vector set to generate a weighted offset vector set; minimizing the sum of squares of the weighted offset vectors is used as the objective function, and combining the three-dimensional space transformation parameter constraints to obtain a three-dimensional space transformation parameter optimization model; In step 44, the objective function is input into the optimization solver, and the parameter estimates of the rotation matrix, translation vector, and scaling factor are initialized; the quasi-Newton optimization algorithm is used to iteratively execute, the objective function gradient vector under the parameters is calculated, the Hessian inverse matrix approximation is updated, and the parameters are updated with step-size adaptiveness in the opposite direction of the gradient; the iteration is terminated when the change in the objective function is lower than the convergence threshold, and the final calibration parameter set consisting of the rotation matrix, translation vector, and scaling factor is output; the iteration is terminated when the change in the objective function is lower than the convergence threshold to obtain the final calibration parameter set.

[0010] Furthermore, in step 5, the predicted motion trajectory is corrected based on the final calibration parameter set. The corrected trajectory is spatially analyzed with the spatial risk field to calculate the collision risk probability and accident severity level, generate a comprehensive risk score, and trigger response actions according to preset thresholds, including: Step 51: Extract the rotation matrix, translation vector, and scaling factor parameters from the final calibration parameter set; input the predicted motion trajectory point set into the space transformer, apply the parameters to perform coordinate transformation calculations, and generate a preliminary correction trajectory; perform time series filtering on the preliminary correction trajectory to eliminate high-frequency jitter noise and generate a final correction trajectory; Step 52: Map the spatial position sequence of the final corrected motion trajectory to the grid coordinate system of the spatial risk field to generate a mapping relationship between trajectory points and grid cells; extract the real-time risk value of the grid cell where each trajectory point is located based on the mapping relationship; Step 53: Calculate a dynamic collision probability value based on the real-time risk value and the geometric distance from the trajectory point to the nearest obstacle; perform an integration operation on the dynamic collision probability value along the trajectory time window to generate an overall collision risk probability value; Step 54 extracts the predefined structural importance coefficient based on the grid cells, and searches the preset accident level comparison table in combination with the overall collision risk probability value to generate the accident severity level; inputs the overall collision risk probability value and the accident severity level into the weighted fusion device to generate a comprehensive risk score; compares the comprehensive risk score with the preset multi-level risk threshold sequence, and outputs a response instruction to trigger the response action.

[0011] Furthermore, in step 4, historical positioning data within the spatial calibration interval is extracted, and a set of position offset vectors between the predicted trajectory points and the actual trajectory points is calculated based on the trajectory points on the predicted motion trajectory; minimizing the weighted sum of the squares of the position offset vectors is used as the optimization goal to obtain a three-dimensional spatial transformation parameter optimization model; based on the optimization model, a quasi-Newton optimization algorithm is used to solve the optimization model to generate a final calibration parameter set, including: Step 61: Based on the timestamp alignment results of the actual trajectory point set and the predicted trajectory point set, perform point-by-point spatial position matching; calculate the straight-line distance difference of each matching point pair in the three-dimensional coordinate system and decompose it into X, Y, and Z axial components; combine the axial components to generate a single-point position offset vector, and summarize the position offset vectors of all timestamps to form a position offset vector set; Step 62: assigning increasing weight coefficients to the position offset vector set according to the time decay characteristic; performing a square operation on the weighted position offset vectors and accumulating the sum of the weighted squares to construct a three-dimensional space transformation parameter optimization model with the goal of minimizing the weighted square sum; Step 63, initialize the rotation angle, translation distance and scaling parameters; iteratively execute based on the quasi-Newton optimization algorithm: calculate the objective function gradient vector under the current parameters, update the Hessian inverse matrix approximation, and perform step-size adaptive parameter adjustment along the opposite direction of the gradient; terminate the iteration when the change in the objective function is lower than the convergence threshold, and output the calibration parameter set consisting of the final rotation angle, translation distance and scaling ratio.

[0012] In a second aspect, a construction equipment operation data intelligent analysis system includes: Acquisition module, used to obtain equipment status data and environmental perception data; The input module is used to obtain the device's predicted motion trajectory based on the device status data and environmental perception data; A fusion module is used to fuse environmental perception data with the preset building information model and real-time obstacle scanning data to generate a spatial risk field; The calculation module is used to select two calibration reference points with fixed spatial positions within the target analysis domain based on the spatial risk field and the predicted motion trajectory, generate a reference direction axis based on the spatial relationship between the two calibration reference points, divide the continuous spatial calibration interval along the reference direction axis, extract historical positioning data within the spatial calibration interval, and calculate the position offset vector set between the predicted trajectory points and the actual trajectory points based on the trajectory points on the predicted motion trajectory; minimize the weighted square sum of the position offset vectors as the optimization goal to obtain a three-dimensional spatial transformation parameter optimization model; and solve the optimization model using a quasi-Newton optimization algorithm based on the optimization model to generate a final calibration parameter set. The output module is used to correct the predicted motion trajectory based on the final calibration parameter set, perform spatial relationship analysis between the corrected trajectory and the spatial risk field, calculate the collision risk probability and accident severity level, generate a comprehensive risk score, and trigger response actions according to preset thresholds.

[0013] According to a third aspect, a computing device includes: one or more processors; The storage device is used to store one or more programs, and when the one or more programs are executed by the one or more processors, the one or more processors implement the method.

[0014] In a fourth aspect, a computer-readable storage medium stores a program, which implements the method when executed by a processor.

[0015] The above solution of the present invention includes at least the following beneficial effects: This method uses a sensor array to collect real-time equipment status data (such as vibration, temperature, and rotational speed) and environmental perception data (such as temperature, humidity, dust concentration, and light intensity). This method combines the 3D spatial coordinate information of a pre-set Building Information Model (BIM) with real-time obstacle scanning data (such as dynamic mapping results from LiDAR and vision sensors), achieving a high-dimensional fusion of the "equipment-environment-space" triad. The BIM model provides spatial parameters for the equipment throughout its lifecycle (such as installation location and planned trajectory), providing a precise spatial reference for environmental data.

[0016] On the other hand, real-time obstacle scanning data dynamically updates information on temporary obstacles (such as construction materials and temporary equipment) in the building environment, avoiding the environmental lag problem of traditional static BIM models, and solving the cumulative error problem caused by multipath effects and signal obstruction in traditional positioning technology in steel structure construction and underground space operation scenarios, providing a reliable position reference for automated construction equipment. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] Figure 1 It is a flow chart of a method for intelligent analysis of construction equipment operation data provided by an embodiment of the present invention.

[0018] Figure 2 This is a schematic diagram of a construction equipment operation data intelligent analysis system provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0019] Exemplary embodiments of the present disclosure will be described in more detail below with reference to the accompanying drawings. Although exemplary embodiments of the present disclosure are shown in the accompanying drawings, it should be understood that the present disclosure can be implemented in various forms and should not be limited by the embodiments set forth herein. Rather, the embodiments are provided to enable a more thorough understanding of the present disclosure and to fully convey the scope of the present disclosure to those skilled in the art.

[0020] like Figure 1 As shown, an embodiment of the present invention provides a method for intelligent analysis of construction equipment operation data, the method comprising the following steps: Step 1: Obtain device status data and environmental perception data; Step 2: Obtain the device's predicted motion trajectory based on the device status data and environmental perception data; Step 3: Fusion the environmental perception data with the preset building information model and real-time obstacle scanning data to generate a spatial risk field; Step 4: Based on the spatial risk field and the predicted motion trajectory, two calibration reference points with fixed spatial positions are selected within the target analysis domain. A reference direction axis is generated based on the spatial relationship between the two calibration reference points. A continuous spatial calibration interval is divided along the reference direction axis. Historical positioning data within the spatial calibration interval is extracted. Based on the trajectory points on the predicted motion trajectory, a set of position offset vectors between the predicted trajectory points and the actual trajectory points is calculated. The optimization goal is to minimize the weighted sum of the squares of the position offset vectors to obtain a three-dimensional spatial transformation parameter optimization model. Based on the optimization model, a quasi-Newton optimization algorithm is used to solve the optimization model and generate the final calibration parameter set. Step 5: Based on the final calibration parameter set, the predicted motion trajectory is corrected, and the spatial relationship between the corrected trajectory and the spatial risk field is analyzed to calculate the collision risk probability and accident severity level, generate a comprehensive risk score, and trigger a response action according to the preset threshold.

[0021] In this embodiment of the present invention, this method uses a sensor array to collect real-time equipment status data (such as vibration, temperature, and rotational speed) and environmental perception data (such as temperature, humidity, dust concentration, and light intensity). This method combines the three-dimensional spatial coordinate information of a pre-defined building information model (BIM) with real-time obstacle scanning data (such as dynamic mapping results from lidar and visual sensors), achieving a high-dimensional fusion of the "equipment-environment-space" triad. On the one hand, the BIM model provides spatial parameters throughout the equipment's lifecycle (such as installation location and planned trajectory path), providing a precise spatial reference for environmental data.

[0022] On the other hand, real-time obstacle scanning data dynamically updates information on temporary obstacles (such as construction materials and temporary equipment) in the building environment, avoiding the environmental lag problem of traditional static BIM models, and solving the cumulative error problem caused by multipath effects and signal obstruction in traditional positioning technology in steel structure construction and underground space operation scenarios, providing a reliable position reference for automated construction equipment.

[0023] In a preferred embodiment of the present invention, step 1, obtaining device status data and environmental perception data, includes: Step 11: Collect raw voltage signals through the strain gauge array and perform analog-to-digital conversion on the raw voltage signals to generate discrete stress data; input the discrete stress data and the number of joint angle pulses collected by the encoder into the kinematic forward solution process to generate real-time spatial pose coordinates; process the real-time spatial pose coordinates, hydraulic pressure sensor readings, motor current sampling values, and control operation instructions to generate a structured device status data matrix; Step 12: Collect original point cloud data in a polar coordinate system using a lidar, and perform coordinate transformation on the original point cloud data to generate a three-dimensional point set in a global coordinate system. Based on the three-dimensional point set, filter out noise points based on the spatial distribution characteristics of adjacent frames to generate de-noised point cloud data. Collect Doppler frequency shift signals using a millimeter-wave radar, perform moving target detection on the frequency shift signals, and calculate obstacle radial velocity vectors. Fusion the de-noised point cloud data, obstacle radial velocity vectors, and the RGB-D depth map collected by the binocular vision sensor generates an environmental perception data packet. Step 13: Time-align the structured device status data matrix with the environment perception data packet, perform interpolation alignment based on the hardware clock synchronization signal, and generate time-synchronized device status data frames and environment perception data frames.

[0024] In the embodiment of the present invention, the above steps can be implemented by the following steps, which are specifically as follows: In step 11 above, the original voltage signal of the building equipment structure is collected in real time through the strain gauge array, and the continuous voltage signal is converted into a discrete digital signal using an analog-to-digital converter (ADC) to form discrete stress data related to the stress distribution of the equipment. The data reflects the stress state of each part of the equipment.

[0025] The discrete stress data is combined with the joint angle pulse number collected by the encoder and processed through the kinematic forward solution algorithm. The process is based on the mechanical structure model of the equipment, and the angles of each joint (converted from the pulse number) are substituted into the kinematic equation to calculate the real-time spatial pose coordinates of the end effector or key components of the equipment.

[0026] Angle conversion: Based on the encoder's resolution (e.g., 1000 pulses per revolution), convert the number of pulses to actual angles. For example, if the encoder outputs 1000 pulses per revolution, each pulse corresponds to 0.36° (360° / 1000). Assuming a joint collects 500 pulses, the rotation angle is 500 × 0.36° = 180°. The encoder's initial position (zero position) must be considered to ensure baseline consistency for angle calculations.

[0027] Starting from the device base (the origin of the coordinate system), the angle of the first joint (converted from the encoder) is substituted into the coordinate transformation (such as rotation or translation) of the joint relative to the base to obtain the position and posture of the end of the first link.

[0028] Taking the end of the first link as the new reference point, substituting the angle of the second joint, and combining the link length parameters, calculate the coordinate transformation of the end of the second link (for example, after joint 2 rotates by an angle of θ2, how does the position of the end of link 2 change in the coordinate system of joint 1); process each joint in turn, and gradually transfer the angle changes of each joint to the end effector through the "coordinate transformation recursion" method. For example, for a three-joint robotic arm, calculate the transformation of joint 1→joint 2→joint 3 in turn, and finally obtain the position (X, Y, Z) and posture (pitch angle, yaw angle, roll angle) of the end effector in the base coordinate system.

[0029] The real-time spatial pose coordinates, hydraulic pressure sensor readings (reflecting the hydraulic load), motor current sampling values ​​(reflecting the motor working status) and control operation instructions (such as movement, start and stop signals) are encoded and integrated into a structured equipment status data matrix according to time series and data type. Each element in the matrix corresponds to a multi-dimensional equipment status parameter at a certain moment.

[0030] In step 12 above, the lidar collects raw point cloud data in a polar coordinate system and converts it into a three-dimensional point set in a global coordinate system through a coordinate transformation matrix (including rotation and translation parameters), so that the point cloud data at different times have a unified spatial reference; dynamic noise reduction is performed based on the spatial distribution characteristics of the point clouds of adjacent frames (such as point density and distance continuity), and outlier noise points are filtered out through clustering or statistical filtering methods to generate noise-reduced point cloud data, thereby improving the accuracy of environmental modeling.

[0031] Millimeter-wave radar collects the Doppler frequency shift signal reflected by the target, extracts the frequency shift characteristics through frequency domain analysis (such as FFT transform), and calculates the radial velocity vector of the obstacle based on the radar detection principle to determine the speed and direction of moving objects in the environment.

[0032] The de-noised point cloud data, obstacle radial velocity vectors, and RGB-D depth maps collected by the binocular vision sensor are fused. Through coordinate registration (aligning the visual data and point cloud data to the same coordinate system) and data association (matching multiple sources of data for the same physical target), an environmental perception data package containing the environment's 3D structure, target position, and motion status is generated.

[0033] In step 13 above, the device status data matrix and the environmental perception data packet are input into the time alignment module. Based on the hardware clock synchronization signal (such as the GPS clock or synchronization pulse), the sensor data with different sampling frequencies are interpolated and aligned (such as linear interpolation or spline interpolation). This eliminates the time offset of each sensor data and generates device status data frames and environmental perception data frames with the same timestamp to ensure data timing consistency.

[0034] The time-synchronized device status data frame is output to the physical-machine learning fusion model for analyzing the device operating status or predicting faults. The environmental perception data frame is output to the spatial risk field construction module for building a real-time environmental model and identifying potential risk areas.

[0035] In an embodiment of the present invention, a sensor group covers both equipment status and environmental perception, enabling comprehensive monitoring of the operating status of construction equipment and improving the real-time and integrity of data collection. Kinematic forward solution and point cloud noise reduction ensure the accuracy of equipment pose calculation and three-dimensional environmental modeling, providing a reliable basis for equipment control and risk assessment. The fusion of LiDAR, millimeter-wave radar, and visual data compensates for the limitations of a single sensor (such as LiDAR's advantages in low light and visual color information), enhancing the ability to perceive dynamic obstacles and complex environments. Hardware clock-based interpolation alignment eliminates data timing deviations, avoids decision-making errors due to time asynchrony, and provides a unified time reference for subsequent model analysis and risk map construction, improving the reliability and security of decision-making.

[0036] In a preferred embodiment of the present invention, step 2, obtaining a predicted motion trajectory of the device based on the device state data and the environmental perception data, includes: Step 21: Receive the time-synchronized device status data frame and extract the real-time spatial pose coordinates and control operation instruction codes from it; input the real-time spatial pose coordinates into the device dynamics equation library and generate initial motion trajectory segments based on the rigid body kinematics principle; input the initial motion trajectory segments and the control operation instruction codes into the motion constraint filter and generate physically compliant trajectory segments in combination with the device mechanical structure parameters; Step 22: Input the physically compliant trajectory segment into a pre-trained trajectory feature extractor to decompose it into a three-dimensional feature vector representing the trajectory characteristics. The three-dimensional feature vector is matched with a library of actual trajectory patterns under similar working conditions to generate a set of trajectory offsets. The set of trajectory offsets is input into an adaptive weighting unit, and based on the environmental features contained in the time-synchronized environmental perception data frame, weights are dynamically assigned to generate a comprehensive correction factor. In step 23, the physically compliant trajectory segments and the comprehensive correction factors are input into the trajectory synthesizer, and the pose correction is performed frame by frame through time axis alignment to obtain the corrected trajectory segments. The corrected trajectory segments are then smoothed and processed to eliminate trajectory jitter, generating an optimized continuous motion trajectory as the predicted motion trajectory of the device.

[0037] In the embodiment of the present invention, the above steps can be implemented by the following steps, which are specifically as follows: In step 21 above, two types of core information are extracted from the time-synchronized device status data frame: real-time spatial pose coordinates, including the device's current three-dimensional position (x, y, z) - representing the absolute position in the global coordinate system, and attitude parameters (such as pitch angle, yaw angle, and roll angle) - describing the device's own direction and rotation state; control operation instruction encoding, such as motor speed instructions and joint angle adjustment instructions, reflecting the action that the control operation expects the device to perform (for example, "bend the left arm 30 degrees" and "move forward at a speed of 0.5m / s").

[0038] The real-time posture coordinates are input into the device dynamic equation library (with a built-in rigid body kinematics model). Based on the rigid body kinematics principle, the predicted motion state for the next time period (such as 0.1 seconds or 0.5 seconds in the future) is calculated in three steps: according to the current posture change rate (such as the position difference and posture change from the previous moment to the current moment), combined with the speed / acceleration target in the operation instruction (for example, the instruction requires "accelerate to 1m / s"), the current motion speed (unit: m / s or degrees / s) and acceleration (unit: m / s) are estimated. 2 or degrees / second 2 The speed range is determined by the maximum operating speed of the device (e.g., the maximum speed of a wheeled robot is 2 m / s, and the maximum rotation speed of a robotic arm joint is 180 degrees / s). The acceleration is limited by the power of the device (e.g., the maximum acceleration corresponding to the maximum torque of the motor).

[0039] Based on the current position, velocity, and acceleration, assuming that the motion is a uniform acceleration or variable acceleration process (smooth transition according to the operation instruction), the three-dimensional position of each time point in the next time period is calculated. For example, if the current velocity is 0.5m / s and the acceleration is 0.2m / s 2 The position in the next 0.1 seconds is calculated as "position = current position + speed × time + 0.5 × acceleration × time 2” logical recursion to ensure that the position coordinates are within the device workspace (for example, the end of the robotic arm must not enter the prohibited area, and the mobile robot must not exceed the preset boundaries).

[0040] Combined with the attitude angle change rate and the attitude adjustment target in the operation instruction (such as "turn right 15 degrees"), the attitude parameters (angles) for the next time period are calculated through Euler angle transformation or quaternion interpolation to ensure continuous attitude changes (such as avoiding non-physical movements such as instantaneous flipping).

[0041] Finally, an initial motion trajectory segment is generated, which is a sequence of continuous predicted position and posture points (such as one point every 0.01 seconds), describing the motion trend of the device under the ideal physical model. The initial trajectory segment and the operation instruction code are input into the motion constraint filter together, and double verification is performed in combination with the mechanical structure parameters of the device (such as the hardware limitations set by the factory): check whether each predicted posture point meets the device hardware limitations: for the robotic arm, verify whether the joint angle is within the allowable range (such as the shoulder joint rotation angle -90° to +90°) and whether there is interference in the connecting rod (such as the end effector must not collide with the base); for mobile equipment, check whether the wheel speed exceeds the maximum speed and whether the steering angle causes the risk of slipping (such as the minimum turning radius limit of Ackerman steering).

[0042] Eliminate or correct the pose points that exceed the constraints (such as adjusting the out-of-bounds joint angle to the threshold boundary); if the speed at a certain point in time suddenly increases beyond the maximum acceleration allowed range (such as the device stipulates that the acceleration shall not exceed 1m / s 2 ), a transition point is inserted to make the speed change gradient meet the limit; for the "jump" in the posture change (such as an instantaneous 180-degree rotation), an intermediate posture is generated by interpolation to ensure smooth and executable motion.

[0043] In step 22 above, the physically compliant trajectory segment is input into a pre-trained trajectory feature extractor (based on deep learning or statistical models), and a three-dimensional feature vector is extracted from the time series data of the trajectory. The features respectively characterize the core attributes of the trajectory: motion trend feature: reflects whether the trajectory as a whole is moving in a straight line, turning in a curve, or reciprocating motion, for example, quantified by information such as trajectory inflection point density and direction change frequency; speed fluctuation feature: describes the smoothness of the speed at each time point, for example, through indicators such as speed standard deviation and acceleration extreme value; spatial distribution feature: characterizes the degree of concentration or dispersion of the trajectory in the posture space, for example, through parameters such as the distribution range of position coordinates and the fluctuation amplitude of posture angles; the final generated three-dimensional feature vector, in which the value range of each dimension is defined according to the specific feature (such as the value after 0-1 normalization), comprehensively describes the behavior pattern of the current trajectory.

[0044] The three-dimensional feature vector is input into the historical behavior analysis module and matched against a library of historically stored actual trajectory patterns. Each record in the pattern library corresponds to a typical operating condition (such as unloaded level ground operation, loaded climbing, high-speed obstacle avoidance, etc.), and each record contains the feature vector and actual motion trajectory for that condition. Similarity calculation logic: Similarity is determined by calculating the "closeness" between the current feature vector and historical feature vectors. For example, this involves calculating the distance between the vectors (the smaller the distance, the higher the similarity) or the consistency of the vector directions (the closer the directions, the higher the similarity). The similarity results are normalized to a value range of 0 to 1 (0 indicates complete dissimilarity, 1 indicates complete consistency).

[0045] Filter out the 3-5 most similar historical working conditions from high to low similarity, retrieve the corresponding actual trajectories, compare the physically compliant trajectories with the historical trajectories point in time, calculate the position deviation (front-back, left-right, up-down distance difference) and attitude deviation (difference in pitch, yaw, and roll angles), and form a set of trajectory offsets (each similar working condition corresponds to a set of offsets). The trajectory offset sets of each similar working condition are input into the adaptive weighting unit. The unit assigns a weight between 0 and 1 (the sum of all weights is 1) to the offset of each similar working condition based on real-time environmental characteristics (such as obstacle distance, equipment load, ground roughness, etc.).

[0046] Working conditions with high similarity and matching environments: If the feature vector similarity of a historical working condition is ≥0.8, and the current environmental parameters (such as load weight) are slightly different from the historical environmental records of the working condition, a higher weight is assigned (such as 0.3-0.4, and the weight of a single working condition does not exceed 0.5); working conditions with medium similarity or environmental differences: If the similarity is between 0.5-0.8, or there is a certain change in the environmental parameters, a medium weight is assigned (such as 0.1-0.3); working conditions with low similarity: When the similarity is <0.5, the weight approaches 0 and the offset is ignored.

[0047] Environmental complexity will adjust the overall weight tendency: when the environmental sensor detects complex situations such as approaching obstacles and bumpy ground, the total weight of all similar working condition offsets will be increased (for example, the total weight will be increased from 0.5 to 0.7) to enhance the correction effect of historical experience on the trajectory; when the environment is stable, the total weight will be reduced (for example, the total weight will be 0.3-0.4), and the physical model prediction will be mainly used; each offset is multiplied by the corresponding weight and then accumulated to generate a comprehensive correction factor that includes the direction and amplitude of position and attitude correction.

[0048] In step 23 above, the physically compliant trajectory segments and the comprehensive correction factors are input into the trajectory synthesizer. The trajectory points are aligned along the time axis and the correction factors are applied to each pose point. Position correction: weighted adjustment of the (x, y, z) coordinates based on the offset; attitude correction: adjustment of the rotation parameters based on the attitude deviation to generate the corrected trajectory segments.

[0049] The corrected trajectory segments are smoothed to eliminate jitter or sudden changes in the trajectory caused by the correction, and finally a continuous, smooth optimized motion trajectory is output and transmitted to the calibration reference point processing process (such as for path planning or control execution).

[0050] In an embodiment of the present invention, an initial trajectory is generated through a dynamic model and combined with mechanical constraint filtering to ensure that the trajectory is physically feasible and avoid mechanical damage caused by excessive device motion. Historical actual trajectories are used to correct deviations, compensating for nonlinear factors not considered by purely physical models (such as friction loss and sensor errors), making the predicted trajectory closer to the actual operating state. Correction weights are dynamically adjusted based on environmental characteristics, allowing trajectory prediction to adapt to different operating conditions (such as complex terrain and load variations), reducing trajectory deviations caused by environmental uncertainty and improving the stability and reliability of device motion. Eliminating trajectory jitter reduces impact loads during device motion and extends mechanical life. Furthermore, a continuous and smooth trajectory helps improve the positioning accuracy of the end effector (for tasks such as robotic grasping and welding), reducing operational errors and ensuring the safety of task execution. The physical model provides basic dynamic constraints, reducing the search space for machine learning; machine learning compensates for modeling errors in the physical model. The combination of the two improves the long-term accuracy of trajectory prediction while ensuring computational efficiency.

[0051] In a preferred embodiment of the present invention, the environmental perception data is integrated with the preset building information model and the real-time obstacle scanning data to generate a spatial risk field, including: Step 31: Receive the time-synchronized environmental perception data frame and extract the de-noised point cloud data from it; input the de-noised point cloud data into the spatial registration process, align it with the coordinate system of the preset building information model, and generate a registered point cloud; input the registered point cloud into the voxelization processor, segment it into equal-sized 3D grid cells, and mark the static obstacle occupied area; Step 32: Extract the obstacle radial velocity vector and RGB-D depth map from the time-synchronized environmental perception data frame; use the obstacle radial velocity vector, RGB-D depth map, and de-noised point cloud data to generate a future time series position set of the dynamic obstacle; and calculate the spatiotemporal distribution probability of the dynamic obstacle in the three-dimensional grid cells based on the future time series position set to generate a spatiotemporal probability distribution of the dynamic obstacle. Step 33: Generate an initial risk field based on the static obstacle occupancy area markers and the spatiotemporal probability distribution of dynamic obstacles; and perform normalization processing on the initial risk field to generate a spatial risk field.

[0052] In the embodiment of the present invention, the above steps can be implemented by the following steps, which are specifically as follows: In step 31 above, after obtaining the original point cloud from the time-synchronized environment perception data frame (such as lidar and millimeter-wave radar data), two filtering processes are performed to remove noise and perform statistical outlier filtering: Set a threshold for the number of neighborhood points (usually 5-50, adjusted according to the density of the environmental point cloud, with a smaller value for dense scenes such as indoors and a larger value for open scenes such as outdoors), and calculate the average distance between each point and the neighboring points; if the distance of a point exceeds "average value + 1-3 times the standard deviation" (1 times filters mild noise, 3 times retains more points), it is determined to be an outlier noise point and removed, retaining the valid point cloud that conforms to statistical laws; voxel grid downsampling: divide the three-dimensional space into equal-sized cubic grids (voxels), and the grid side length is set according to sensor accuracy and computational efficiency (usually 0.05-0.5 meters, small size is used for high-precision scenes such as robot navigation, and large size is used for macro modeling). Only one representative point (such as the center point) is retained for each voxel, reducing the amount of data while retaining the outline of the environment.

[0053] The denoised point cloud is spatially matched with the preset building information model (BIM): significant structural features in the point cloud (such as wall corners, column vertices, and stair corners) and corresponding features in the BIM model are extracted, and feature point pairs are established through manual labeling or automatic recognition by algorithms (such as point A in the point cloud corresponds to point A′ in the BIM).

[0054] Based on the position difference of the feature point pairs, the rotation and translation parameters are calculated. Rotation parameters: adjust the point cloud direction through Euler angles (pitch angle, yaw angle, roll angle) to make the axis of the point cloud coordinate system consistent with the BIM global coordinate system (for example, the X axis points to the east side of the building, and the Z axis is vertically upward). The angle range is limited to -180° to 180° to avoid direction confusion; Translation parameters: calculate the overall position offset of the point cloud on the X, Y, and Z axes (for example, in a building scene, it is usually -100 meters to 100 meters), and ensure numerical stability through normalization processing (scaling according to the maximum size of the scene). Finally, through the rotation and translation transformation matrix, the point cloud coordinate system is aligned to the BIM global coordinate system to generate a registered point cloud (the spatial positions of the two completely overlap).

[0055] The registered point cloud is input into the voxelization processor, which divides the three-dimensional space into grid cells of equal size (e.g., 0.1m × 0.1m × 0.1m, adjusted according to the scene accuracy requirements): each grid cell is checked one by one. If it contains point cloud data (i.e., occupied by static objects such as walls and fixed equipment), it is marked as a "static obstacle occupied area"; if it is empty (no point cloud), it is marked as a "passable area".

[0056] Through processing, the continuous point cloud data is converted into a discrete three-dimensional grid matrix, each grid carries a "place occupied" or "passing" status label, forming a structured representation of the static environment.

[0057] In step 32 above, dynamic obstacles are separated from the time-synchronized frames of environmental perception data (e.g., two consecutive frames of lidar point clouds or camera RGB-D images) through the following steps: comparing the point cloud data of the current frame with the previous frame, marking the points whose positions have changed (i.e., dynamic points); the point cloud positions of stationary objects should basically coincide (allowing for slight noise errors); using a density clustering algorithm (e.g., clustering adjacent dynamic points into a group based on the distance between the point clouds), the dynamic points are divided into independent obstacle targets (e.g., pedestrians, vehicles, mobile robotic arms), with each cluster corresponding to one mobile obstacle.

[0058] Extract the coordinates of the cluster center point of the obstacle in the current frame (such as the three-dimensional position (x1, y1, z1)) and the corresponding center point coordinates in the previous frame (x0, y0, z0), and calculate the displacement vector between the two frames (Δx=x1-x0, Δy=y1-y0, Δz=z1-z0). The direction of the vector is the movement direction of the obstacle (for example, from (0, 0, 0) to (1, 0, 0) indicates movement along the positive direction of the X axis).

[0059] It is known that the time interval between two frames of data is Δt (such as 0.1 second), and the velocity is the modulus of the displacement vector divided by Δt, that is, the velocity is equal to the square root of the sum of the squares of Δx, Δy, and Δz divided by Δt, and the unit is meter per second (if it is a two-dimensional plane motion, ignore the Z axis).

[0060] If the calculated speed exceeds the reasonable range preset by the device (for example, a pedestrian speed exceeding 5m / s is considered noise), the current value is replaced by the average speed of the previous frames to ensure the rationality of the speed data.

[0061] Combine the RGB-D depth map (which provides obstacle shape and distance information) with the aforementioned velocity vector: Use the depth map pixel grayscale values ​​to identify obstacle boundaries and determine their dimensions (e.g., a pedestrian is approximately a 1.8-meter-tall, 0.5-meter-wide cylinder). Assuming the obstacle maintains its current direction and speed (uniform linear motion) in the short term, generate multiple predicted positions in a time series (e.g., 0.5 seconds, 1 second, and 1.5 seconds in the future) (position at each time point = current position + speed × time), forming a future time series position set.

[0062] Random perturbations are added to each predicted position (simulating possible speed fluctuations or changes in direction), and the probability of each three-dimensional grid cell being occupied by an obstacle at different times is calculated. The closer the grid is to the current position and the shorter the time, the higher the probability of being occupied (for example, the probability of occupying the grid where the current position is located in the next 0.5 seconds is set to 0.8, and after 1 second it drops to 0.6 due to increased uncertainty). The probability value is represented by color depth (0 is white, 1 is red), forming a three-dimensional heat map with a time dimension (XY axes are spatial coordinates, and Z axis is the time axis), which intuitively shows high-risk areas where dynamic obstacles may appear in the future.

[0063] In step 33 above, the static obstacle occupied area (marked as high risk) and the spatiotemporal distribution heat map are input into the risk superimposer, and the comprehensive risk value is calculated for each grid cell. The risk value of the grid occupied by static obstacles is directly set to the maximum value. The risk value of the grid that may be occupied by dynamic obstacles is linearly mapped according to the probability of the heat map (for example, a probability of 0.8 corresponds to a risk value of 0.8). The original risk distribution map is generated (each grid cell has a risk value between 0 and 1).

[0064] Combined with the load-bearing structure data in the building information model (such as the location of load-bearing walls and columns), the original risk distribution map is strengthened in key areas: the risk value of grids close to load-bearing walls and columns is increased by 20%-50%, regardless of whether there are obstacles (because collisions may cause structural damage). In key passage areas such as stairs and elevator shafts, the risk value of dynamic obstacles is additionally weighted (for example, multiplied by 1.5 times) to ensure safety priority.

[0065] The enhanced risk distribution map is globally normalized, mapping the risk values ​​of all grids to a range of 0-1 (e.g., using a linear transformation to make the maximum value 1 and the minimum value 0), generating a standardized spatial risk field. The map includes the location of static obstacles, the future risk distribution of dynamic obstacles, and the risk level of key areas of the building, and is output to the calibration benchmark processing process (e.g., for avoiding high-risk areas during path planning). In an embodiment of the present invention, the spatiotemporal distribution of static building structures and dynamic obstacles is taken into account simultaneously, avoiding the omission of dynamic risks (such as the sudden intrusion of moving objects) caused by relying solely on static maps, and improving the real-time and integrity of environmental modeling. Combined with the load-bearing structure information of the BIM model, the key areas are risk-weighted, and priority is given to protecting building safety (such as avoiding collisions with load-bearing walls). It is suitable for scenarios such as factories and buildings that have high requirements for structural protection. The uncertainty of dynamic obstacles is quantified through a probability diffusion model, and the risk distribution is intuitively represented in the form of a heat map, which facilitates the planning algorithm to dynamically adjust the path according to the risk level (such as bypassing high-risk areas and passing through low-risk areas). The normalized risk map uses a unified numerical standard (0-1) to represent the risk level, which can be directly connected to modules such as path planning and motion control, reducing the data conversion cost between different areas and improving overall decision-making efficiency.

[0066] In a preferred embodiment of the present invention, step 4 selects two calibration reference points with fixed spatial positions within the target analysis domain based on the spatial risk field and the predicted motion trajectory, generates a reference direction axis based on the spatial relationship between the two calibration reference points, divides a continuous spatial calibration interval along the reference direction axis, extracts historical positioning data within the spatial calibration interval, and calculates a set of position offset vectors between the predicted trajectory points and the actual trajectory points based on the trajectory points on the predicted motion trajectory; minimizes the weighted sum of the squares of the position offset vectors as the optimization goal to obtain a three-dimensional spatial transformation parameter optimization model; and solves the optimization model using a quasi-Newton optimization algorithm based on the optimization model to generate a final calibration parameter set, including: Step 41: Based on the three-dimensional spatial distribution characteristics of the spatial risk field, two calibration reference points with fixed spatial positions are selected within the target analysis domain; a reference direction axis is generated based on the spatial coordinates of the two calibration reference points, and direction cosines are calculated; and equal-spaced segmentation is performed along the reference direction axis, with the segmentation points as boundaries to generate continuous spatial calibration intervals. Step 42: extract the historical positioning data of the device in each spatial calibration interval to obtain an actual trajectory point set; extract a predicted trajectory point set that matches the timestamp of the historical positioning data based on the predicted motion trajectory; calculate the spatial offset between the predicted trajectory point and the actual trajectory point point by point to generate a position offset vector set; Step 43: performing normalized weighted processing on the position offset vector set to generate a weighted offset vector set; minimizing the sum of squares of the weighted offset vectors is used as the objective function, and combining the three-dimensional space transformation parameter constraints to obtain a three-dimensional space transformation parameter optimization model; In step 44, the objective function is input into the optimization solver, and the parameter estimates of the rotation matrix, translation vector, and scaling factor are initialized. The quasi-Newton optimization algorithm is used for iterative execution to calculate the objective function gradient vector under the parameters, update the Hessian inverse matrix approximation, and perform step-size adaptive parameter updates in the opposite direction of the gradient. When the change in the objective function is lower than the convergence threshold, the iteration is terminated, and the final calibration parameter set consisting of the rotation matrix, translation vector, and scaling factor is output.

[0067] In the embodiment of the present invention, the above steps can be implemented by the following steps, which are specifically as follows: In step 41 above, the selection of static structural feature points (such as the tops of building columns, corners of fixed equipment, and intersections of walls) from the spatial risk field must meet three conditions: fixed position: they must be permanent and immovable structures, excluding dynamic objects (such as pedestrians and movable equipment); recognizability: they must be stably detected by sensors (such as lidar and visual cameras) and have obvious features (such as right-angle turning points and height mutation points); and reasonable spacing: the distance between the two points must be moderate (usually 5-20 meters), ensuring the pointing accuracy of the direction axis while avoiding overly dense calibration interval division due to being too close. For example, in a factory scenario, the top endpoints of the two load-bearing columns in the workshop are selected as reference points A and B.

[0068] Vector definition: With reference point A as the starting point and reference point B as the end point, a three-dimensional space vector is formed (i.e., a directed line segment pointing from the coordinates of A to the coordinates of B). The components of the vector on the X, Y, and Z coordinate axes are calculated (i.e., the coordinates of point B minus the coordinates of point A to obtain Δx, Δy, and Δz). The total length of the vector (i.e., the three-dimensional distance of the line segment) is calculated. The components of the X, Y, and Z axes are divided by the total length to obtain the cosine values ​​of the three directions. The three values ​​represent the cosines of the angles between the vector and the X, Y, and Z axes, respectively, reflecting the "projection ratio" of the vector in the directions of the three coordinate axes.

[0069] Each direction cosine value is between -1 and 1. When the vector is exactly in the same direction as a coordinate axis, the corresponding direction cosine is 1 (in the same direction) or -1 (in the opposite direction); when the vector is perpendicular to the coordinate axis, the corresponding direction cosine is 0.

[0070] Perform equal-interval segmentation along the reference direction axis (i.e., the line connecting A to B): Set the interval distance (usually 0.5-2 meters) based on the accuracy requirements of the scene. For example, select 0.5-meter interval in high-precision positioning scenarios and 1-meter interval in large-scale scenarios.

[0071] With each segmentation point as the center, it expands toward the plane perpendicular to the reference direction axis to form a continuous cylindrical calibration interval (such as a cylinder with a radius of 1 meter and the direction axis as the center axis). Each interval covers a certain range of three-dimensional space for subsequent local calibration.

[0072] The intervals are numbered in the order of division (such as interval 1, interval 2, etc.) to ensure that the position of each interval on the reference direction axis is unique and continuous.

[0073] In step 42 above, the actual positioning trajectory point set within the current calibration interval (position points collected in real time by sensors such as GPS and lidar with timestamps) is extracted, and the predicted motion trajectory point set with the corresponding timestamps is retrieved at the same time, ensuring that the two sets of points correspond one-to-one in chronological order and that the time difference does not exceed a preset threshold (e.g., 50 milliseconds).

[0074] For each pair of corresponding points, the 3D offset between the predicted and actual trajectory points is calculated (Δx = actual x - predicted x, Δy = actual y - predicted y, Δz = actual z - predicted z), forming a set of position offset vectors. For example, if the actual coordinates of a point are (10, 5, 2) and the predicted coordinates are (10.1, 4.9, 2.1), the offset vector is (-0.1, 0.1, -0.1).

[0075] In step 43, the offset vector is weighted according to the risk value of each point in the spatial risk field (the higher the risk value, the greater the weight). For example, the risk value is mapped to a weight coefficient of 0.5-1.5 (a risk value of 1 corresponds to a weight of 1.5, and a risk value of 0 corresponds to a weight of 0.5), so that the offset of high-risk areas (such as those near load-bearing walls) is corrected first.

[0076] With "minimizing the sum of squares of weighted offset vectors" as the objective function, a three-dimensional space transformation parameter optimization model is constructed, including rotation matrix parameters: the three Euler angles (pitch, yaw, and roll) that describe the rotation of the coordinate system; translation vector parameters: the three-dimensional coordinate offset (tx, ty, tz); and scaling factor parameters: the global scale adjustment coefficient (to handle sensor scale error, usually with an initial value of 1 and a range of 0.95-1.05).

[0077] The model must meet the following constraints: the rotation matrix is ​​orthogonal (to ensure the rationality of the transformation) and the scaling factor is non-negative.

[0078] In step 44 above, the rotation matrix parameters are initially set to the "no rotation" state, that is, the directions of the X, Y, and Z axes in the three-dimensional space are completely consistent with the global coordinate system (equivalent to the object not undergoing any rotation); the translation vector parameters are initially set to (0, 0, 0), indicating that there is no position offset between the predicted trajectory coordinate system and the actual trajectory coordinate system in the initial stage; the scaling factor parameters are initially set to 1, assuming that the predicted trajectory and the actual trajectory have the same scale (no magnification or reduction error).

[0079] The initialization parameters choose the "zero error" assumption as the starting point of iterative optimization, and are subsequently gradually corrected through data-driven methods.

[0080] Calculate the gradient of the objective function (weighted sum of squared offsets) with respect to the rotation angle, translation, and scaling factor under the parameters. The gradient is a vector whose direction indicates the fastest growth direction of the objective function. During iteration, search for the final parameters in the opposite direction of the gradient (the fastest descent direction). For example, if the gradient of the rotation angle θ is positive, it means that reducing θ can reduce the objective function, so the next step is to adjust in the direction of reducing θ.

[0081] The quasi-Newton method does not directly calculate the complex Hessian matrix (second-order derivative matrix). Instead, it iteratively approximates the inverse matrix (used to adjust the "curvature" of the search direction) through historical gradient information. It records the difference (Δx) between the current parameter vector and the previous parameter vector. It also records the difference (Δg) between the current gradient and the previous gradient. Δx and Δg are used to update the approximate value of the inverse Hessian matrix. The approximate value of the inverse Hessian matrix is ​​used to convert the gradient vector into a more optimal search direction, thereby accelerating convergence.

[0082] Try different step sizes along the search direction (the opposite direction of the gradient combined with the inverse Hessian matrix) to find the step size that makes the objective function decrease the most and the calculation stable; preset a maximum step size, and then decrease it proportionally (such as multiplying by 0.5 each time); for each step size, calculate the objective function value under the new parameters, requiring that the "sufficient decrease" condition is met (such as the decrease in the function value exceeds a certain threshold related to the step size); select the maximum step size that meets the conditions as the update step size for this iteration; avoid parameter oscillation caused by too large a step size, or slow convergence caused by too small a step size.

[0083] Add "search direction × final step size" to the current parameters to obtain new rotation matrix parameters, translation vector parameters, and scaling factor parameters. For example, if the current rotation angle is 0.1°, the search direction is to decrease by 0.05°, and the step size is 1, the new rotation angle is updated to 0.05°.

[0084] If the change in the objective function is less than the preset threshold (for example, the function value change between two iterations is less than 0.001), it means that the parameter adjustment can no longer significantly improve the calibration accuracy; the number of iterations reaches the upper limit (for example, 100) to avoid excessive calculation time.

[0085] When the termination conditions are met, the final set of calibration parameters is output. For example, the rotation matrix parameters correspond to three Euler angles (such as pitch angle 0.1°, yaw angle -0.2°, and roll angle 0.3°) to correct the directional deviation of the predicted trajectory; the translation vector parameters, such as (0.05m, -0.03m, 0.01m), compensate for the position offset of the predicted trajectory; and the scaling factor parameter, such as 1.02, adjusts the overall scale of the predicted trajectory to be consistent with the actual one.

[0086] In an embodiment of the present invention, the static structure point selection based on the environmental risk map ensures that the reference point position is fixed and can be repeatedly identified, avoids interference from dynamic obstacles, and improves the long-term stability of the calibration model. The calibration interval is divided along the reference axis, and the offset characteristics of different positions are optimized in sections (such as the difference between near offset and far offset) to solve the problem that global calibration cannot cover local errors. Higher weights are given to trajectory offsets in high-risk areas, and positioning errors that may cause collisions (such as offsets when approaching obstacles) are corrected first, optimizing the motion trajectory while ensuring safety. The quasi-Newton method reduces the amount of calculation by approximating the Hessian matrix. The rotation, translation, and scaling parameters are optimized at the same time, which can handle sensor installation deviations (rotation / translation) and scale errors (scaling), and is suitable for joint calibration of different types of positioning equipment (such as lidar and visual cameras).

[0087] In a preferred embodiment of the present invention, step 5 corrects the predicted motion trajectory based on the final calibration parameter set, performs spatial relationship analysis between the corrected trajectory and the spatial risk field, calculates the collision risk probability and accident severity level, generates a comprehensive risk score, and triggers a response action based on a preset threshold, including: Step 51: Extract the rotation matrix, translation vector, and scaling factor parameters from the final calibration parameter set; input the predicted motion trajectory point set into the space transformer, apply the parameters to perform coordinate transformation calculations, and generate a preliminary correction trajectory; perform time series filtering on the preliminary correction trajectory to eliminate high-frequency jitter noise and generate a final correction trajectory; Step 52: Map the spatial position sequence of the final corrected motion trajectory to the grid coordinate system of the spatial risk field to generate a mapping relationship between trajectory points and grid cells; extract the real-time risk value of the grid cell where each trajectory point is located based on the mapping relationship; Step 53: Calculate a dynamic collision probability value based on the real-time risk value and the geometric distance from the trajectory point to the nearest obstacle; perform an integration operation on the dynamic collision probability value along the trajectory time window to generate an overall collision risk probability value; Step 54 extracts the predefined structural importance coefficient based on the grid cells, and searches the preset accident level comparison table in combination with the overall collision risk probability value to generate the accident severity level; inputs the overall collision risk probability value and the accident severity level into the weighted fusion device to generate a comprehensive risk score; compares the comprehensive risk score with the preset multi-level risk threshold sequence, and outputs a response instruction to trigger the response action.

[0088] In the embodiment of the present invention, the above steps can be implemented by the following steps, which are specifically as follows: In step 51 , three types of parameter rotation matrices are extracted from the final calibration parameter set to correct the directional deviation of the predicted trajectory (e.g., rotating the predicted trajectory by 0.1° around the X-axis to align it with the actual coordinate system); a translation vector is used to compensate for the position offset of the predicted trajectory (e.g., translating all predicted points along the Y-axis by -0.03 meters to eliminate sensor installation position errors); and a scaling factor is used to adjust the overall scale of the predicted trajectory (e.g., amplifying the predicted distance by 1.02 times to match the actual sensor's ranging accuracy).

[0089] Three types of parameters are applied to each predicted trajectory point to generate a preliminary corrected trajectory point (the coordinate transformation process is similar to the combined operation of "first rotate, then scale, and finally translate"). Time series filtering (such as sliding average filtering and median filtering) is performed on the preliminary corrected trajectory point sequence: the filter window length is set (for example, including the current point and 2 points before and after, for a total of 5 points); the coordinates of the current point are replaced by the average value (or median value) of the coordinates of the points in the window to eliminate high-frequency jitter caused by fluctuations in the calibration parameters (such as millimeter-level position jumps), and the final smooth corrected motion trajectory is generated.

[0090] In step 52 , each continuous position point (x, y, z) in the final corrected trajectory is converted to the grid coordinate system of the spatial risk field: given the grid size of the risk map (e.g., 0.1 m × 0.1 m × 0.1 m) and the global coordinate origin, the grid cell index corresponding to each position point is calculated (e.g., the x-coordinate is divided by the grid size and rounded to obtain the grid number in the x-axis direction).

[0091] If the location point is exactly at the intersection of two grids, it will be mapped to the adjacent grids at the same time (risk values ​​are assigned according to distance weights). Based on the mapping results, the risk value of each grid cell is extracted (a value between 0 and 1, with 0 indicating no risk and 1 indicating the highest risk): grids occupied by static obstacles are directly marked with a risk value of 1; for grids in the dynamic obstacle heat map, the risk value is updated in real time according to the probability distribution (for example, if the probability of a grid being occupied in the next 1 second is 0.7, the risk value is marked as 0.7).

[0092] In step 53 , for each grid point, the geometric distance to the nearest obstacle in the grid (e.g., the distance to a static wall, the distance to the predicted position of a dynamic pedestrian) is calculated. The closer the distance, the higher the probability of a dynamic collision (e.g., the probability is 0.3 at a distance of 0.5 meters, and increases to 0.8 at a distance of 0.2 meters). Combining the grid risk value (basic risk) and the geometric distance (dynamic correction) to generate a dynamic collision probability value (range 0-1) for a single trajectory point.

[0093] Set the length of the time window (e.g., trajectory points within the next 2 seconds) and integrate (sum or weighted average) the dynamic collision probability values ​​of all trajectory points within the window: recent trajectory points (e.g., within the next 0.5 seconds) are given a higher weight (e.g., weight 0.8), while distant points (e.g., after 1.5 seconds in the future) are given a lower weight (e.g., weight 0.2); the overall collision risk probability is obtained (e.g., 0.6 means there is a 60% probability of a collision within the next 2 seconds). In step 54 above, the “structural importance coefficient” (predefined parameters, such as 1.5 for load-bearing wall areas and 1.0 for general areas) of the grid where the trajectory point is located is extracted from the risk map. Combined with the preset accident level comparison table (e.g., a coefficient ≥ 1.5 corresponds to “major accident risk” and 1.0-1.5 corresponds to “general accident risk”), the accident severity level (divided into three levels: high, medium, and low) is generated.

[0094] The collision risk probability (0-1) and the accident severity level (quantified by a coefficient of 0.5-1.5) are weighted and fused (e.g., the risk probability accounts for 60% of the weight and the severity accounts for 40%) to obtain a comprehensive score (0-1.5), which is compared with the preset three-level threshold sequence (e.g., high risk threshold 1.2, medium risk 0.8, low risk 0.5): Score ≥ 1.2: triggers an emergency stop command and the equipment brakes immediately; 0.8 ≤ score < 1.2: triggers a deceleration command and the equipment reduces its operating speed by 50%; 0.5 ≤ score < 0.8: triggers an early warning prompt and sends a risk warning to the control operation; score < 0.5: no action, normal operation.

[0095] In this embodiment of the present invention, calibration parameters compensate for sensor installation deviations and scale errors, improving the consistency between the corrected trajectory and actual motion and reducing the collision risk caused by inaccurate positioning. By combining the real-time location of trajectory points, obstacle distances, and the importance of environmental structures, collision risk is converted into quantifiable probability values ​​and severity levels, avoiding the limitations of traditional "unsafe / safe" binary judgments and enabling refined risk management. Based on the comprehensive risk score, different levels of response actions (from warning to emergency stop) are triggered, maximizing equipment efficiency while ensuring safety (e.g., normal operation in low-risk scenarios, slowing down for medium-risk scenarios, and immediate stop in high-risk scenarios). This system not only considers the immediate risk of dynamic obstacles but also incorporates the long-term safety priorities of static structures (e.g., protecting load-bearing walls), forming a dual protection mechanism of "real-time dynamic risk calculation + structural static risk enhancement." This mechanism is suitable for industrial and architectural scenarios with extremely high safety requirements. Temporal filtering eliminates trajectory jitter and avoids misjudgments caused by sensor noise. A weighted fusion algorithm balances different risk factors, enhancing adaptability to complex environments (e.g., areas with multiple obstacles and complex structures).

[0096] In a preferred embodiment of the present invention, step 4 extracts historical positioning data within the spatial calibration interval, calculates a set of position offset vectors between the predicted trajectory points and the actual trajectory points based on the trajectory points on the predicted motion trajectory, minimizes the weighted sum of the squares of the position offset vectors as the optimization goal, and obtains a three-dimensional spatial transformation parameter optimization model; solves the optimization model using a quasi-Newton optimization algorithm based on the optimization model to generate a final calibration parameter set, including: Step 61: Based on the timestamp alignment results of the actual trajectory point set and the predicted trajectory point set, perform point-by-point spatial position matching; calculate the straight-line distance difference of each matching point pair in the three-dimensional coordinate system and decompose it into X, Y, and Z axial components; combine the axial components to generate a single-point position offset vector, and summarize the position offset vectors of all timestamps to form a position offset vector set; Step 62: assigning increasing weight coefficients to the position offset vector set according to the time decay characteristic; performing a square operation on the weighted position offset vectors and accumulating the sum of the weighted squares to construct a three-dimensional space transformation parameter optimization model with the goal of minimizing the weighted square sum; Step 63, initialize the rotation angle, translation distance and scaling parameters; iterate based on the quasi-Newton optimization algorithm, calculate the objective function gradient vector under the current parameters, update the Hessian inverse matrix approximation, and perform step-size adaptive parameter adjustment along the opposite direction of the gradient; terminate the iteration when the objective function change is lower than the convergence threshold, and output the calibration parameter set consisting of the final rotation angle, translation distance and scaling.

[0097] In the embodiment of the present invention, the above steps can be implemented by the following steps, which are specifically as follows: In step 61 above, the actual trajectory point set is the position data collected in real time by the device sensor (such as lidar, visual camera), and each point contains three-dimensional coordinates (x a ,y a , z a ) and timestamp t a ; Prediction trajectory point set: through the theoretical motion trajectory, each point contains the corresponding timestamp t p The three-dimensional coordinates (x p ,y p , z p ).

[0098] Find the point pair with the smallest time difference (usually no more than 50 milliseconds) to ensure that each pair of trajectory points represents the actual position and predicted position of the device at the same time (for example, the actual point t a =10:00:00.123 and the predicted point t p =10:00:00.125 is considered a matching pair).

[0099] For each pair of matching points, calculate the position deviation between the predicted trajectory point and the actual trajectory point in the three coordinate axis directions. The difference in the X-axis direction (left and right offset) is: the X coordinate of the actual point minus the X coordinate of the predicted point, that is, Δx=x a -x p ; If Δx is positive, it means the predicted point is on the left side of the actual point; if it is negative, it means the predicted point is on the right side.

[0100] Y-axis direction difference (front and back offset): the Y coordinate of the actual point minus the Y coordinate of the predicted point, that is, Δy=y a -y p If Δy is positive, it means the predicted point is behind the actual point; if it is negative, it means the predicted point is in front (assuming the Y axis is defined as the direction of the device's forward movement).

[0101] Z-axis direction difference (up and down offset): the Z coordinate of the actual point minus the Z coordinate of the predicted point, that is, Δz=z a -z p If Δz is positive, it means the predicted point is below the actual point; if it is negative, it means the predicted point is above the actual point (applicable to scenarios with height changes, such as vertical movement of a robotic arm).

[0102] The differences between the three coordinate axes are combined into a three-dimensional vector (Δx, Δy, Δz). The direction of the vector represents the offset direction of the predicted trajectory relative to the actual trajectory, and the length represents the straight-line distance offset. For example, if the actual point coordinates are (10, 5, 2) and the predicted point coordinates are (10.1, 4.9, 2.1), then Δx = -0.1 (the predicted point is 0.1 meter to the left), Δy = +0.1 (the predicted point is 0.1 meter behind), and Δz = -0.1 (the predicted point is 0.1 meter below), forming an offset vector (-0.1, 0.1, -0.1).

[0103] The offset vectors of all matching point pairs are aggregated to form a set of position offset vectors containing time series information (for example, 100 offset vectors containing 100 time points, each vector records the three-dimensional offset at the corresponding moment).

[0104] In step 62, each offset vector is assigned an increasing weight coefficient based on its timestamp (e.g., the weight of the most recent 10% of time points is 1, the weight of the next most recent 20% is 0.8, and so on, decreasing in sequence). The logic is that the closer the data is to the current moment, the greater its impact on the real-time calibration and the higher its weight, to prevent outdated data from interfering with the current error judgment.

[0105] For each weighted offset vector, calculate the square value of the X, Y, and Z components (such as (Δx×weight) 2), and then add the square values ​​of the three components to obtain the weighted square value of a single vector; accumulate the weighted square values ​​of all time points to obtain a total value, which is the optimization target value. The smaller the value, the smaller the overall deviation between the predicted trajectory and the actual trajectory.

[0106] With "minimizing the optimization target value" as the core, a three-dimensional spatial transformation model containing three types of parameters is constructed: rotation angle: corrects the directional deviation of the predicted trajectory (such as the rotation angle around the X-axis, Y-axis, and Z-axis); translation distance: compensates for the position deviation of the predicted trajectory (the translation amount in the X-axis, Y-axis, and Z-axis directions); scaling: adjusts the overall scale of the predicted trajectory (such as the global scaling caused by sensor ranging error).

[0107] In the above step 63, the rotation angle is initially set to 0°, assuming that the predicted trajectory and the actual trajectory are in exactly the same direction in the initial state (no rotation deviation).

[0108] Translation distance: Initially set to (0, 0, 0) meters, that is, the origin of the predicted trajectory coordinate system coincides with the origin of the actual trajectory coordinate system, with no position offset; Scaling ratio: Initially set to 1, assuming that the scale of the predicted trajectory is the same as the actual trajectory (for example, a predicted distance of 1 meter corresponds to an actual distance of 1 meter); Key logic: Initialization parameters use the "zero error" assumption as the starting point for iterative optimization, and then gradually correct the deviation based on historical data.

[0109] Calculate the rate of change (gradient) of the optimized target value (weighted squared offset) with respect to the rotation angle, translation distance, and scaling ratio under the current parameters. For example, if the gradient of the rotation angle θ is positive, it means that increasing θ will increase the target value, so the next iteration should adjust in the direction of decreasing θ. If the gradient of the translation distance on the X-axis is negative, it means that increasing the X-axis translation will decrease the target value, so the adjustment should be in the positive direction. The direction of the gradient vector represents the direction of the fastest increase in the target value, and the final parameters are searched in the opposite direction (the direction of the fastest decrease) during iteration.

[0110] The quasi-Newton method approximates the curvature of the target function through historical iteration data (instead of second-order derivative calculation): Record the difference between the current parameter vector and the previous parameter vector (for example, the rotation angle changes from 0.1° to 0.08°, the difference is -0.02°); record the difference between the current gradient and the previous gradient (for example, the X-axis gradient changes from 0.5 to 0.3, the difference is -0.2); use the two sets of differences to update the curvature approximation matrix (similar to "memorizing" the curvature of the objective function) so that the next search direction is closer to the final path. The curvature approximation matrix helps the algorithm determine the "curvature" direction of the parameter adjustment and avoid falling into a local end.

[0111] Combine the reverse gradient with the curvature approximation matrix to obtain the final parameter adjustment direction (such as adjusting the rotation angle and translation distance at the same time), and dynamically adjust the step size: preset the initial step size (such as adjusting the rotation angle by 0.1° and the translation distance by 0.01 meters each time); try different step sizes, calculate the target value under the new parameters, and require the target value to drop by a certain amount to meet the preset conditions (such as a decrease of at least 0.01); if the target value drops significantly, increase the step size to speed up convergence; if it drops slowly or increases, reduce the step size to avoid oscillation (such as multiplying the step size by 0.5). For example: the first iteration step size is 0.1, the target value drops by 0.1, and the next step size is increased to 0.15; if the target value only drops by 0.02, the step size is reduced to 0.05.

[0112] Add "adjustment direction × final step size" to the current parameters to get the new parameter values: Rotation angle = current angle + adjustment direction × step size (for example, adjusting from 0.1° to 0.05°); Translation distance = current distance + adjustment amount of each axis × step size (for example, adjusting the X axis from 0.02 meters to 0.025 meters); Scaling ratio = current ratio + adjustment amount × step size (for example, adjusting from 1.01 to 1.012).

[0113] If the target value changes for three consecutive iterations are all less than the set threshold (such as 0.001), it means that the parameter adjustment has not significantly improved the optimization target. The number of iterations reaches the upper limit (such as 100) to avoid excessive calculation time. When the termination condition is met, the parameters of the final iteration are extracted: rotation angle: such as 0.3° rotation around the X-axis and -0.2° rotation around the Y-axis to correct the directional deviation of the predicted trajectory; translation distance: such as 0.03 meters translation on the X-axis and -0.01 meters translation on the Y-axis to compensate for position offset; scaling ratio: such as 1.02 to adjust the scale of the predicted trajectory to be consistent with the actual one. The parameters are combined to generate the final calibration parameter set.

[0114] In an embodiment of the present invention, by assigning a higher weight to recent data, trajectory deviations in the current environment are corrected first, preventing interference from early noise data on real-time calibration and improving the dynamic adaptability of the calibration model. The three-dimensional offset is decomposed into the X, Y, and Z axes for separate optimization. The quasi-Newton algorithm reduces the amount of computation through curvature approximation, and typically converges within 10-20 iterations, making it suitable for online real-time calibration of devices (e.g., updating calibration parameters once per second), avoiding the time-consuming issues of traditional algorithms. Multiple error types cover the optimization of three types of parameters: rotation, translation, and scaling. It can handle multiple errors, including sensor installation angle deviation (rotation), fixed position offset (translation), and ranging scale error (scaling), and is suitable for joint calibration of positioning devices of different brands and types. Automatic optimization based on historical trajectory data eliminates the need for manually pre-setting error models and can adapt to error characteristics in different scenarios (e.g., ranging deviation caused by multiple reflections indoors, GPS signal drift outdoors), improving reliability in complex environments.

[0115] like Figure 2 As shown, an embodiment of the present invention further provides a construction equipment operation data intelligent analysis system, comprising: Acquisition module, used to obtain equipment status data and environmental perception data; The input module is used to obtain the device's predicted motion trajectory based on the device status data and environmental perception data; A fusion module is used to fuse environmental perception data with the preset building information model and real-time obstacle scanning data to generate a spatial risk field; The calculation module is used to select two calibration reference points with fixed spatial positions within the target analysis domain based on the spatial risk field and the predicted motion trajectory, generate a reference direction axis based on the spatial relationship between the two calibration reference points, divide the continuous spatial calibration interval along the reference direction axis, extract historical positioning data within the spatial calibration interval, and calculate the position offset vector set between the predicted trajectory points and the actual trajectory points based on the trajectory points on the predicted motion trajectory; minimize the weighted square sum of the position offset vectors as the optimization goal to obtain a three-dimensional spatial transformation parameter optimization model; and solve the optimization model using a quasi-Newton optimization algorithm based on the optimization model to generate a final calibration parameter set. The output module is used to correct the predicted motion trajectory based on the final calibration parameter set, perform spatial relationship analysis between the corrected trajectory and the spatial risk field, calculate the collision risk probability and accident severity level, generate a comprehensive risk score, and trigger response actions according to preset thresholds.

[0116] It should be noted that this system is a system corresponding to the above method, and all implementation methods in the above method embodiment are applicable to this embodiment and can achieve the same technical effects.

[0117] An embodiment of the present invention further provides a computing device comprising: a processor and a memory storing a computer program, wherein the computer program, when executed by the processor, performs the above-described method. All implementations in the above-described method embodiments are applicable to this embodiment and can achieve the same technical effects.

[0118] The embodiment of the present invention further provides a computer-readable storage medium storing instructions, which, when executed on a computer, causes the computer to execute the above-described method. All implementations in the above-described method embodiment are applicable to this embodiment and can achieve the same technical effects.

[0119] The above is a preferred embodiment of the present invention. It should be pointed out that for ordinary technicians in this technical field, several improvements and modifications can be made without departing from the principles of the present invention. These improvements and modifications should also be regarded as within the scope of protection of the present invention.

Claims

1. A method for intelligent analysis of construction equipment operation data, characterized in that: The method comprises Step 1: Obtain device status data and environmental perception data; Step 2: Obtain the device's predicted motion trajectory based on the device status data and environmental perception data; Step 3: Fusion the environmental perception data with the preset building information model and real-time obstacle scanning data to generate a spatial risk field; Step 4: Based on the spatial risk field and the predicted motion trajectory, two calibration reference points with fixed spatial positions are selected within the target analysis domain, and a reference direction axis is generated according to the spatial relationship between the two calibration reference points. Divide the continuous spatial calibration intervals along the reference direction axis; Extract historical positioning data within the spatial calibration interval and calculate the position offset vector set between the predicted and actual trajectory points based on the trajectory points on the predicted motion trajectory. Minimize the weighted sum of the squares of the position offset vectors as the optimization goal to obtain a three-dimensional spatial transformation parameter optimization model. Based on the optimization model, use the quasi-Newton optimization algorithm to solve the optimization model and generate the final calibration parameter set. Step 5: Based on the final calibration parameter set, the predicted motion trajectory is corrected, and the spatial relationship between the corrected trajectory and the spatial risk field is analyzed to calculate the collision risk probability and accident severity level, generate a comprehensive risk score, and trigger a response action according to the preset threshold.

2. A method for intelligent analysis of construction equipment operation data according to claim 1, characterized in that: Step 1: Obtain device status data and environmental perception data, including: Step 11: Collect raw voltage signals through the strain gauge array and perform analog-to-digital conversion on the raw voltage signals to generate discrete stress data; input the discrete stress data and the number of joint angle pulses collected by the encoder into the kinematic forward solution process to generate real-time spatial pose coordinates; process the real-time spatial pose coordinates, hydraulic pressure sensor readings, motor current sampling values, and control operation instructions to generate a structured device status data matrix; Step 12: Collect original point cloud data in a polar coordinate system using a lidar, and perform coordinate transformation on the original point cloud data to generate a three-dimensional point set in a global coordinate system. Based on the three-dimensional point set, filter out noise points based on the spatial distribution characteristics of adjacent frames to generate de-noised point cloud data. Collect Doppler frequency shift signals using a millimeter-wave radar, perform moving target detection on the frequency shift signals, and calculate obstacle radial velocity vectors. Fusion the de-noised point cloud data, obstacle radial velocity vectors, and the RGB-D depth map collected by the binocular vision sensor generates an environmental perception data packet. Step 13: Time-align the structured device status data matrix with the environment perception data packet, perform interpolation alignment based on the hardware clock synchronization signal, and generate time-synchronized device status data frames and environment perception data frames.

3. The intelligent analysis method for construction equipment operation data according to claim 2, characterized in that: Step 2: Obtain the device's predicted motion trajectory based on the device status data and environmental perception data, including: Step 21: Receive the time-synchronized device status data frame and extract the real-time spatial pose coordinates and control operation instruction codes from it; input the real-time spatial pose coordinates into the device dynamics equation library and generate initial motion trajectory segments based on the rigid body kinematics principle; input the initial motion trajectory segments and the control operation instruction codes into the motion constraint filter and generate physically compliant trajectory segments in combination with the device mechanical structure parameters; Step 22: Input the physically compliant trajectory segment into a pre-trained trajectory feature extractor to decompose it into a three-dimensional feature vector representing the trajectory characteristics. The three-dimensional feature vector is matched with a library of actual trajectory patterns under similar working conditions to generate a set of trajectory offsets. The set of trajectory offsets is input into an adaptive weighting unit, and based on the environmental features contained in the time-synchronized environmental perception data frame, weights are dynamically assigned to generate a comprehensive correction factor. In step 23, the physically compliant trajectory segments and the comprehensive correction factors are input into the trajectory synthesizer, and the pose correction is performed frame by frame through time axis alignment to obtain the corrected trajectory segments. The corrected trajectory segments are then smoothed and processed to eliminate trajectory jitter, generating an optimized continuous motion trajectory as the predicted motion trajectory of the device.

4. The intelligent analysis method for construction equipment operation data according to claim 3, characterized in that: Step 3: Fusion of environmental perception data with the preset building information model and real-time obstacle scanning data to generate a spatial risk field, including: Step 31: Receive the time-synchronized environmental perception data frame and extract the de-noised point cloud data from it; input the de-noised point cloud data into the spatial registration process, align it with the coordinate system of the preset building information model, and generate a registered point cloud; input the registered point cloud into the voxelization processor, segment it into equal-sized 3D grid cells, and mark the static obstacle occupied area; Step 32: Extract the obstacle radial velocity vector and RGB-D depth map from the time-synchronized environmental perception data frame; use the obstacle radial velocity vector, RGB-D depth map, and de-noised point cloud data to generate a future time series position set of the dynamic obstacle; and calculate the spatiotemporal distribution probability of the dynamic obstacle in the three-dimensional grid cells based on the future time series position set to generate a spatiotemporal probability distribution of the dynamic obstacle. Step 33: Generate an initial risk field based on the static obstacle occupancy area markers and the spatiotemporal probability distribution of dynamic obstacles; and perform normalization processing on the initial risk field to generate a spatial risk field.

5. The intelligent analysis method for construction equipment operation data according to claim 4, characterized in that: Step 4: Based on the spatial risk field and the predicted motion trajectory, two calibration reference points with fixed spatial positions are selected within the target analysis domain, and a reference direction axis is generated according to the spatial relationship between the two calibration reference points. Divide the continuous spatial calibration intervals along the reference direction axis; The historical positioning data within the spatial calibration interval is extracted, and based on the trajectory points on the predicted motion trajectory, the position offset vector set between the predicted trajectory points and the actual trajectory points is calculated. The optimization goal is to minimize the weighted square sum of the position offset vectors to obtain a three-dimensional spatial transformation parameter optimization model. According to the optimization model, the quasi-Newton optimization algorithm is used to solve the optimization model and generate the final calibration parameter set, including: Step 41: Based on the three-dimensional spatial distribution characteristics of the spatial risk field, two calibration reference points with fixed spatial positions are selected within the target analysis domain; a reference direction axis is generated based on the spatial coordinates of the two calibration reference points, and direction cosines are calculated; and equal-spaced segmentation is performed along the reference direction axis, with the segmentation points as boundaries to generate continuous spatial calibration intervals. Step 42: extract the historical positioning data of the device in each spatial calibration interval to obtain an actual trajectory point set; extract a predicted trajectory point set that matches the timestamp of the historical positioning data based on the predicted motion trajectory; calculate the spatial offset between the predicted trajectory point and the actual trajectory point point by point to generate a position offset vector set; Step 43: performing normalized weighted processing on the position offset vector set to generate a weighted offset vector set; minimizing the sum of squares of the weighted offset vectors is used as the objective function, and combining the three-dimensional space transformation parameter constraints to obtain a three-dimensional space transformation parameter optimization model; In step 44, the objective function is input into the optimization solver, and the parameter estimates of the rotation matrix, translation vector, and scaling factor are initialized; the quasi-Newton optimization algorithm is used to iteratively execute, the objective function gradient vector under the parameters is calculated, the Hessian inverse matrix approximation is updated, and the parameters are updated with step-size adaptiveness in the opposite direction of the gradient; the iteration is terminated when the change in the objective function is lower than the convergence threshold, and the final calibration parameter set consisting of the rotation matrix, translation vector, and scaling factor is output; the iteration is terminated when the change in the objective function is lower than the convergence threshold to obtain the final calibration parameter set.

6. The intelligent analysis method for construction equipment operation data according to claim 5, characterized in that: Step 5: Based on the final calibration parameter set, the predicted trajectory is corrected. The corrected trajectory is then spatially analyzed with the spatial risk field to calculate the collision risk probability and accident severity level. A comprehensive risk score is generated, and response actions are triggered according to preset thresholds, including: Step 51: Extract the rotation matrix, translation vector, and scaling factor parameters from the final calibration parameter set; input the predicted motion trajectory point set into the space transformer, apply the parameters to perform coordinate transformation calculations, and generate a preliminary correction trajectory; perform time series filtering on the preliminary correction trajectory to eliminate high-frequency jitter noise and generate a final correction trajectory; Step 52: Map the spatial position sequence of the final corrected motion trajectory to the grid coordinate system of the spatial risk field to generate a mapping relationship between trajectory points and grid cells; extract the real-time risk value of the grid cell where each trajectory point is located based on the mapping relationship; Step 53: Calculate a dynamic collision probability value based on the real-time risk value and the geometric distance from the trajectory point to the nearest obstacle; perform an integration operation on the dynamic collision probability value along the trajectory time window to generate an overall collision risk probability value; Step 54 extracts the predefined structural importance coefficient based on the grid cells, and searches the preset accident level comparison table in combination with the overall collision risk probability value to generate the accident severity level; inputs the overall collision risk probability value and the accident severity level into the weighted fusion device to generate a comprehensive risk score; compares the comprehensive risk score with the preset multi-level risk threshold sequence, and outputs a response instruction to trigger the response action.

7. A method for intelligent analysis of construction equipment operation data according to claim 6, characterized in that: Step 4: Extract historical positioning data within the spatial calibration interval and calculate the position offset vector set between the predicted trajectory points and the actual trajectory points based on the trajectory points on the predicted motion trajectory. Minimize the weighted square sum of the position offset vectors as the optimization goal to obtain a three-dimensional spatial transformation parameter optimization model. According to the optimization model, the quasi-Newton optimization algorithm is used to solve the optimization model and generate the final calibration parameter set, including: Step 61: Based on the timestamp alignment results of the actual trajectory point set and the predicted trajectory point set, perform point-by-point spatial position matching; calculate the straight-line distance difference of each matching point pair in the three-dimensional coordinate system and decompose it into X, Y, and Z axial components; combine the axial components to generate a single-point position offset vector, and summarize the position offset vectors of all timestamps to form a position offset vector set; Step 62: assigning increasing weight coefficients to the position offset vector set according to the time decay characteristic; performing a square operation on the weighted position offset vectors and accumulating the sum of the weighted squares to construct a three-dimensional space transformation parameter optimization model with the goal of minimizing the weighted square sum; Step 63, initialize the rotation angle, translation distance and scaling parameters; iterate based on the quasi-Newton optimization algorithm, calculate the objective function gradient vector under the current parameters, update the Hessian inverse matrix approximation, and perform step-size adaptive parameter adjustment along the opposite direction of the gradient; terminate the iteration when the objective function change is lower than the convergence threshold, and output the calibration parameter set consisting of the final rotation angle, translation distance and scaling.

8. A construction equipment operation data intelligent analysis system, the system implementing the method according to any one of claims 1 to 7, characterized in that: include: Acquisition module, used to obtain equipment status data and environmental perception data; The input module is used to obtain the device's predicted motion trajectory based on the device status data and environmental perception data; A fusion module is used to fuse environmental perception data with the preset building information model and real-time obstacle scanning data to generate a spatial risk field; A calculation module is used to select two calibration reference points with fixed spatial positions within the target analysis domain based on the spatial risk field and the predicted motion trajectory, and generate a reference direction axis according to the spatial relationship between the two calibration reference points; Divide the continuous spatial calibration intervals along the reference direction axis; Extract historical positioning data within the spatial calibration interval and calculate the position offset vector set between the predicted and actual trajectory points based on the trajectory points on the predicted motion trajectory. Minimize the weighted sum of the squares of the position offset vectors as the optimization goal to obtain a three-dimensional spatial transformation parameter optimization model. Based on the optimization model, use the quasi-Newton optimization algorithm to solve the optimization model and generate the final calibration parameter set. The output module is used to correct the predicted motion trajectory based on the final calibration parameter set, perform spatial relationship analysis between the corrected trajectory and the spatial risk field, calculate the collision risk probability and accident severity level, generate a comprehensive risk score, and trigger response actions according to preset thresholds.

9. A computing device, characterized in that include: one or more processors; A storage device 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 method according to any one of claims 1 to 7.

Images