A safety production risk assessment method and system based on AI intelligent decision-making

Abstract

This invention relates to the technical field of safety risk assessment, and discloses a safety production risk assessment method and system based on AI intelligent decision-making. The method includes: acquiring global absolute pose, depth matrix, synchronous brightness matrix, radar point cloud, and vehicle velocity vector; determining blinding states based on brightness mean and variance to configure a spatial mask, or extracting gradient edges from the depth matrix to construct a three-dimensional blind zone and mapping it to a spatial mask; converting the depth point cloud corresponding to the depth matrix and the radar point cloud to the chassis coordinate system, and fusing them after time compensation to generate a point cloud to be determined; associating the trajectory to generate a valid trajectory and calculating the absolute velocity vector; if preset filtering conditions are not met, extracting instantaneous spatial features and inputting them into a classifier to output confidence; calculating collision margin when the confidence meets a preset threshold, and performing emergency braking when braking conditions are met. This invention can solve tracking failure caused by blind zone clutter and achieve stable collision avoidance.

Description

Technical Field

[0001] This invention relates to the technical field of safety risk assessment, and in particular to a safety production risk assessment method and system based on AI intelligent decision-making. Background Technology

[0002] In warehousing and production environments with dense metal shelving, mobile platforms are commonly used to transport large non-metallic goods. To ensure operational safety in human-machine collaborative areas, mobile platforms are generally equipped with environmental perception and collision avoidance systems to monitor for intrusion by personnel and the status of obstacles in real time. In such operating environments with narrow aisles, when mobile platforms are transporting large goods and turning or moving straight, the size of the goods can easily obstruct the detection field of the onboard sensors, creating dynamic blind spots around the vehicle. Simultaneously, the densely packed metal shelving and supports have strong reflective properties, which can cause complex interference with radio waves.

[0003] Chinese patent document CN112257522A discloses a multi-sensor fusion environmental perception method based on environmental features. This scheme acquires current ambient light intensity and rain information to switch between lidar, millimeter-wave radar, and visual sensors. When the scheme determines to use millimeter-wave radar and visual sensor fusion, it employs a convolutional neural network algorithm to process the visual image data and a Kalman filter combined with a two-degree-of-freedom vehicle model to process the millimeter-wave radar signal. Finally, a weighted average information fusion algorithm is used to fuse the processed visual image and radar signal to obtain obstacle information.

[0004] However, the aforementioned existing technologies have certain limitations in real-world warehousing and logistics scenarios. First, when processing heterogeneous sensor data fusion, existing solutions, if they do not isolate the physical differences between visual and radar data at the underlying level, forcibly substituting visual point clouds lacking Doppler radial velocity and echo intensity fields into a state matrix containing radar-specific physical features for filtering can easily lead to dimensionality mismatch in the heterogeneous state matrices, causing filter divergence. Second, metal shelves in the environment can generate specular reflections of microwaves, resulting in multipath propagation. Existing multipath filtering mechanisms often only rely on whether the target's motion state is in the same direction as the vehicle to remove clutter. However, when following other vehicles in the aisle, the motion state of the actual worker and the vehicle ahead may coincide with the vehicle's height, easily leading to misjudgment and rejection by the system as clutter, thus causing serious collision hazards. Third, when performing static map filtering and coordinate mapping, existing solutions lack the necessary prerequisite of introducing global absolute positioning, resulting in a discontinuity in the benchmark for mapping local observation coordinates to the global environmental map. Therefore, there is an urgent need to provide a method that can isolate the physical features of heterogeneous point clouds at the underlying level and eliminate radar mirror multipath artifacts by combining global positioning and physical space mapping principles, thereby achieving accurate risk assessment. Summary of the Invention

[0005] To address tracking failures caused by blind zone clutter and achieve stable collision avoidance, this invention provides a safety production risk assessment method and system based on AI intelligent decision-making.

[0006] In a first aspect, the present invention provides a safety production risk assessment method based on AI intelligent decision-making, which adopts the following technical solution: A safety production risk assessment method based on AI intelligent decision-making, comprising:

[0007] The global absolute pose, depth matrix, synchronization brightness matrix, radar point cloud, and vehicle velocity vector of the mobile vehicle at a reference time are obtained, and the mean and variance of the synchronization brightness matrix are calculated.

[0008] If the mean value is greater than the preset extreme brightness threshold and the variance is less than the preset blinding limit, the detection field of view will be configured as a spatial mask, the speed limit will be controlled and near-field braking will be enabled.

[0009] Otherwise, the gradient edges of the depth matrix are extracted to construct a three-dimensional blind zone, which is then mapped to generate a dynamic blind zone and configured as the spatial mask;

[0010] The depth point cloud corresponding to the depth matrix and the radar point cloud are transformed to the chassis coordinate system; the radar point cloud is time-compensated based on the vehicle velocity vector; the depth point cloud that does not fall into the spatial mask and the compensated radar point cloud that falls into the spatial mask are fused with additional data source attributes to generate the undetermined point cloud.

[0011] The undecided point cloud is associated with historical flight tracks to generate a valid flight track. Based on the attributes of the data source, measurement features are extracted to calculate the relative velocity vector, which is then added to the vehicle's velocity vector to generate an absolute velocity vector.

[0012] If the absolute velocity vector does not meet the preset filtering conditions, extract the instantaneous spatial features and input them into the classifier to output the confidence score;

[0013] When the confidence level meets a preset threshold, the collision margin is calculated based on the relative velocity vector, and emergency braking is performed when the braking conditions are met.

[0014] Optionally, extracting the gradient edges of the depth matrix to construct a 3D blind zone includes:

[0015] Calculate the depth difference between adjacent grids in the depth matrix;

[0016] When the absolute value of the depth difference is greater than the size threshold, the corresponding adjacent meshes are extracted as edge points and connected to generate closed gradient edges;

[0017] A closed envelope surface is formed by radiating rays from the projection center toward the gradient edge;

[0018] Starting from the depth section of the gradient edge, extending along the envelope surface to the distance boundary, a closed polyhedral cone is generated as the three-dimensional blind zone.

[0019] Optionally, time compensation is performed on the radar point cloud based on the vehicle velocity vector, including:

[0020] Extract the acquisition time of the radar point cloud and calculate the time difference between the acquisition time and the reference time;

[0021] The offset vector is generated by multiplying the vehicle's speed vector by the time difference.

[0022] Time compensation is achieved by superimposing the offset vector onto the spatial coordinates of the radar point cloud in the chassis coordinate system.

[0023] Optionally, extracting measurement features based on the data source attributes to calculate the relative velocity vector includes:

[0024] When the data source attribute is radar point cloud, the first measurement equation is constructed using radar installation offset, the spatial coordinates of the historical track and radial velocity.

[0025] When the data source attribute is a depth point cloud, a second measurement equation that does not include the radial velocity is constructed using the spatial coordinates of the historical track.

[0026] The relative velocity vector is generated by fusing solutions using a Kalman filter.

[0027] Optionally, the preset filtering conditions include static conditions, and determining whether the static conditions are met includes:

[0028] Calculate the magnitude of the absolute velocity vector, and use the global absolute pose to map the coordinates of the effective track to absolute coordinates;

[0029] When the data source attribute is radar point cloud, calculate the average echo intensity of the point cloud to be decided;

[0030] If the modulus is less than the noise floor threshold, the absolute coordinates match the preset metal spectrum, and the data source attribute is a depth point cloud or the average echo intensity is greater than the reflection threshold, the valid track is discarded.

[0031] Optionally, the preset filtering conditions also include dynamic conditions. When the static conditions are not met, determining that the dynamic conditions are met includes:

[0032] Calculate the angle and magnitude difference between the absolute velocity vector and the vehicle velocity vector;

[0033] When the absolute coordinates fall within the impassable physical boundary corresponding to the preset metal map, and the included corners fall into the same direction interval and the difference in magnitude is less than the difference threshold, it is determined that the dynamic condition is met and the valid track is removed.

[0034] Optionally, extract instantaneous spatial features, including:

[0035] Extract the absolute coordinates of the effective track, and use the global absolute pose transformation to generate a relative position vector relative to the moving vehicle;

[0036] Extract the three-dimensional coordinate set of the point cloud to be decided contained in the effective track, calculate the covariance matrix of the three-dimensional coordinate set and extract the feature values ​​to generate spatial distribution features;

[0037] The relative position vector, the relative velocity vector, and the spatial distribution features are concatenated to reconstruct the instantaneous spatial features.

[0038] Optionally, when the confidence level meets a preset threshold, the collision margin is calculated based on the relative velocity vector, including:

[0039] The classifier, constructed using a multilayer sensing mechanism, maps the instantaneous spatial features to generate a target confidence score, which is then used as the confidence score. The classifier is generated through iterative training using a loss function based on historical instantaneous spatial features and supervision labels calibrated using actual physical collision times.

[0040] A risk signaling is generated when the target confidence level is greater than the probability threshold.

[0041] The collision margin is calculated based on the risk signaling and the relative velocity vector.

[0042] Optionally, calculating the collision margin based on the risk signaling and the relative velocity vector includes:

[0043] In response to the risk signaling, the driving trajectory of the mobile vehicle is extracted;

[0044] Calculate the lateral and longitudinal relative distances of the absolute coordinates corresponding to the effective trajectory relative to the moving vehicle, and calculate the projection component of the relative velocity vector in the direction of the travel trajectory;

[0045] When the lateral relative distance is less than the lateral safety boundary and the projection component is less than zero, the ratio of the absolute value of the longitudinal relative distance to the projection component is taken as the collision margin; otherwise, the collision margin is set as the safety extreme value.

[0046] Secondly, the present invention provides a safety production risk assessment system based on AI intelligent decision-making, which adopts the following technical solution: A safety production risk assessment system based on AI intelligent decision-making, comprising:

[0047] The blind zone construction module acquires the global absolute pose, depth matrix, synchronous brightness matrix, radar point cloud, and vehicle velocity vector of the mobile vehicle at a reference time, and calculates the mean and variance of the synchronous brightness matrix. If the mean is greater than a preset extreme brightness threshold and the variance is less than a preset blinding threshold, the detection field of view is configured as a spatial mask and the speed limit and near-field braking are controlled. Otherwise, the gradient edges of the depth matrix are extracted to construct a three-dimensional blind zone, mapped as a dynamic blind zone, and configured as the spatial mask.

[0048] The point cloud processing module converts the depth point cloud corresponding to the depth matrix and the radar point cloud to the chassis coordinate system; performs time compensation on the radar point cloud based on the vehicle speed vector; fuses the depth point cloud that does not fall into the spatial mask and the compensated radar point cloud that falls into the spatial mask with additional data source attributes to generate a point cloud to be resolved; associates the point cloud to be resolved with historical tracks to generate a valid track; extracts measurement features based on the data source attributes to calculate the relative speed vector, and adds it to the vehicle speed vector to generate an absolute speed vector.

[0049] The intelligent decision-making module extracts instantaneous spatial features and inputs them into a classifier to output confidence scores when the absolute velocity vector does not meet preset filtering conditions; it calculates collision margins when the confidence scores meet preset thresholds.

[0050] The safety execution module controls the mobile carrier to brake urgently when the collision margin meets the braking conditions.

[0051] In summary, the present invention has the following beneficial technical effects:

[0052] 1. By using a spatial masking mechanism to isolate the physical features of depth point clouds and radar point clouds at the bottom layer, the depth point clouds that do not fall into the spatial mask and the radar point clouds that fall into the spatial mask and are time-compensated are fused together with their respective data source attributes to generate the point cloud to be decided. In the subsequent filtering, the corresponding measurement features are extracted based on the data source attribute branches to calculate the velocity vector. This avoids the Kalman filter divergence caused by forcibly substituting heterogeneous sensor data into a unified state matrix due to differences in physical characteristics, and effectively improves the stability of state estimation in the multi-sensor fusion process.

[0053] 2. By using global absolute pose to map the coordinates of effective tracks to absolute coordinates, and combining the hash hit judgment of the preset metal spectrum, effective tracks are destroyed when the absolute velocity vector magnitude is less than the noise floor threshold and the data source attribute is a depth point cloud or the average echo intensity of radar point cloud is greater than the reflection threshold. This allows the mirror multipath reflection clutter of static metal shelves to be accurately identified and eliminated in the global physical space, reducing the probability of static strong reflection structures being misjudged as dynamic obstacles.

[0054] 3. By pre-extracting historical instantaneous spatial features in an offline environment and combining them with supervised labels based on actual physical collision time calibration, the multilayer perceptron classifier is trained under supervision. This enables the classifier to map the target confidence score to instantaneous spatial features with kinematic causality as input. In the online inference stage, it can uniformly assess the obstacle risk represented by different data sources without relying on manually set rules, thus expanding the applicability of the assessment method in different sensing configuration scenarios. Attached Figure Description

[0055] Figure 1 A flowchart of a safety production risk assessment method based on AI intelligent decision-making provided in an embodiment of the present invention;

[0056] Figure 2 An architecture diagram of a safety production risk assessment system based on AI intelligent decision-making provided in an embodiment of the present invention;

[0057] Figure 3 The diagram illustrates the classifier structure and feature transfer principle provided in this embodiment of the invention. Detailed Implementation

[0058] The following combination Figures 1-3 The present invention will be described in further detail below.

[0059] like Figure 1 As shown in the figure, this invention discloses a safety production risk assessment method based on AI intelligent decision-making. The method achieves frame-by-frame continuous operation by executing multi-level conditional branches and looping back to the starting node after a single judgment flow. It is applied to warehousing or production environments with dense metal shelves, metal supports and narrow aisles, and the mobile carrier is used to transport large non-metallic goods.

[0060] Combination Figure 2The mobile vehicle is equipped with depth vision sensing nodes, millimeter-wave Doppler sensing nodes, real-time positioning sensing nodes, chassis-mounted drive-by-wire execution nodes, and a system main control unit. The three sensing nodes are connected in parallel topology, unidirectionally feeding the collected data into the system main control unit. After making a decision, the system main control unit unidirectionally outputs control signals to the chassis-mounted drive-by-wire execution nodes. The real-time positioning sensing node is a LiDAR SLAM estimation module or a UWB ultra-wideband positioning device, used to calculate and output the global absolute pose of the mobile vehicle in a global environment map in real time. The depth vision sensing node is an RGB-D binocular camera, a ToF camera, or a solid-state LiDAR, outputting a depth matrix with a system timestamp and a synchronously acquired brightness matrix. The millimeter-wave Doppler sensing node is a 77GHz millimeter-wave radar or a 4D imaging millimeter-wave radar, outputting a radar point cloud with a system timestamp. The radar point cloud data fields include spatial coordinates, relative distance, azimuth angle, elevation angle, radial velocity, and echo intensity. The chassis drive-by-wire actuators feed back the vehicle's own speed vector via the CAN FD bus or Ethernet TSN bus, and receive control signals for emergency braking, speed limit, and near-field braking modes. The system's main control unit includes a processor, memory, clock synchronization interface, and sensor communication interface. Internally, it stores preset extrinsic parameter matrices, preset radar installation offsets, preset metal spectra, preset classifiers, preset filtering conditions, preset braking conditions, preset extreme brightness thresholds, preset blinding limits, preset metal reflection thresholds, preset near-field braking parameters, and lateral safety boundaries.

[0061] Each hardware node uses IEEE 1588 PTP or equivalent hardware synchronization pulses to unify system time. The output frame rate of the depth vision perception node is 15Hz to 60Hz, and the output frame rate of the millimeter-wave Doppler perception node is 10Hz to 30Hz. The absolute value of the timestamp synchronization tolerance between the radar point cloud and the depth matrix is ​​preferably no greater than 10ms. The vehicle velocity vector fed back by the chassis drive-by-wire execution node includes longitudinal and lateral velocity components in the chassis coordinate system. In differential chassis or forklift chassis, yaw rate is added for short-term motion compensation. The global absolute pose includes absolute translation coordinates and absolute heading angle.

[0062] The system's main control unit reads the depth matrix corresponding to the reference time from the depth vision sensing node, simultaneously reads the global absolute pose at the corresponding time from the instantaneous localization sensing node, and calculates the mean and variance of the brightness matrix acquired synchronously with the depth matrix. Each grid in the depth matrix records the depth value corresponding to the projected pixel, and the brightness matrix records the synchronous exposure brightness.

[0063] In response to situations where strong light can cause blindness, the system's main control unit prioritizes preliminary anomaly detection. A preset extreme brightness threshold is set, for example, 240 for an 8-bit image, and a preset blindness threshold is set, for example, a variance of 100. When the average brightness exceeds the preset extreme brightness threshold and the brightness variance is less than the preset blindness threshold, it is determined that the camera has encountered strong light, causing widespread overexposure and blindness in the field of view. In this case, the vision system cannot effectively extract the true physical gradient edges. Therefore, a fallback logic is triggered: the system's main control unit configures the preset global detection field of view (i.e., the radar hardware's native maximum detection field of view boundary) as a spatial mask, activates the global microwave sensing mode, and controls the moving vehicle to limit its speed and activate the near-field braking mode.

[0064] When the conditions of mean brightness exceeding a preset extreme brightness threshold and brightness variance falling below a preset blinding threshold are not met, the system's main control unit calculates the depth difference between adjacent grids using the depth matrix as input, for example, calculating the depth difference along the horizontal or vertical direction. Preset jump conditions include the absolute value of the depth difference exceeding a preset size threshold. The preset size threshold is set based on the minimum occlusion thickness of the edge of large non-metallic goods, preferably 0.35m. Adjacent grids satisfying the preset jump conditions are extracted as edge points, and all edge points are connected to generate a closed gradient edge. A closed gradient edge represents a complete depth abrupt change contour formed between the foreground goods boundary and its background.

[0065] The system's main control unit utilizes a vision node with a projection center to emit multiple rays from the projection center towards the gradient edge, forming a closed envelope surface. Starting from the depth section where the gradient edge is located, the closed envelope surface extends away from the vision node to a predetermined distance boundary, generating a multifaceted cone with an internal closed volume as a 3D blind zone. This physical deduction process strictly limits the geometric starting point of the 3D blind zone to behind the obstacle, ensuring that the non-metallic obstacle body that causes the depth jump and its unobstructed area directly in front are not included in the spatial mask. This geometric deduction, which excludes the obstacle body from the mask and only defines the physical shadow area caused by occlusion, can completely preserve the high-precision visual depth point cloud of the foreground obstacle during subsequent point cloud clipping, effectively avoiding self-blinding of the obstacle by the system due to mask miscutting. The distance boundary is set according to the channel width and the maximum braking distance of the moving vehicle, preferably 3m to 12m. The preset extrinsic parameter matrix obtained from the factory joint calibration is called to map the 3D blind zone to the chassis coordinate system to generate a dynamic blind zone, and the dynamic blind zone is configured as a spatial mask as the point cloud clipping boundary. The spatial mask's data structure is a three-dimensional set of voxels, recording all spatial raster indices falling within the dynamic blind zone.

[0066] The millimeter-wave Doppler sensing node outputs radar point cloud data to the system's main control unit, while the chassis drive-by-wire execution node feeds back the vehicle's velocity vector. The main control unit calls a preset extrinsic parameter matrix to uniformly transform the depth point cloud contained in the depth matrix and the radar point cloud to the chassis coordinate system. Subsequently, the main control unit extracts the acquisition time of the radar point cloud, calculates the time difference between the acquisition time and the reference time, calculates the product of the vehicle's velocity vector and the time difference to generate an offset vector, and directly superimposes the offset vector onto the radar point cloud spatial coordinates transformed to the chassis coordinate system to complete time compensation. This step is based on the physical laws of rigid body space transformation, avoiding the incorrect substitution of translational offsets from the chassis coordinate system into the radar's local coordinate system, which has angular extrinsic parameters. Then, heterogeneous data region fusion logic is executed in the chassis coordinate system. Spatial partitioning is performed based on a spatial mask: In the unobstructed area outside the spatial mask, depth point clouds whose spatial coordinates do not fall within the spatial mask are extracted and labeled with data source attributes indicating depth point clouds; in the area within the spatial mask, radar point clouds whose spatial coordinates fall within the spatial mask (the determination algorithm is based on whether the 3D coordinates of the point cloud are contained within the index range of the 3D voxel set) and have undergone time compensation are extracted and labeled with data source attributes indicating radar point clouds. The system merges the extracted depth point clouds and the time-compensated radar point clouds to generate a unified undecided point cloud carrying the corresponding data source attributes. This logic provides a benchmark for subsequent branch solutions through physical spatial alignment and labeling of hardware physical features.

[0067] The system's main control unit performs Euclidean clustering on the point cloud to be resolved and associates it with historical tracks to generate valid tracks. During Kalman tracking calculations, the system updates the state using the rate of change of the chassis coordinate system's three-dimensional spatial coordinates over time, which is contained in the historical tracks. Subsequently, based on the data source attributes, it extracts the measurement features corresponding to the point cloud to be resolved and performs branch measurement updates: when the data source attributes indicate radar point cloud, it extracts the radial velocity contained in the measurement features and constructs a first measurement equation containing the depth dimension using the preset radar installation offset, three-dimensional spatial coordinates, and radial velocity.

[0068]

[0069] in These are the three-dimensional components of the spatial coordinates in the chassis coordinate system. The three-dimensional coordinate components of the preset radar installation offset. , , This represents the three-dimensional components of the relative velocity vector. This equation compensates for the translational extrinsic parameter error between the radar phase center and the chassis origin, avoiding the divergence of the extended Kalman filter covariance matrix caused by angular misalignment during close-range observation. When the data source attribute indicates a depth point cloud, due to the objective lack of Doppler frequency shift features in visual characteristics, the Doppler channel is adaptively skipped, and a second measurement equation that does not include radial velocity is constructed using three-dimensional spatial coordinates: ,in The measurement vector represents the visual target, and the corresponding measurement matrix is ​​a third-order identity matrix. This isolation mechanism avoids matrix calculation anomalies caused by extracting Doppler features from the visual point cloud. After calculating the relative velocity vector in the chassis coordinate system, it is vector-superimposed with the vehicle's velocity vector to generate the target's absolute velocity vector in the global environment map.

[0070] The preset filtering conditions include static and dynamic conditions. The system's main control unit calculates the magnitude of the absolute velocity vector and, using the global absolute pose combined with the rigid body homogeneous transformation matrix, maps the current local coordinates of the chassis of the valid track to the absolute coordinates of the environmental map. The preset metal map is a set of absolute coordinates of the static metal structure in the environment, pre-calibrated and stored using a voxel hash table. When the data source attribute indicates a radar point cloud, the average echo intensity of the undecided point cloud associated with the valid track is calculated; this calculation is skipped for depth point clouds. When the magnitude is less than the preset noise floor threshold, preferably between 0.05 m / s and 0.20 m / s, the absolute coordinates successfully match the hash of the preset metal map, and the data source attribute indicates a depth point cloud, or the data source attribute indicates a radar point cloud and the average echo intensity is greater than the preset metal reflection threshold, the static condition is met, and the valid track is removed. This step improves the mathematical derivation basis for coordinate system mapping.

[0071] When static conditions are not met, multipath dynamic condition judgment is performed. When radar microwaves encounter the side surface of a smooth metal shelf and generate a dynamic illusion of vehicle movement, based on the principle of optical and microwave virtual mirroring, the spatial absolute coordinates calculated by the multipath virtual image will be projected inside the impassable physical boundary behind the metal reflector. The system calculates the angle and magnitude difference between the absolute velocity vector and the vehicle's velocity vector to determine whether the absolute coordinates fall inside the impassable physical boundary corresponding to the preset metal pattern. When the absolute coordinates fall inside the impassable physical boundary, and the angle falls within the same direction range (preferably -5 degrees to +5 degrees) and the magnitude difference is less than the difference threshold (preferably 0.15 m / s), it is determined that the internal mirror clutter meets the dynamic conditions, and the valid track is discarded. If the absolute coordinates fall within the legal passage, the system retains the real obstacle ahead, avoiding the problem of falsely discarding real targets caused by filtering only based on same-direction velocity.

[0072] When the absolute velocity vector does not meet the preset filtering conditions, the system extracts the absolute coordinates of the effective track in the current frame and uses the global absolute pose of the current frame combined with the rigid body homogeneous transformation matrix to inversely transform it to the chassis coordinate system to generate a relative position vector relative to the moving vehicle. Simultaneously, the system extracts the three-dimensional coordinate set of the undetermined point cloud contained in the effective track in the current frame, calculates the covariance matrix of this three-dimensional coordinate set, and extracts eigenvalues ​​to generate spatial distribution features; these spatial distribution features reflect the discrete distribution scale of the target's length, width, and height geometric contours.

[0073] Furthermore, such as Figure 3 As shown, the system uses a relative position vector containing the target's relative distance, a relative velocity vector containing the target's approaching situation, and spatial distribution features reflecting the target's geometric size as three independent parallel inputs. These are then concatenated into instantaneous spatial features, which are subsequently input into a classifier with a multi-level fully connected hidden layer topology for forward propagation. Finally, the inputs converge and a single-node output layer generates the target confidence score. This mechanism incorporates the relative spatial position and relative approaching velocity between the vehicle and the target, conforming to kinematic principles and providing complete input conditions with a physical basis for subsequent collision risk prediction.

[0074] To overcome the problem of not being able to directly call open-source pre-trained models for non-standard heterogeneous data and to ensure the feasibility of the decision-making algorithm, the classifier deployed inside the system's main control unit undergoes white-box supervised training in an offline environment before the system is put into operation. The specific offline training closed-loop mechanism is as follows:

[0075] First, a heterogeneous training sample set is constructed. A large number of historical valid tracks collected by the mobile vehicle under similar historical operating conditions are acquired. Following rules strictly consistent with online inference, historical absolute coordinates are extracted and combined with historical global absolute pose to generate historical relative position vectors through inverse transformation. The 3D coordinate set of historical valid tracks is extracted, and the eigenvalues ​​of the covariance matrix are calculated to generate historical spatial distribution features. The historical relative position vector, historical relative velocity vector, and historical spatial distribution features are concatenated to reconstruct historical instantaneous spatial features, which serve as input samples for the network's forward propagation.

[0076] Secondly, objective physical calibration of the supervision labels is performed. A global high-precision trajectory reconstructed by offline high-precision LiDAR is introduced as the ground truth benchmark to calculate the actual physical collision time of the target relative to the moving carrier in the input sample. Rigid physical labeling rules are established: when the actual physical collision time is less than a preset safety lower limit time, or when manual intervention is recorded in the system's historical logs, the supervision label of the corresponding input sample is labeled with a value of one, representing a high-risk collision state; when the actual physical collision time is greater than or equal to the preset safety lower limit time and no physical intervention occurs, the supervision label is labeled with a value of zero, representing a safe state. This rule solves the technical problem of the inability to manually label non-standard spatial heterogeneous features.

[0077] Finally, network iteration and parameter fixing are performed. The preset classifier adopts a multilayer perceptron architecture, whose network topology includes an input layer, hidden layers using a linear rectified activation function, and an output layer using a logistic activation function. The input samples are forward-propagated to output the predicted target confidence, and the error loss between the predicted target confidence and the corresponding supervision label is calculated using a binary cross-entropy loss function. An adaptive moment estimation optimization algorithm is used to perform backpropagation along the gradient descent direction, iteratively updating the network weights and bias parameters of the hidden and output layers until the loss function reaches its convergence limit. The network weights are then fixed to generate the classifier.

[0078] During the online risk assessment phase, the system's main control unit inputs the instantaneous spatial features generated in the current frame into a classifier that has completed offline training for network mapping, and outputs the current target confidence score. When the target confidence score is greater than a preset probability threshold, a risk signal is generated.

[0079] In response to risk signals, the system extracts the preset travel trajectory of the mobile vehicle. It calculates the lateral relative distance of the absolute coordinates of the effective trajectory relative to the mobile vehicle in the lateral direction, and the longitudinal relative distance in the travel trajectory direction. It also calculates the longitudinal velocity projection component of the relative velocity vector in the travel trajectory direction. To avoid ghost braking caused by absolutely safe stationary obstacles in the lateral direction within the passage, the system introduces a lateral safety boundary determination. The lateral safety boundary is set based on the physical width of the mobile vehicle plus a buffer tolerance, preferably 0.8 meters. When the lateral relative distance is less than the preset lateral safety boundary and the projection component is less than zero, it indicates that the target is within the collision envelope directly in front and is physically approaching. In this case, the ratio of the longitudinal relative distance to the absolute value of the projection component is calculated to generate the actual collision margin. When the lateral relative distance is greater than or equal to the lateral safety boundary, or the projection component is greater than or equal to zero, it indicates that the target is in the lateral safety zone or far from the mobile vehicle, and the collision margin is set to a preset safety limit. When the collision margin meets the braking conditions, such as less than 1.5 seconds, the chassis drive-by-wire execution node controls emergency braking.

[0080] This invention also discloses a safety production risk assessment system based on AI intelligent decision-making, including a blind spot construction module, a point cloud processing module, an intelligent decision-making module, and a safety execution module.

[0081] The blind zone construction module is used to obtain the global absolute pose, depth matrix, synchronous brightness matrix, radar point cloud, and vehicle velocity vector of the mobile vehicle at the reference time, and calculate the mean brightness and variance of the synchronous brightness matrix. When the mean brightness is greater than the preset extreme brightness threshold and the variance of brightness is less than the preset blinding limit, the detection field of view is configured as a spatial mask to control speed limit and enable the near-field braking mode. If the conditions are not met, the gradient edges in the depth matrix that meet the preset jump conditions are extracted to construct a three-dimensional blind zone. The preset extrinsic parameter matrix is ​​called to map the three-dimensional blind zone to the chassis coordinate system to generate a dynamic blind zone, and the dynamic blind zone is configured as a spatial mask.

[0082] The point cloud processing module is used to transform the depth point cloud and radar point cloud contained in the depth matrix to the chassis coordinate system using the extrinsic parameter matrix; to perform time compensation on the radar point cloud transformed to the chassis coordinate system based on the vehicle's velocity vector to align it to the reference time; to extract the depth point cloud whose spatial coordinates do not fall into the spatial mask, and to fuse it with the time-compensated radar point cloud whose spatial coordinates fall into the spatial mask, after adding the corresponding data source attributes, to generate the point cloud to be decided; to associate with historical tracks to generate valid tracks, to extract measurement features based on data source attributes to calculate the relative velocity vector, and to add it to the vehicle's velocity vector to generate the absolute velocity vector.

[0083] When the absolute velocity vector does not meet the preset filtering conditions, the intelligent decision-making module extracts instantaneous spatial features and inputs them into the classifier to output confidence; when the confidence meets the preset threshold, it calculates the collision margin.

[0084] When the collision margin meets the braking conditions, the safety execution module controls the moving vehicle to perform emergency braking. Each hardware and software coordination step is implemented by the built-in processor executing relevant program instructions.

[0085] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A safety production risk assessment method based on AI intelligent decision-making, characterized in that, include: The global absolute pose, depth matrix, synchronization brightness matrix, radar point cloud, and vehicle velocity vector of the mobile vehicle at a reference time are obtained, and the mean and variance of the synchronization brightness matrix are calculated. If the mean value is greater than the preset extreme brightness threshold and the variance is less than the preset blinding limit, the detection field of view will be configured as a spatial mask, the speed limit will be controlled and near-field braking will be enabled. Otherwise, the gradient edges of the depth matrix are extracted to construct a three-dimensional blind zone, which is then mapped to generate a dynamic blind zone and configured as the spatial mask; The depth point cloud corresponding to the depth matrix and the radar point cloud are transformed to the chassis coordinate system; the radar point cloud is time-compensated based on the vehicle velocity vector; the depth point cloud that does not fall into the spatial mask and the compensated radar point cloud that falls into the spatial mask are fused with additional data source attributes to generate the undetermined point cloud. The undecided point cloud is associated with historical flight tracks to generate a valid flight track. Based on the attributes of the data source, measurement features are extracted to calculate the relative velocity vector, which is then added to the vehicle's velocity vector to generate an absolute velocity vector. If the absolute velocity vector does not meet the preset filtering conditions, extract the instantaneous spatial features and input them into the classifier to output the confidence score; When the confidence level meets a preset threshold, the collision margin is calculated based on the relative velocity vector, and emergency braking is performed when the braking conditions are met.

2. The method according to claim 1, characterized in that, Extracting the gradient edges of the depth matrix to construct a 3D blind zone includes: Calculate the depth difference between adjacent grids in the depth matrix; When the absolute value of the depth difference is greater than the size threshold, the corresponding adjacent meshes are extracted as edge points and connected to generate closed gradient edges; A closed envelope surface is formed by radiating rays from the projection center toward the gradient edge; Starting from the depth section of the gradient edge, extending along the envelope surface to the distance boundary, a closed polyhedral cone is generated as the three-dimensional blind zone.

3. The method according to claim 1, characterized in that, Time compensation of the radar point cloud based on the vehicle velocity vector includes: Extract the acquisition time of the radar point cloud and calculate the time difference between the acquisition time and the reference time; The offset vector is generated by multiplying the vehicle's speed vector by the time difference. Time compensation is achieved by superimposing the offset vector onto the spatial coordinates of the radar point cloud in the chassis coordinate system.

4. The method according to claim 1, characterized in that, Based on the attributes of the data source, the measurement features are extracted and the relative velocity vector is calculated, including: When the data source attribute is radar point cloud, the first measurement equation is constructed using radar installation offset, the spatial coordinates of the historical track and radial velocity. When the data source attribute is a depth point cloud, a second measurement equation that does not include the radial velocity is constructed using the spatial coordinates of the historical track. The relative velocity vector is generated by fusing solutions using a Kalman filter.

5. The method according to claim 1, characterized in that, The preset filtering conditions include static conditions, and determining whether the static conditions are met includes: Calculate the magnitude of the absolute velocity vector, and use the global absolute pose to map the coordinates of the effective track to absolute coordinates; When the data source attribute is radar point cloud, calculate the average echo intensity of the point cloud to be decided; If the modulus is less than the noise floor threshold, the absolute coordinates match the preset metal spectrum, and the data source attribute is a depth point cloud or the average echo intensity is greater than the reflection threshold, the valid track is discarded.

6. The method according to claim 5, characterized in that, The preset filtering conditions also include dynamic conditions. When the static conditions are not met, determining that the dynamic conditions are met includes: Calculate the angle and magnitude difference between the absolute velocity vector and the vehicle velocity vector; When the absolute coordinates fall within the impassable physical boundary corresponding to the preset metal map, and the included corners fall into the same direction interval and the difference in magnitude is less than the difference threshold, it is determined that the dynamic condition is met and the valid track is removed.

7. The method according to claim 1, characterized in that, Extracting instantaneous spatial features, including: Extract the absolute coordinates of the effective track, and use the global absolute pose transformation to generate a relative position vector relative to the moving vehicle; Extract the three-dimensional coordinate set of the point cloud to be decided contained in the effective track, calculate the covariance matrix of the three-dimensional coordinate set and extract the feature values ​​to generate spatial distribution features; The relative position vector, the relative velocity vector, and the spatial distribution features are concatenated to reconstruct the instantaneous spatial features.

8. The method according to claim 1, characterized in that, When the confidence level meets a preset threshold, the collision margin is calculated based on the relative velocity vector, including: The classifier, constructed using a multilayer sensing mechanism, maps the instantaneous spatial features to generate a target confidence score, which is then used as the confidence score. The classifier is generated through iterative training using a loss function based on historical instantaneous spatial features and supervision labels calibrated using actual physical collision times. A risk signaling is generated when the target confidence level is greater than the probability threshold. The collision margin is calculated based on the risk signaling and the relative velocity vector.

9. The method according to claim 8, characterized in that, Calculating the collision margin based on the risk signaling and the relative velocity vector includes: In response to the risk signaling, the driving trajectory of the mobile vehicle is extracted; Calculate the lateral and longitudinal relative distances of the absolute coordinates corresponding to the effective trajectory relative to the moving vehicle, and calculate the projection component of the relative velocity vector in the direction of the travel trajectory; When the lateral relative distance is less than the lateral safety boundary and the projection component is less than zero, the ratio of the absolute value of the longitudinal relative distance to the projection component is taken as the collision margin; otherwise, the collision margin is set as the safety extreme value.

10. A safety production risk assessment system based on AI intelligent decision-making, used to implement the method described in any one of claims 1-9, characterized in that, include: The blind zone construction module acquires the global absolute pose, depth matrix, synchronous brightness matrix, radar point cloud, and vehicle velocity vector of the mobile vehicle at a reference time, and calculates the mean and variance of the synchronous brightness matrix. If the mean is greater than a preset extreme brightness threshold and the variance is less than a preset blinding threshold, the detection field of view is configured as a spatial mask and the speed limit and near-field braking are controlled. Otherwise, the gradient edges of the depth matrix are extracted to construct a three-dimensional blind zone, mapped as a dynamic blind zone, and configured as the spatial mask. The point cloud processing module converts the depth point cloud corresponding to the depth matrix and the radar point cloud to the chassis coordinate system; performs time compensation on the radar point cloud based on the vehicle speed vector; fuses the depth point cloud that does not fall into the spatial mask and the compensated radar point cloud that falls into the spatial mask with additional data source attributes to generate a point cloud to be resolved; associates the point cloud to be resolved with historical tracks to generate a valid track; extracts measurement features based on the data source attributes to calculate the relative speed vector, and adds it to the vehicle speed vector to generate an absolute speed vector. The intelligent decision-making module extracts instantaneous spatial features and inputs them into a classifier to output confidence scores when the absolute velocity vector does not meet preset filtering conditions; it calculates collision margins when the confidence scores meet preset thresholds. The safety execution module controls the mobile carrier to brake urgently when the collision margin meets the braking conditions.

Images