Low-speed unmanned vehicle operation status data recording and analysis system

Abstract

This invention provides a system for recording and analyzing low-speed unmanned vehicle operating status data, relating to the field of data processing technology. The system includes: analyzing a multi-dimensional state matrix based on a decision geometry structure to generate an event trigger sequence; extracting data slices within the event time neighborhood from the multi-dimensional state matrix based on the time reference in the event trigger sequence to construct a spatiotemporal data cube; extracting features from the spatiotemporal data cube to obtain event feature fingerprints and generate event feature codes; inputting the event feature codes into a pre-trained deep belief network to obtain event risk level assessment results; and fusing the event trigger sequence, spatiotemporal data cube, event feature codes, and event risk level assessment results to generate a structured event analysis report. This invention can improve the accuracy of abnormal event detection, the reliability of risk assessment, and the efficiency of analysis in the operating status of low-speed unmanned vehicles.

Description

Technical Field

[0001] This invention relates to the field of data processing technology, and in particular to a system for recording and analyzing the operating status data of low-speed unmanned vehicles. Background Technology

[0002] In the actual operation of unmanned logistics vehicles in the park, for example, when they drive to a road intersection and need to turn left or right, the vehicles generate multi-dimensional spatiotemporal data, including GPS heading angle changes, IMU yaw rate, wheel speed difference, and visual sensor data. Currently, the analysis of this type of operation usually relies on setting fixed thresholds for a single data dimension for judgment. For example, abnormal steering events may only be triggered when the yaw rate exceeds a certain set value. However, in reality, the normal agile steering of the vehicle at the intersection and the abnormal yaw caused by slippery ground or instantaneous sensor interference may show similar numerical characteristics in a single dimension (such as yaw rate). This judgment method based on a single or a few isolated thresholds is often difficult to effectively distinguish and quantify such complex behavioral patterns that are tightly coupled in spatiotemporal space and dynamically evolve in a multi-dimensional feature space, which may lead to misjudgment or omission of potential risks. Summary of the Invention

[0003] The technical problem to be solved by the present invention is to provide a system for recording and analyzing the operating status data of low-speed unmanned vehicles, which can improve the accuracy of abnormal event detection, the reliability of risk assessment and the efficiency of analysis in the operating status of low-speed unmanned vehicles.

[0004] To solve the above-mentioned technical problems, the technical solution of the present invention is as follows:

[0005] The first aspect is a low-speed unmanned vehicle operation status data recording and analysis system, including:

[0006] The acquisition module is used to collect multi-dimensional spatiotemporal data streams of vehicle operating status;

[0007] The fusion module is used to fuse multidimensional spatiotemporal data streams, generate a multidimensional state matrix, and convert it into discrete feature topology through mapping;

[0008] The partitioning module is used to construct a dynamic reference manifold based on discrete feature topology, and determine the analysis boundary on the dynamic reference manifold. According to the geodesic distribution of the discrete feature topology, the analysis boundary is partitioned, the vertex sequence of the analysis boundary is obtained, the adjacent edge vectors are calculated based on the vertex sequence, the concavity and convexity of the vertices are determined according to the cross product sign of the adjacent edge vectors, and the analysis boundary is divided into convex segments and concave segments based on the concavity and convexity of the vertices, generating topological analysis cells.

[0009] The computation module is used to map discrete feature topology to corresponding topology analysis cells and calculate manifold geometric correction quantities; based on the manifold geometric correction quantities, a decision geometry structure is constructed.

[0010] The parsing module is used to parse the multidimensional state matrix based on the decision geometry structure and generate an event trigger sequence; based on the time reference in the event trigger sequence, it extracts data slices in the event time neighborhood from the multidimensional state matrix and constructs a spatiotemporal data cube.

[0011] The extraction module is used to extract features from the spatiotemporal data cube, obtain event feature fingerprints, and generate event feature codes.

[0012] The training module is used to input event feature encodings into a pre-trained deep belief network to obtain event risk level assessment results;

[0013] The generation module is used to integrate event trigger sequences, spatiotemporal data cubes, event feature codes, and event risk level assessment results to generate a structured event analysis report.

[0014] In a second aspect, a computing device includes:

[0015] One or more processors;

[0016] A storage device for storing one or more programs that, when executed by one or more processors, enable the one or more processors to implement the system.

[0017] Thirdly, a computer-readable storage medium storing a program that, when executed by a processor, implements the system.

[0018] The above-described solution of the present invention has at least the following beneficial effects:

[0019] By constructing dynamic reference manifolds and topological analysis cells based on geodesics and concavity / convexity partitioning, the normal operating space of vehicles can be modeled in detail, thereby enabling sensitive and accurate identification of subtle abnormal patterns and achieving earlier risk warnings. Not only does it output reliable risk levels through deep belief networks, but it also preserves the complete analysis chain from raw data and spatiotemporal cubes to feature encoding, making the risk assessment results traceable and enhancing the system's credibility and debugging convenience. Attached Figure Description

[0020] Figure 1 This is a schematic diagram of a low-speed unmanned vehicle operation status data recording and analysis system provided in an embodiment of the present invention.

[0021] Figure 2 This is a schematic diagram of the process of inputting event feature encoding into a pre-trained deep belief network to obtain the event risk level assessment result, provided by an embodiment of the present invention. Detailed Implementation

[0022] Exemplary embodiments of the present disclosure will now be described in more detail with reference to the accompanying drawings. While exemplary embodiments of the present disclosure are shown in the drawings, it should be understood that the present disclosure may be implemented in various forms and should not be limited to the embodiments set forth herein. Rather, these embodiments are provided so that this disclosure will be thorough and complete, and will fully convey the scope of the disclosure to those skilled in the art.

[0023] like Figure 1 As shown, an embodiment of the present invention proposes a low-speed unmanned vehicle operating status data recording and analysis system, including:

[0024] The acquisition module is used to collect multi-dimensional spatiotemporal data streams of vehicle operating status;

[0025] The fusion module is used to fuse multidimensional spatiotemporal data streams, generate a multidimensional state matrix, and convert it into discrete feature topology through mapping;

[0026] The partitioning module is used to construct a dynamic reference manifold based on discrete feature topology, and determine the analysis boundary on the dynamic reference manifold. According to the geodesic distribution of the discrete feature topology, the analysis boundary is partitioned, the vertex sequence of the analysis boundary is obtained, the adjacent edge vectors are calculated based on the vertex sequence, the concavity and convexity of the vertices are determined according to the cross product sign of the adjacent edge vectors, and the analysis boundary is divided into convex segments and concave segments based on the concavity and convexity of the vertices, generating topological analysis cells.

[0027] The computation module is used to map discrete feature topology to corresponding topology analysis cells and calculate manifold geometric correction quantities; based on the manifold geometric correction quantities, a decision geometry structure is constructed.

[0028] The parsing module is used to parse the multidimensional state matrix based on the decision geometry structure and generate an event trigger sequence; based on the time reference in the event trigger sequence, it extracts data slices in the event time neighborhood from the multidimensional state matrix and constructs a spatiotemporal data cube.

[0029] The extraction module is used to extract features from the spatiotemporal data cube, obtain event feature fingerprints, and generate event feature codes.

[0030] The training module is used to input event feature encodings into a pre-trained deep belief network to obtain event risk level assessment results;

[0031] The generation module is used to integrate event trigger sequences, spatiotemporal data cubes, event feature codes, and event risk level assessment results to generate a structured event analysis report.

[0032] In this embodiment of the invention, by constructing a dynamic reference manifold and topological analysis cells based on geodesics and concavity / convexity partitioning, the normal operating space of a vehicle can be modeled precisely, thereby enabling sensitive and accurate identification of subtle abnormal patterns and achieving earlier risk warnings. Not only does it output reliable risk levels through deep belief networks, but it also preserves the complete analysis chain from raw data and spatiotemporal cubes to feature encoding, making the risk assessment results traceable and enhancing the system's credibility and debugging convenience.

[0033] In a preferred embodiment of the present invention, the acquisition of multi-dimensional spatiotemporal data streams of vehicle operating status specifically includes: deploying multiple sensors at key locations of the low-speed unmanned vehicle, and using these sensors to collaboratively acquire multi-dimensional spatiotemporal data streams of vehicle operating status; fixing a GPS positioning module at the center of the vehicle's top to acquire the vehicle's longitude, latitude, altitude, and heading angle in real time, acquiring data ten times per second to ensure that the deviation between the positioning result and the actual position does not exceed 0.5 meters; and installing an inertial measurement unit at the center of gravity of the vehicle chassis to acquire the vehicle's acceleration and angular velocity one hundred times per second, with the acceleration measurement range controlled within ±10 meters per second squared and the angular velocity measurement range controlled within ±500 degrees per second. Wheel speed sensors are installed on the hubs of all four wheels, collecting the rotational speed of each wheel fifty times per second. The wheel speed difference is calculated by comparing the rotational speeds of the four wheels, and then combined with the wheel radius to calculate the vehicle speed, ensuring that the speed measurement deviation does not exceed 0.1 km / h. Visual cameras are fixed on the inside of the front and rear windshields, with the front camera facing the direction of travel and the rear camera facing the opposite direction, collecting thirty frames of road environment images per second, with an image resolution set to 1920 x 1080 pixels. A lidar is installed on the front of the vehicle's roof, collecting distance and orientation data of obstacles around the vehicle ten times per second, with a detection range covering 0.1 meters to 100 meters and an angular resolution controlled within 0.1 degrees. All data collected by the sensors are transmitted in real time to the onboard data processing terminal, forming a multi-dimensional spatiotemporal data stream.

[0034] This embodiment, by deploying multiple sensors at key locations and specifying the parameters to be collected, can capture key information such as the vehicle's position, attitude, speed, and environment.

[0035] In a preferred embodiment of the present invention, fusing multidimensional spatiotemporal data streams to generate a multidimensional state matrix, and then converting it into a discrete feature topology through mapping, may include:

[0036] The process involves performing timestamp alignment and spatial coordinate system transformation on asynchronous sensor data from a multidimensional spatiotemporal data stream to generate a fused data stream. Based on this fused data stream, a sliding window of preset time length is used to continuously extract data, and the time-series data within each window is arranged and combined according to the state dimension to generate a structured multidimensional state matrix. Specifically, this includes: first, timestamp alignment of the multidimensional spatiotemporal data stream, using the timestamp of the GPS positioning module output data as a reference, generated at a frequency of ten times per second (i.e., each timestamp is 0.1 seconds apart); for asynchronous data output from other sensors such as the inertial measurement unit and wheel speed sensor, linear interpolation is used for completion. Specifically, the target timestamp of the data to be completed is first selected, and then the nearest valid data point before and after the target timestamp is found in the historical data of that sensor. Assuming the timestamp of the previous valid data point is... The value is The timestamp of the subsequent valid data points is The value is Given a target timestamp of t, first calculate the numerical difference. minus The time difference is minus Divide the numerical difference by the time difference to obtain the change per unit time, and then subtract the timestamps of the previous valid data points from the target timestamp t. The time interval is obtained, multiplied by the change per unit time, and then the value of the previous valid data point is added. This process yields the padded value corresponding to the target timestamp; all sensor data is processed in this way, ultimately unifying the timestamp interval of all sensor data to 0.1 seconds.

[0037] After timestamp alignment, a spatial coordinate system transformation is performed, converting the image coordinate system data acquired by the visual camera and the polar coordinate system data acquired by the LiDAR into the vehicle's body coordinate system. During the transformation, the visual camera's image coordinate system uses the lens optical center as its origin. First, the target's coordinates in the image coordinate system are obtained by converting pixel coordinates to physical coordinates. Then, a rotation matrix with zero degrees rotation around the X, Y, and Z axes is used to eliminate angular deviations. Next, translation vectors of 0 meters on the X-axis, 0 meters on the Y-axis, and 0.8 meters on the Z-axis are superimposed to obtain the coordinates in the vehicle's body coordinate system. The LiDAR's polar coordinate system uses the radar center as its origin. First, the polar coordinate distance and angle are converted to rectangular coordinate values, and then the same rotation and translation operations are performed to unify the data from different coordinate systems and generate fused data. After the spatial coordinate system transformation is completed, a sliding window is used to capture the fused data stream. The window time length is set to five seconds and the step size is one second. That is, the first window covers the first five seconds of data from the initial moment, the second window covers the data from the first second to the sixth second, and so on. Each captured window contains complete data for fifty time points. The time series data in each window are arranged according to the state dimension, which includes longitude, latitude, altitude, driving speed, acceleration, heading angle, yaw rate, wheel speed difference, obstacle distance, and obstacle orientation, for a total of ten dimensions. When arranging, each dimension is a row, and the data of the fifty time points under that dimension are arranged in chronological order as the fifty values ​​of that row. The ten dimensions form ten rows, and finally, they are arranged into a multi-dimensional state matrix of ten rows and fifty columns.

[0038] Based on the structured multidimensional state matrix, the state vector corresponding to each time step in the multidimensional state matrix is ​​mapped to a preset multidimensional feature space. The coordinate position of each state vector in the feature space is calculated to generate a feature space trajectory. Based on the feature space trajectory, the feature distance between adjacent coordinate points in the feature space trajectory is calculated. According to a preset distance threshold, dense coordinate points with feature distances less than the threshold are aggregated into a feature node. All generated feature nodes are connected in chronological order to generate a discrete feature topology. Specifically, this includes: extracting the ten-dimensional state vector corresponding to each time step from the multidimensional state matrix, with each time step corresponding to a column of data in the matrix, and using this column of data as the ten-dimensional state vector for that time step; then mapping these state vectors to a preset eight-dimensional feature space, where the dimensions of the eight-dimensional feature space correspond to the dimensions of the original state, with velocity features corresponding to rows. The system considers the following parameters: driving speed and wheel speed difference; attitude features corresponding to heading angle and yaw rate; position change features corresponding to changes in longitude, latitude, and altitude; obstacle perception features corresponding to obstacle distance and orientation; motion stability features corresponding to acceleration fluctuations; steering features corresponding to the rate of change of heading angle and yaw rate; acceleration features corresponding to raw acceleration data; and environment matching features corresponding to the results of simple processing of environmental data collected by visual cameras and LiDAR. For each state vector, the corresponding raw dimension data is extracted and normalized. Taking the driving speed dimension in the speed features as an example, the minimum and maximum values ​​of this dimension are first calculated across all time points. Then, the minimum value is subtracted from the driving speed data at that time point, and the difference is divided by the difference between the maximum and minimum values ​​to obtain the normalized value. This value is the coordinate of the state vector in the speed feature dimension. All eight feature dimensions are calculated in this way. Each state vector corresponds to a coordinate point in the feature space. Connecting the coordinate points of all time points in chronological order forms the feature space trajectory.

[0039] To calculate the Euclidean distance between adjacent coordinate points in the feature space trajectory, take two adjacent coordinate points and extract their coordinate values ​​in eight feature dimensions. Calculate the numerical difference in each dimension (the difference between the previous and subsequent dimensions), square the difference in each dimension, and sum the squared values ​​of all eight dimensions. Finally, take the square root of this sum to obtain the feature distance between the two adjacent coordinate points. Set a distance threshold of 0.3, iterate through the feature distances of all adjacent coordinate points, and group together dense coordinate points with consecutive feature distances less than 0.3 into a single feature node. The coordinates of each feature node are calculated by taking the values ​​of all coordinate points in each feature dimension within the group, summing these values, and dividing by the number of coordinate points in the group to obtain the average value for each dimension. These average values ​​constitute the coordinates of the feature node. Connect all feature nodes in chronological order using directed line segments, with the feature node in the previous time step as the starting point of the line segment and the feature node in the next time step as the ending point. All directed line segments and feature nodes together form a discrete feature topology.

[0040] In this embodiment, the heterogeneity of data from different sensors is eliminated by timestamp alignment and coordinate system transformation, so that the fused data stream has consistency; the sliding window and matrix arrangement make the time series data structured, which is convenient for batch processing.

[0041] In a preferred embodiment of the present invention, a dynamic reference manifold is constructed based on discrete feature topology, and an analysis boundary is determined on the dynamic reference manifold; the analysis boundary is divided according to the geodesic distribution of the discrete feature topology, and a vertex sequence of the analysis boundary is obtained; adjacent edge vectors are calculated based on the vertex sequence; the concavity / convexity of the vertices is determined based on the cross product sign of the adjacent edge vectors; and the analysis boundary is divided into convex and concave segments based on the concavity / convexity of the vertices to generate a topology analysis cell, which may include:

[0042] Based on the spatial coordinates of feature nodes in the discrete feature topology, a dynamic reference manifold is fitted and generated. Based on the spatial distribution density of feature nodes on the dynamic reference manifold, density contour lines are calculated and plotted, with the outermost closed density contour line selected as the analysis boundary. Specifically, this involves: first, collecting the complete eight-dimensional coordinates of all feature nodes in the discrete feature topology, with each feature node corresponding to specific values ​​in eight dimensions such as velocity and attitude features; organizing these coordinates into a structured dataset for fitting; and then using radial basis function interpolation to generate the dynamic reference manifold, using the eight feature dimensions as interpolation variables, and determining the distribution positions of feature nodes in eight-dimensional space. As interpolation node values, a Gaussian function is selected as the radial basis function. The shape parameter ε is determined by the average distance between the interpolation nodes. First, the pairwise Euclidean distances between all interpolation nodes are calculated. These distances are summed and divided by the total number of distances to obtain the average distance. The reciprocal of this average distance is taken as the specific value of ε. In practice, all feature nodes are first determined as interpolation nodes. Then, multiple interpolation points required for manifold fitting are selected. The selection of interpolation points covers the value range of the eight feature dimensions and is evenly distributed. The interval between adjacent interpolation points in each dimension is one-fiftieth of the value range of the corresponding interpolation node.

[0043] For each interpolation point, calculate the Euclidean distance between that point and all interpolation nodes. During the calculation, extract the eight-dimensional coordinates of the interpolation point and the interpolation nodes, calculate the numerical difference in each dimension and square it, then sum the eight squared values ​​and take the square root to obtain the distance value *r*. When substituting *r* into the Gaussian function calculation, first calculate the product of *r* and ε, then square this product, take the negative of the squared result, and calculate the natural exponent to obtain the corresponding basis function value. Construct an interpolation matrix based on the basis function values, where each row of the matrix corresponds to an interpolation point, each column corresponds to an interpolation node, and the matrix elements are the basis function values ​​of the interpolation point and the corresponding interpolation node. The interpolation nodes... The eight-dimensional coordinates of a point are used as the right-hand vector. The interpolation coefficient vector is solved using the LU decomposition method. First, the interpolation matrix is ​​decomposed into a lower triangular matrix and an upper triangular matrix. Then, the lower triangular matrix and the right-hand vector are used to solve for the intermediate vector. Finally, the upper triangular matrix and the intermediate vector are used to solve for the coefficient vector. The basis function values ​​of the points to be interpolated are weighted and summed with the interpolation coefficient vector. Each basis function value is multiplied by the coefficient of the corresponding interpolation node and then summed to obtain the fitted coordinates of the points to be interpolated. The fitted coordinates of all the points to be interpolated and the coordinates of the interpolation nodes together constitute the dynamic reference manifold. After fitting, the actual coordinates of each feature node and the fitted coordinates of the corresponding position on the manifold are calculated. The difference between the target and the target is calculated by squaring the difference in each dimension, summing the results, and averaging them to obtain the average fitting error, ensuring that this error is controlled within 0.05. The spatial distribution density of feature nodes on the dynamic reference manifold is calculated using the K-nearest neighbor method with K set to 20. For each target feature node, the Euclidean distance between it and all other feature nodes in the dataset is calculated. This is done by first extracting the coordinates of the two nodes in eight dimensions, calculating the numerical difference in each dimension, squaring the differences, summing the eight squared values, and taking the square root to obtain the Euclidean distance. The top twenty nodes, sorted by distance from smallest to largest, are selected as the nearest neighbors. The distances between these twenty nodes and the target node are then compared. The maximum Euclidean distance between target nodes is squared, and the spatial distribution density value of the target node is obtained by dividing the square by 20. After calculating the density values ​​of all feature nodes in the same way, density contour lines are drawn based on these density values. First, the maximum and minimum values ​​of all density values ​​are counted, the difference between the two is calculated and divided into five segments, resulting in five density intervals, i.e., five density levels. Each level corresponds to a fixed density threshold. Contour lines are drawn according to the correspondence between the density values ​​at each position on the manifold and the threshold. Each level corresponds to a closed contour line, and the outermost closed contour line is selected as the analysis boundary.

[0044] Based on the projection of directed edges in the discrete feature topology onto the dynamic reference manifold, the geodesic distribution on the dynamic reference manifold is calculated. Based on the intersection points of the geodesic distribution and the analysis boundary, the analysis boundary is divided into multiple continuous boundary segments, and all intersection points are arranged sequentially to generate a vertex sequence of the analysis boundary. Specifically, this includes: first, extracting all directed edges from the discrete feature topology, with each directed edge corresponding to two adjacent feature nodes, and recording the eight-dimensional coordinates of these two nodes as the coordinates of the two endpoints of the edge. Projecting the coordinates of the two endpoints of each directed edge onto the dynamic reference manifold, calculating the Euclidean distance between the endpoint coordinates and each point on the manifold, and selecting the point on the manifold with the smallest distance as the projection endpoint position, thus obtaining the two projection endpoints corresponding to each edge; finding the shortest path connecting these two projection endpoints on the dynamic reference manifold, specifically through iterative calculation, using the two projection endpoints as the start and end points, selecting multiple candidate path points on the manifold, calculating the total length of each candidate path (i.e., the sum of the Euclidean distances between adjacent points on the path), and selecting the candidate path with the smallest total length as the geodesic of the directed edge on the manifold; all directed edges in the discrete feature topology are then processed. The geodesics corresponding to the edges together constitute the geodesic distribution on the dynamic reference manifold. The intersection point of each geodesic with the analysis boundary is determined. Multiple sampling points are uniformly selected along the geodesic from the starting point to the ending point. The distance from each sampling point to the analysis boundary is calculated one by one. The sampling point with the smallest distance that is less than a set small threshold is the intersection point of the geodesic with the analysis boundary. Each intersection point is used as a boundary division point. All division points divide the analysis boundary into multiple continuous boundary segments. Starting from any intersection point on the analysis boundary, all intersection points are traversed in a clockwise direction and sorted. Each intersection point is used as a vertex, forming an ordered sequence of analysis boundary vertices.

[0045] Based on three consecutive vertices in the vertex sequence, calculate the first edge vector from the previous vertex to the middle vertex and the second edge vector from the middle vertex to the next vertex. Perform a cross product of the first and second edge vectors to determine the concavity / convexity attribute of the middle vertex. Specifically, select three consecutive vertices in the sorted vertex sequence, designated as the previous vertex, middle vertex, and next vertex, and extract their complete eight-dimensional coordinates. These eight dimensions correspond to velocity features, attitude features, position change features, obstacle perception features, motion stability features, steering features, acceleration features, and environment matching features. The first three dimensions are the core spatial dimensions of the vehicle's operating state, and their spatial relationships accurately reflect the concavity / convexity of the boundary; therefore, subsequent calculations focus on these three dimensions. Calculate the first edge vector from the previous vertex to the middle vertex. The three core components of this vector are the velocity feature coordinates of the middle vertex minus the velocity feature coordinates of the previous vertex, the attitude feature coordinates of the middle vertex minus the attitude feature coordinates of the previous vertex, and the position change feature coordinates of the middle vertex minus the position change feature coordinates of the previous vertex. The remaining five dimensional components are not involved in the concavity / convexity determination.

[0046] Calculate the second side vector from the middle vertex to the next vertex, again taking the core three-dimensional components: the velocity feature coordinates of the next vertex minus the velocity feature coordinates of the middle vertex, the attitude feature coordinates of the next vertex minus the attitude feature coordinates of the middle vertex, and the position change feature coordinates of the next vertex minus the position change feature coordinates of the middle vertex. Perform a cross product operation on the core three-dimensional components of the two side vectors. First, multiply the second component of the first side vector by the third component of the second side vector, then multiply the third component of the first side vector by the second component of the second side vector, and subtract the result of the second from the first result to obtain the first component of the cross product. Next, multiply the third component of the first side vector by the first component of the second side vector, and subtract the result of the first component multiplied by the second component of the second side vector. The three components are used to obtain the second component of the cross product. Finally, the first component of the first side vector is multiplied by the second component of the second side vector, and then the product of the second component of the first side vector and the first component of the second side vector is subtracted to obtain the third component of the cross product. The third component of the cross product is used as the criterion for determining concavity and convexity. If the value of this component is greater than zero, it means that the boundary at the middle vertex is convex outward, and it is determined to be a convex vertex. If the value of this component is less than zero, it means that the boundary at the middle vertex is concave inward, and it is determined to be a concave vertex. If the value of this component is equal to zero, it means that the three vertices are collinear, and it is determined to be a collinear vertex. At this time, the Euclidean distance from the collinear vertex to the last vertex of the previous marked segment is calculated, and then the Euclidean distance to the first vertex of the next segment to be marked is calculated. The collinear vertex is then assigned to the segment with the smaller distance.

[0047] Based on the concavity / convexity properties, boundary segments formed by consecutive vertices with the same convex property are marked as convex segments, and boundary segments formed by consecutive vertices with the same concave property are marked as concave segments. Based on all marked convex and concave segments on the analysis boundary, a topological analysis cell is generated. Specifically, this includes: traversing all vertices in the order of the vertex sequence, synchronously recording the concavity / convexity properties of each vertex, and establishing segment marking starting from the first vertex; when two or more consecutive convex vertices appear, marking starts from the first convex vertex and ends when a non-convex vertex is encountered, marking the continuous boundary segment formed by sequentially connecting these consecutive convex vertices as a convex segment; when two or more consecutive convex vertices appear... When dealing with concave vertices, marking begins from the first concave vertex and ends when a non-concave vertex is encountered. This continuous boundary segment formed by sequentially connecting these concave vertices is marked as a concave segment. For collinear vertices, if the segment before and after them are both marked with convex vertices, it is directly classified as a convex segment; if the segment before and after them are both marked with concave vertices, it is directly classified as a concave segment; if the segment before and after them are convex and concave respectively, it is classified as the segment closer to the collinear vertex. Combining the distribution of feature nodes within the discrete feature topology, the convex and concave segments on the analysis boundary are used as the basic boundary, and the region is segmented according to the association between the internal feature nodes and each boundary segment.

[0048] For each internal feature node, calculate its vertical distance to each convex and concave segment. The vertical distance is calculated as the shortest distance from the node to the straight line containing the boundary segment. The initial region containing the boundary segment with the smallest vertical distance is determined as the temporary home region of the node. The temporary home regions of all nodes are verified. If the number of nodes in a certain initial region is too small (less than five), the nodes in that region are reassigned to the adjacent region with the next smallest distance, ultimately ensuring that there are a sufficient number of feature nodes in each region. Each region is a closed region formed by the corresponding part of the convex segment, part of the concave segment, and internal feature nodes. This closed region is the topology analysis cell, and each cell contains at least three different boundary segments (convex or concave segments) and corresponding internal feature nodes, ensuring that the cell can completely reflect the distribution of a specific vehicle operating state feature.

[0049] In this embodiment, the determination of the analysis boundary clarifies the effective analysis range; the geodesic division and concavity / convexity judgment make the boundary segmentation more closely match the data characteristics; the topological analysis cell divides the complex feature space into simple sub-regions, realizing refined analysis by region and improving the accuracy of recognizing complex vehicle operation behavior patterns.

[0050] In a preferred embodiment of the present invention, discrete feature topology is mapped to corresponding topology analysis cells, and manifold geometry correction is calculated; based on the manifold geometry correction, a decision geometry structure is constructed, which may include:

[0051] Based on the topological analysis cells and discrete feature topologies, a mapping operation is performed to obtain the attribution relationships. Based on these attribution relationships, the set of normal distances from each feature node to the boundaries of its respective topological analysis cell is calculated to obtain the manifold geometric correction. Specifically, this involves: first, extracting the boundary coordinate information of all topological analysis cells and the eight-dimensional coordinate information of all feature nodes in the discrete feature topology. The boundary coordinate information of the topological analysis cells requires specifying the eight-dimensional coordinates of the two endpoints of each boundary segment, while the feature node coordinates of the discrete feature topology are eight-dimensional feature coordinates. Based on this information, a mapping operation is performed to determine the attribution relationships. For each feature node… Calculate the shortest distance from each feature node to each boundary segment of each topological analysis cell. Specifically, treat each boundary segment as a line segment and construct its parametric equation. Using the feature node coordinates as known points, calculate the projection point of that point onto the line. The projection point is calculated by first finding the vector between the feature node and the starting point of the boundary segment, then taking the dot product of this vector and the direction vector of the boundary segment. Divide the dot product by the square of the magnitude of the direction vector of the boundary segment to obtain the parameter value. If the parameter value is between zero and one, the projection point is a point on the boundary segment. If the parameter value is less than zero, the projection point is the starting point of the boundary segment. If the parameter value is greater than one, the projection point is the ending point of the boundary segment. The projection point is determined as the point on the boundary segment closest to the feature node. Then, the Euclidean distance between the feature node coordinates and the projection point coordinates is calculated, i.e., the numerical difference is calculated dimension by dimension and squared. The square root of the eight squared values ​​is taken; this distance is the shortest distance from the feature node to the boundary segment. For each topology analysis cell, the shortest distances of all its boundary segments are compared, and the smallest distance value is selected. The topology analysis cell corresponding to this distance value is the cell to which the feature node belongs. After all feature nodes have completed cell matching, a complete attribution table of feature node number and its corresponding cell number is formed. Based on the attribution table, the set of normal distances from each feature node to each boundary of its corresponding topology analysis cell is calculated as the manifold geometry correction. Before the normal distance calculation... First, determine the direction of the normal vector of the boundary segment. Take the direction vector of the boundary segment and rotate it 90 degrees clockwise around the midpoint of the segment to obtain two candidate normal vectors. Select a reference point inside the cell and calculate the dot product of the two candidate normal vectors with the reference point and the vector of the midpoint of the boundary segment. The candidate normal vector with a positive dot product is the normal vector pointing into the cell. Calculate the perpendicular distance from the feature node to the boundary segment in the direction of the normal vector. The calculation method is the projection length of the vector of the feature node and the starting point of the boundary segment onto the normal vector. The projection length is obtained by dividing the vector dot product by the normal vector magnitude. Each boundary segment corresponds to a normal distance value. Arrange these distance values ​​in clockwise order in the cell to form the normal distance set of the feature node, which is the manifold geometric correction.

[0052] Based on the manifold geometric corrections corresponding to all feature nodes within a topological analysis cell, a local decision geometry is constructed. The spatial adjacency relationships of all topological analysis cells and their local decision geometry are then integrated to construct a decision geometry. Specifically, this involves: for each topological analysis cell, selecting the numbers of all feature nodes within that cell from the attribution table; extracting the corresponding manifold geometric corrections based on these numbers; calculating statistical parameters for each dimension of the correction (i.e., the number of boundary segments of the cell); for each correction dimension, summing the normal distances of all nodes within that dimension to obtain a total distance; dividing the total distance by the number of nodes to obtain the average value for that dimension; calculating the difference between the normal distance of each node in that dimension and the average value; squaring all differences and summing them to obtain a sum of squares; and dividing the sum of squares by the number of nodes. The number of nodes yields the variance for that dimension, with each dimension corresponding to a set of mean and variance. Simultaneously, the boundary shape parameters of the cell are extracted. The boundary segment length is calculated using the Euclidean distance between the two endpoints of the boundary segment, i.e., the difference in endpoint coordinates is calculated dimension by dimension and squared, with the square root of the eight squared values ​​added together. When calculating the angle between adjacent boundary segments, the direction vectors of two adjacent boundary segments are first taken. The first vector is the coordinates of the current boundary segment's endpoint minus the starting point coordinates, and the second vector is the coordinates of the next boundary segment's endpoint minus the starting point coordinates. The cross product and dot product of the two vectors are calculated. The magnitude of the cross product is divided by the product of the magnitudes of the two vectors to obtain the sine value, and the dot product is divided by the product of the magnitudes of the two vectors to obtain the cosine value. The arctangent function is used to convert the sine and cosine values ​​into an angle, with the angle range controlled between 0 and 180 degrees.

[0053] When extracting the distribution parameters of internal feature nodes, the center coordinates of nodes are calculated by dimension. The center coordinates of each dimension are the sum of the corresponding dimension coordinates of all feature nodes within that cell, divided by the number of nodes. The eight dimensions form the center coordinates. The statistical parameters such as the mean and variance of each dimension, the shape parameters such as the length of each boundary segment and the adjacent angle, and the distribution parameters such as the center coordinates are integrated and arranged in the order of statistical parameters, shape parameters, and distribution parameters to construct the local decision geometry of that cell. The local decision geometry of all topological analysis cells and the boundary coordinate information of each cell are collected. Adjacent cells are determined by calculating the intersection points of the boundary segments of different cells. For any two different cells, each boundary segment of the first cell is intersected with... For each boundary segment of the second cell, the intersection point is calculated. If there is one or more intersection points on the line segment, the two cells are determined to be adjacent. The connection boundary segment number and intersection point coordinates of the adjacent cells are marked as overlapping area information. The local decision geometry of the adjacent cells are fused. For statistical parameters, the difference in mean and variance of the same dimension of the adjacent cells are taken. For shape parameters, the difference in length and angle of the adjacent boundary segments are taken. For distribution parameters, the Euclidean distance of the center coordinates is taken. These differences and distances are used as interaction parameters. All local decision geometry structures are arranged into a network structure according to spatial adjacency. Interaction parameters are added at the connection positions of adjacent structures to form a global decision geometry structure containing local features and spatial correlation information.

[0054] In this embodiment, the attribution relationship is determined by calculating the distance from the feature node to the cell boundary, ensuring the accuracy of the mapping; the manifold geometric correction can accurately reflect the deviation between the node and the boundary.

[0055] In a preferred embodiment of the present invention, based on the decision geometry structure, the multidimensional state matrix is ​​analyzed to generate an event trigger sequence; based on the time reference in the event trigger sequence, data slices within the event time neighborhood are extracted from the multidimensional state matrix to construct a spatiotemporal data cube, which may include:

[0056] Based on the decision geometry and multidimensional state matrix, each state vector of the multidimensional state matrix is ​​mapped to the decision geometry to obtain the geometric position coordinates corresponding to each state vector. Specifically, this includes: extracting the ten-dimensional state vector corresponding to each row of the multidimensional state matrix. The ten-dimensional state vector is, in order, longitude, latitude, altitude, driving speed, acceleration, heading angle, yaw rate, wheel speed difference, obstacle distance, and obstacle orientation. First, according to the feature space mapping rules, the ten-dimensional state vector is converted into eight-dimensional feature space coordinates. The correspondence between the eight-dimensional features and the ten-dimensional states is fixed: speed features correspond to driving speed and wheel speed difference, attitude features correspond to heading angle and yaw rate, position change features correspond to changes in longitude, latitude, and altitude, obstacle perception features correspond to obstacle distance and orientation, motion stability features correspond to acceleration fluctuations, steering features correspond to the rate of change of heading angle and yaw rate, acceleration features correspond to the original acceleration data, and environment matching features correspond to the visual and lidar environmental data processing results. During the transformation, the coordinates of each feature dimension are obtained through normalization calculation. Taking the speed feature as an example, the values ​​of driving speed and wheel speed difference in all state vectors are first counted. The minimum value of the two is taken as the global minimum value and the maximum value is taken as the global maximum value (in low-speed autonomous vehicle scenarios, the global range of driving speed is usually 0 to 10 kilometers per hour, and the global range of wheel speed difference is usually 0 to 0.5 meters per second). The average value of driving speed and wheel speed difference in the current state vector is subtracted from the global minimum value. The difference is divided by the difference between the global maximum value and the global minimum value to obtain the coordinates of the speed feature dimension. The other seven feature dimensions are calculated according to the same logic.

[0057] Based on the spatial extent of the decision geometry, the eight-dimensional feature space coordinates are mapped onto the global decision geometry. The Euclidean distance between these coordinates and the center coordinates of all local decision geometry structures is calculated, and the local region with the smallest distance is the region to which the coordinates belong. The center coordinates of a local region are the sum of the coordinates of each dimension of all feature nodes within that region divided by the number of nodes. The scaling factor is determined by the boundary range of the local region, taking the reciprocal of the difference between the maximum and minimum values ​​of each dimension's boundary to ensure that the adjusted coordinates fall within the range of 0 to 1 in the local region (this range is a general range for feature space standardization and can be fine-tuned to 0.1 to 0.9 according to the feature distribution density to avoid subsequent calculation deviations caused by the coordinates being at the boundary). The adjustment method is to subtract the corresponding dimension of the center coordinates of the local region from each dimension of the feature space coordinates, and then multiply the result by the scaling factor of the corresponding dimension to obtain the geometric position coordinates of the state vector.

[0058] Based on the sequence of changes in geometric position coordinates over time, the time points where abrupt changes in geometric position coordinates occur are detected, resulting in a set of event boundary points. Based on this set, the continuous timeline is divided into multiple time periods, and a unique event identifier is assigned to each period to generate an event trigger sequence. Specifically, this includes: arranging the geometric position coordinates of each state vector in ascending order of timestamps to form a sequence of changes in geometric position coordinates over time; and calculating the Euclidean distance between the geometric position coordinates of two adjacent time points: extracting the eight-dimensional values ​​of the coordinates of the previous and subsequent time points, and calculating the distance dimension by dimension. The difference between the value at the next time point and the value at the previous time point is calculated. The squares of each difference are then summed, and the square root of the sum is taken to obtain the distance value. The distance threshold is set to 1.2 to 1.8 times the average distance value of all adjacent time points. The specific value is dynamically selected according to the vehicle operation scenario—1.2 times is used in low-speed and stable driving scenarios to reduce false judgments, and 1.8 times is used in complex road conditions (such as intersections and curves) to avoid missed judgments. The overall value conforms to the coordinate change pattern of the vehicle during stable operation. When the distance value between a certain time point and the previous time point is greater than the threshold, the time point is determined to be the starting point of the event boundary.

[0059] Starting from the next time point after the starting point, monitor subsequent distance values, continuously counting the number of time points where the distance value is less than a threshold. When three or more consecutive time points have distance values ​​less than the threshold (the number of consecutive values ​​can be adjusted between 2 and 4; for low-speed scenarios, use 3 to ensure stability), the first of these three time points is determined as the event boundary termination point. If no three consecutive time points less than the threshold appear after the starting point, the last time point of the current sequence is taken as the termination point. All starting points are paired with their corresponding termination points to form an event boundary point set. If a single starting point has no subsequent termination point, it is paired with the end of the sequence. Based on the boundary point set, the continuous timeline is divided into multiple time periods, each time period corresponding to a vehicle operation event. Each event is assigned a 16-bit binary event identifier: the first six bits are the event type prefix, determined by the range of speed variation within the event period. For example, a speed increase exceeding 0.4 to 0.6 meters per second (0.4 meters per second for small low-speed autonomous vehicles, and 0.6 meters per second for medium-sized vehicles) corresponds to 000001, a speed decrease exceeding 0.4 to 0.6 meters per second corresponds to 000010, and constant speed operation (speed fluctuation less than 0.2 meters per second) corresponds to 000100. The last ten bits are the time code suffix, converting the number of seconds of the event start time into a ten-bit binary number, padding with zeros at the high bits if less than ten bits. The start timestamp, end timestamp, and event identifier of each event are recorded in the order of the start timestamp, and then integrated to generate an event trigger sequence.

[0060] Based on the start and end time references corresponding to each event identifier in the event trigger sequence, data rows covering the time windows before and after the event are extracted from the multidimensional state matrix to obtain event data slices. Based on the event data slices, multiple data slices are aligned and integrated along the time dimension to construct a spatiotemporal data cube. Specifically, according to the start and end timestamps of each event in the event trigger sequence, the event time neighborhood is set to 1.5 to 2.5 seconds before the start timestamp and 2.5 to 3.5 seconds after the end timestamp. The specific range is adjusted according to the event type. For sudden events (such as emergency braking and steering), a larger range (2.5 seconds before and 3.5 seconds after) is used to cover the complete state transition, while for stable events (such as constant speed driving), a smaller range (1.5 seconds before and 2.5 seconds after) is used. This range can fully cover the vehicle state transition process before and after the event, which meets the time requirements for low-speed autonomous vehicle event analysis.

[0061] The timestamp interval of the multidimensional state matrix is ​​0.1 seconds (the interval can be adjusted between 0.05 and 0.2 seconds; 0.05 seconds is used for high-precision analysis, and 0.1 seconds for routine analysis). The timestamp range corresponding to the time neighborhood is calculated as follows: subtract 15 to 25 time intervals (corresponding to 1.5 to 2.5 seconds) from the start timestamp to obtain the neighborhood start timestamp; add 25 to 35 time intervals (corresponding to 2.5 to 3.5 seconds) to the end timestamp to obtain the neighborhood end timestamp. All data rows with timestamps between the neighborhood start and end timestamps are extracted from the multidimensional state matrix. Each data row corresponds to ten-dimensional state data at a given time point. These data rows are arranged in ascending order of timestamps to form a data slice for that event. Each slice contains 50 to 60 time point data points (corresponding to 5 to 6 seconds of duration). Time dimension alignment is performed on all event data slices: using the start timestamp of each event as the base time, the time dimension of each data slice within the slice is calculated. The difference between the data row timestamp and the reference time is used as the time offset. A negative offset indicates before the reference time, and a positive offset indicates after the reference time. The time offset of all slices is uniformly adjusted to range from -2.5 seconds to +3.5 seconds, corresponding to 50 to 60 fixed offset positions. When offset data is missing, linear interpolation is used to complete it (the interpolation error is controlled within 0.01 to ensure the accuracy of the completed data). The aligned data slices are integrated to construct a spatiotemporal data cube: the time dimension consists of 50 to 60 fixed offset positions, corresponding to the time series before and after the event; the spatial dimension consists of three vehicle position-related dimensions: longitude, latitude, and altitude, directly extracting the corresponding dimensions from the state data; the state dimension consists of ten original state dimensions such as driving speed, acceleration, and heading angle; after the three dimensions are cross-integrated, a three-dimensional spatiotemporal data cube of 50 to 60 (time) multiplied by three (space) multiplied by ten (state) is formed.

[0062] This embodiment maps state vectors to decision geometry to associate data with the analysis framework; distance mutation detection can accurately identify event boundaries and ensure precise event segmentation.

[0063] In a preferred embodiment of the present invention, feature extraction from a spatiotemporal data cube to obtain event feature fingerprints and generate event feature codes may include:

[0064] Based on the spatiotemporal data cube, statistical features are calculated along the time, space, and state dimensions to obtain a multi-dimensional statistical feature set. Based on this multi-dimensional statistical feature set, a subset of key features is selected through feature importance ranking and redundancy analysis to form an event feature fingerprint. Specifically, this includes calculating statistical features along the time, space, and state dimensions of the spatiotemporal data cube. All calculations are based on the actual state data of vehicle operation within the cube to ensure that the features align with monitoring requirements. In the time dimension, fifty fixed-offset position data points are selected (corresponding to five seconds before and after the event). For each of the ten state dimensions (longitude, latitude, altitude, etc.), six statistical characteristics are calculated. Taking the driving speed dimension as an example, the average value is calculated by adding the speed data at fifty time points and dividing by fifty; the maximum value is the largest value among the fifty data points, and the minimum value is the smallest value; the standard deviation is calculated by adding the squares of the differences between each speed data point and the average value, dividing the sum by fifty, and then taking the square root; the first difference is calculated by subtracting the speed data at the previous time point from the speed data at the next time point, resulting in forty-nine differences, and taking the largest absolute value among these differences as the maximum value of the first difference and the smallest value as the minimum value of the first difference.

[0065] Ten state dimensions are calculated one by one according to this logic, resulting in a total of sixty time statistical features. In the spatial dimension, all data in the three spatial dimensions of longitude, latitude, and altitude are extracted from the cube, and two features, distribution entropy and cluster center distance, are calculated. For distribution entropy calculation, the operating range of the low-speed autonomous vehicle is first divided into grids: longitude is one grid for every 0.0001 degrees (approximately 10 meters of precision), latitude is one grid for every 0.0001 degrees, and altitude is one grid for every 0.1 meters. The number of spatial data points in each grid is counted, and this number is divided by the total number of data points to obtain the grid probability. Then, the probability of each grid is multiplied by its natural logarithm, and the sum of all results is taken as the negative value to obtain the distribution entropy. For cluster center distance calculation, the K-means clustering algorithm (K value is 3 to 5, adjusted according to data density) is used to iteratively divide the spatial data into K clusters. The center coordinates of each cluster are calculated by summing the coordinates of all data points in that cluster according to their corresponding dimensions and dividing by the number of data points. The Euclidean distance from each data point to its cluster center is calculated, and the sum of all distances is divided by the total number of data points to obtain the cluster center distance, ultimately yielding two spatial statistical features. In terms of state dimensions, the traces of the correlation coefficient matrix and covariance matrix among the ten state dimensions are calculated. When calculating the correlation coefficient matrix, any two state dimensions (such as driving speed and acceleration) are selected, and their covariance is calculated. The covariance is the sum of the products of the differences between each data point in the two dimensions and its own mean, divided by the number of data points. The covariance is then divided by the product of the standard deviations of the two dimensions to obtain the correlation coefficient. All dimensions are combined pairwise to form a 10x10 correlation coefficient matrix.

[0066] When calculating the covariance matrix, each diagonal element represents the variance of the corresponding state dimension, and the off-diagonal elements represent the covariance of the two corresponding dimensions. The trace of the covariance matrix is ​​obtained by summing all diagonal elements, ultimately yielding two state statistical features. Sixty temporal features, two spatial features, and two state features are integrated to form a 64-dimensional multi-dimensional statistical feature set. A random forest algorithm is used to select key features. First, 200 to 300 decision trees are constructed (the number of trees is adjusted according to the feature dimensions). Each tree is constructed by randomly selecting a portion of samples and features, with 70% of the samples and 50% of the features being randomly selected. For each feature, its value across all decision trees is calculated. The reduction in classification error is taken as the average reduction across all trees, which is then used as the importance score for that feature. An importance threshold of 0.01 to 0.05 is set (0.01 for complex scenarios and 0.05 for simple scenarios). Features with scores greater than the threshold are retained to form a preliminary feature set. The Pearson correlation coefficient between any two features in the preliminary feature set is calculated in the same way as the correlation coefficient for the state dimension. A correlation coefficient threshold of 0.7 to 0.8 is set (0.7 for high accuracy requirements and 0.8 for high efficiency requirements). If the correlation coefficient between two features is greater than the threshold, the feature with the higher importance score is retained, and the feature with the lower score is deleted. The final subset of key features is the event feature fingerprint.

[0067] Based on the event feature fingerprint, each feature value in the key feature subset is standardized and discretized to generate an event feature code. Specifically, this involves: first standardizing the event feature fingerprint, then performing discretization to generate a fixed-format event feature code; during standardization, for each feature in the fingerprint, the feature values ​​of all vehicle operation events are iterated through, recording the minimum and maximum values; for example, if the minimum value of a feature across all events is 0.2 and the maximum value is 1.8, and the current event's feature value is 1.0, then the standardization calculation is to subtract 0.2 from 1.0 to get the difference 0.8, and then divide 0.8 by the difference 1.8 minus 0.2 to get 1.6, resulting in a standardized value of 0.5. All features are converted in this way to ensure the values ​​fall within the 0 to 1 range, eliminating dimensional differences between different features; after standardization, discretization is performed, dividing the 0 to 1 range into ten equal intervals. The sub-intervals range sequentially from 0 to 0.1, 0.1 to 0.2, ..., 0.9 to 1.0. Each sub-interval corresponds to a unique four-bit binary code, with the following fixed encoding rules: 0 to 0.1 corresponds to 0000, 0.1 to 0.2 corresponds to 0001, 0.2 to 0.3 corresponds to 0010, 0.3 to 0.4 corresponds to 0011, 0.4 to 0.5 corresponds to 0100, 0.5 to 0.6 corresponds to 0101, 0.6 to 0.7 corresponds to 0110, 0.7 to 0.8 corresponds to 0111, 0.8 to 0.9 corresponds to 1000, and 0.9 to 1.0 corresponds to 1001. Following the fixed order of features in the event fingerprint (e.g., time features first, then spatial features, and finally state features), the four-bit binary codes corresponding to the intervals to which the standardized values ​​of each feature belong are concatenated sequentially. If the feature fingerprint contains fifteen key features, the concatenation results in a sixty-bit binary sequence, which is the event feature code of the current event.

[0068] In this embodiment, the standardized and discretized event feature encoding unifies the feature format, which facilitates subsequent model processing and improves the reliability and efficiency of feature application.

[0069] like Figure 2 As shown, in another preferred embodiment of the present invention, inputting event feature encoding into a pre-trained deep belief network to obtain an event risk level assessment result may include:

[0070] Based on the event feature encoding, the numerical sequence of the event feature encoding is input into the input layer of the pre-trained deep belief network. Based on the numerical sequence of the event feature encoding received by the input layer, the feature activation state of the first hidden layer is obtained through the operation of the first hidden layer. Specifically, this includes: first, clarifying the correspondence between the event feature encoding and the input layer nodes; the sixty-bit binary encoding corresponds to the sixty nodes of the input layer sequentially from left to right, i.e., the first bit of the encoding corresponds to input layer node 1, the second bit corresponds to input layer node 2, and so on, up to the sixtieth bit corresponding to input layer node 60; converting the sixty-bit binary sequence into a numerical sequence, directly converting 0 in the encoding to the numerical value 0, and 1 to the numerical value 1, forming an input sequence composed of sixty 0s or 1s; inputting this sequence bit by bit into the input layer of the pre-trained deep belief network, with each input layer node only receiving the numerical value converted from the corresponding encoded bit.

[0071] The input value of each node in the first hidden layer is calculated. Taking node 1 of the first hidden layer as an example, the input value calculation logic is as follows: the value of input node 1 is multiplied by the weight from input node 1 to the first hidden layer node 1, plus the value of input node 2 multiplied by the weight from input node 2 to the first hidden layer node 1, and so on, up to the value of input node 60 multiplied by the weight from input node 60 to the first hidden layer node 1. Finally, the bias term of the first hidden layer node 1 is added. Here, the weights and bias term are fixed values ​​obtained through iterative optimization of the first restricted Boltzmann machine in the pre-training stage of the deep belief network using the contrastive divergence algorithm. The weights range from -0.05 to +0.05, and the bias term... The range of terms is concentrated between -0.1 and +0.1. The input value of each node in the first hidden layer is substituted into the sigmoid activation function. The calculation method is to divide 1 by 1 plus the sum of the negative input value of the natural exponent. For example, when the input value is 1, the result is 1 divided by 1 plus the sum of the negative 1 of the natural exponent, which is approximately 0.731. This result is the activation probability of the node. An activation threshold of 0.5 is set. If the activation probability is greater than 0.5, the node's activation state is marked as 1; if the activation probability is less than or equal to 0.5, the activation state is marked as 0. This logic is used to complete the calculation of the ninety nodes in the first hidden layer, forming a sequence of ninety 0s or 1s. This sequence is the feature activation state of the first hidden layer.

[0072] Based on the feature activation states of the first hidden layer, the high-level feature states of the second hidden layer are obtained through calculations in the second hidden layer. Based on the high-level feature states of the second hidden layer, the probability distribution of the event corresponding to multiple risk levels is obtained through calculations in the output layer. Specifically, this includes: inputting the feature activation states of the ninety nodes of the first hidden layer sequentially into the ninety corresponding nodes of the second hidden layer, i.e., inputting the activation state of node 1 in the first hidden layer into node 1 of the second hidden layer, up to inputting the activation state of node 90 in the first hidden layer into node 90 of the second hidden layer; and calculating the input value for each node in the second hidden layer. Taking the second hidden layer node 1 as an example, the input value is the activation state of the first hidden layer node 1 multiplied by the weight from the first hidden layer node 1 to the second hidden layer node 1, plus the activation state of the first hidden layer node 2 multiplied by the weight from the first hidden layer node 2 to the second hidden layer node 1, and so on, until the activation state of the first hidden layer node 90 multiplied by the weight from the first hidden layer node 90 to the second hidden layer node 1. Finally, the bias term of the second hidden layer node 1 is added. Here, the weights and bias terms are fixed values ​​after optimization of the second restricted Boltzmann machine in the pre-training stage, and their value range is consistent with the weight bias of the first hidden layer. Substituting the input value into the sigmoid activation function, the calculation method is the same as in step 701. For example, when the input value is negative 1, the calculation result is 1 divided by 1 plus the natural exponent of 1, which is approximately equal to 0.269. This is then compared with the threshold of 0.5 to determine the activation state. After completing the calculation of the forty-five nodes of the second hidden layer, a sequence of forty-five 0s or 1s is formed, which is the high-level feature state of the second hidden layer.

[0073] Forty-five high-level feature states are input sequentially into the four nodes of the output layer. The input value of each output node needs to be accumulated by the contribution values ​​of the forty-five high-level feature states. Taking output node 1 (corresponding to low risk) as an example, the input value is the activation state of the second hidden layer node 1 multiplied by the weight from the second hidden layer node 1 to the output layer node 1, plus the activation state of the second hidden layer node 2 multiplied by the weight from the second hidden layer node 2 to the output layer node 1, and so on, up to the activation state of the second hidden layer node 45 multiplied by the weight from the second hidden layer node 45 to the output layer node 1. Finally, the bias term of output layer node 1 is added. Here, the weights and bias terms are optimized by gradient descent in the backpropagation layer during the pre-training stage. The input values ​​of the four output nodes are fixed after optimization. The input values ​​of the four output nodes are substituted into the softmax activation function. The calculation logic is as follows: calculate the natural index for the input value of each output node, and then divide the natural index by the sum of the natural indices of the four output nodes to obtain the probability of the node. For example, if the input values ​​of the four output nodes are 0.8, 0.5, 0.3, and 0.1, the corresponding natural indices are approximately 2.225, 1.648, 1.349, and 1.105, with a sum of approximately 6.327. The probability of output node 1 is 2.225 divided by 6.327, which is approximately 0.352. The sum of the probabilities of the four nodes is always 1, forming a probability distribution of low risk, medium risk, high risk, and very high risk for the event.

[0074] Based on the probability distribution of an event corresponding to multiple risk levels, the risk level corresponding to the highest probability value is selected to obtain the event risk level assessment result. Specifically, this includes: clarifying the fixed correspondence between output layer nodes and risk levels: output layer node 1 corresponds to low risk, node 2 to medium risk, node 3 to high risk, and node 4 to extremely high risk; extracting the probability values ​​of the four nodes and recording them as low risk probability, medium risk probability, high risk probability, and extremely high risk probability; sorting the four probability values ​​by size and selecting the probability value with the largest value, whose corresponding risk level is the preliminary event risk level assessment result; if two or more probability values ​​are equal and all are the maximum value, a secondary judgment mechanism is initiated: first, the nodes with an activation state of 1 in the high-level feature state of the second hidden layer are counted. The number of active nodes reflects the complexity of the event characteristics; the more active nodes, the richer the abnormal characteristics of the event. If there are fewer than twenty active nodes, the event is considered to have low complexity, and a lower risk level is selected from the risk levels corresponding to equal probabilities. If there are more than twenty active nodes, the event is considered to have high complexity, and a higher risk level is selected from the risk levels corresponding to equal probabilities. If there are exactly twenty active nodes, the activation status of the encoding bits corresponding to the core safety features of low-speed autonomous vehicles (such as obstacle distance and changes in driving speed) is considered. If the number of core feature encoding bits that are 1 is less than 3, a lower risk level is selected; otherwise, a higher risk level is selected. This logic ensures that each event receives a unique risk level assessment result that fits the actual operating scenario.

[0075] This embodiment, by appropriately setting the layers and number of nodes in the deep belief network and combining it with ample training on historical data, enables the network to accurately learn the correlation between feature encoding and risk level. Layered computation progressively extracts high-level features, improving the accuracy of risk assessment.

[0076] In a preferred embodiment of the present invention, the event trigger sequence, spatiotemporal data cube, event feature encoding, and event risk level assessment results are fused to generate a structured event analysis report, which may include:

[0077] Based on the event trigger sequence, spatiotemporal data cube, event feature encoding, and event risk level assessment results, the event's time stamp, spatial range, feature vector, and risk level are extracted to obtain a set of key event elements. Specifically, key event elements are extracted in a fixed order of temporal and spatial risk characteristics. All extraction operations are completed through the element extraction module of the vehicle-mounted data processing terminal to ensure the consistency between the elements and the source data. During timestamp extraction, the start and end timestamp fields are located from the structured data of the event trigger sequence. The field format is YYYY-MM-DDHH:MM:SS.fff, accurate to milliseconds. After extracting the 16-bit binary event identifier, it is converted to a decimal value through bit weight calculation. The calculation method is to multiply each bit by 2 raised to the power of 1 from the first bit to the sixteenth bit, and then sum them up. For example, when converting binary 0000010000000001 to decimal, the first bit is 0 multiplied by 2 to the power of 15, the sixth bit is 1 multiplied by 2 to the power of 10, and the sixteenth bit is 1 multiplied by 2 to the power of 0, resulting in 1025. During spatial range extraction, the spatial retrieval interface of the spatiotemporal data cube is called. The start and end timestamps of the event time neighborhood are input, and the longitude, latitude, and altitude fields of all data rows within this time range are traversed. The maximum and minimum values ​​of each field are calculated using data statistics tools. For example, after traversing the longitude, the maximum value is 116.4520 degrees and the minimum value is 116.451. 0 degrees, forming a spatial boundary description of longitude 116.4510 to 116.4520, latitude 39.9050 to 39.9060, and altitude 20.5 to 21.0; during feature vector extraction, a 60-bit binary sequence is read from the storage address of the event feature code, and split into groups according to the order of the first four bits, the middle four bits, and the last four bits, with each group corresponding to a key feature, maintaining the original left-to-right order; during risk level label extraction, the risk level field is located from the output file of the deep belief network, and a unique label is obtained from low risk, medium risk, high risk, and extremely high risk; after extraction, each element is organized in the format of attribute name followed by attribute value, for example, start time colon 2025-12-03 10:00:00.000 and end longitude colon 116.4520, and then encapsulated into event key element set tags using XML tags. Each element is arranged in the order of time-marked spatial range feature vector risk level, forming a structured event key element set.

[0078] Based on the set of key event elements, a standardized data structure is organized to obtain an initial event reporting framework. According to the time correspondence between the event trigger sequence and the spatiotemporal data cube, the state change data of the spatiotemporal data cube is filled into the corresponding data fields of the initial event reporting framework, resulting in an event reporting framework containing state data. Specifically, this includes: based on the element types of the key event element set, using the standardized data structure specified in GB / T30146-2013 Vehicle Information Interaction System, an initial event reporting framework is constructed using a report framework generation tool; the four core data fields of the framework are all set with attributes such as field identifier, data type, length, and whether they are required: the basic information field contains an event identifier decimal value (10 digits required), a start timestamp time format (23 digits required), and an end... The fields include: a 23-digit timestamp (required), a 100-digit spatial range string (required), and a 15-digit vehicle terminal number string (required); the event status data field has subfields arranged in order of time offset from -2.5 seconds to +3.5 seconds, each named offset_X seconds, where X is the specific offset, such as -2.5, -2.4, up to 3.5. Each subfield contains ten status dimension subfields, such as longitude floating-point type with four decimal places and latitude floating-point type with four decimal places, all of which are required; the feature encoding field is set according to the bit order of sixty bits, with encoding bits 1 to 60 being Boolean type (required), and a reserved feature type label subfield; the risk conclusion field includes a 10-digit risk level label string (required) and a 200-digit risk description string (required). After the framework is built, the correspondence between the starting timestamp of the event trigger sequence and the time offset of the spatiotemporal data cube is extracted using a time association tool. A mapping table is established where the event starting timestamp plus the offset equals the specific time point. For example, if the event starting time is 10:00:00:00:00, the offset of -2.5 seconds corresponds to the specific time of 9:59:57:50:00. Based on the mapping table, the ten-dimensional state data of the corresponding time point is called from the spatiotemporal data cube. The data is then filled into the ten state dimensions of the corresponding offset_X seconds subfield in the event state data field in a fixed order of longitude, latitude, altitude, driving speed, acceleration, heading angle, yaw rate, wheel speed difference, obstacle distance, and obstacle orientation. After filling in the data, the numerical range of each subfield, such as driving speed from 0 to 10 kilometers per hour, is checked using a field validation tool to ensure the data is valid, resulting in an event reporting framework containing state data.

[0079] Based on the numerical values ​​of the event feature codes and the labels of the event risk level assessment results, the feature code values ​​and risk level labels are filled into the feature and conclusion fields of the event report framework containing state data to obtain a fully filled event report framework. Specifically, when filling the feature codes, the mapping configuration file between the feature codes and fields is called. This file presets the correspondence rules between sixty-bit codes and feature types: code bits 1 to 4 correspond to the average value of the time dimension, 5 to 8 correspond to the maximum value of the time dimension, 9 to 12 correspond to the minimum value of the time dimension, 13-16 correspond to the standard deviation of the time dimension, 17 to 20 correspond to the maximum value of the first difference of the time dimension, 21 to 24 correspond to the minimum value of the first difference of the time dimension, 25 to 28 correspond to the distribution entropy of the spatial dimension, 29 to 32 correspond to the cluster center distance of the spatial dimension, 33 to 36 correspond to the trace of the covariance matrix of the state dimension, and 37 to 60 correspond to the key values ​​of the correlation coefficient matrix of the state dimension, which are selected from high to low correlation.

[0080] According to this rule, the 60-bit binary sequence is filled bit by bit into the coding bits 1 to 60 of the feature coding field, and the corresponding feature type is filled in the feature type labeling subfield. For example, when coding bits 1 to 4 are filled with 0100, the average value of the time dimension is labeled. When filling in the risk level, the risk level label is filled into the risk level label field of the risk conclusion field, and then the risk description is supplemented according to the preset risk level description rules: Low risk: The vehicle's speed fluctuation is less than 0.2 meters per second, the obstacle distance is greater than 5 meters, there is no abnormal attitude change, and the vehicle is running smoothly; Medium risk: The vehicle's speed fluctuation is 0.2 to 0.5 meters per second, or the obstacle distance is 3 to 5 meters, and the attitude change is gentle, requiring slight attention; High risk: The vehicle's speed changes suddenly by more than 0.5 meters per second, resulting in sudden braking or acceleration, or the heading angle changes by more than 30 degrees, resulting in sudden turning, and the obstacle distance is 1 to 3 meters, requiring close attention; Very high risk: The vehicle's speed changes suddenly by more than 1 meter per second, or the obstacle distance is less than 1 meter, or there is an abnormal yaw rate greater than 10 degrees per second, indicating a safety hazard, requiring immediate intervention. After the filling is completed, the integrity verification process is started: verify that there are no empty values ​​in the required fields of the basic information field, that all offset quantum fields of the event status data field have ten-dimensional data, that the sixty-bit code of the feature encoding field is not missing, and that the label and description of the risk conclusion field match. If there are missing or mismatched values, return to the previous steps to correct them. After the verification is passed, the fully filled event report framework is obtained.

[0081] Based on the fully populated event report framework, a formatted conversion is performed to generate a structured event analysis report. Specifically, this includes: formatting the report using Extensible Markup Language (Extreme Markup Language) version 1.0, with the conversion process automatically executed by the report generation module of the vehicle terminal, outputting a structured event analysis report; generating report header information, obtaining the current time as the report generation time through the system time interface, in the format of year, month, day, hour, minute, and second; the report number is a combination of the event identifier's decimal value and the generation timestamp, such as 1025-20251203100000; and generating a device identifier to read the unique hardware of the vehicle terminal. The item number, such as LVUV20250001, is wrapped in a report header label, and each field is labeled with a sub-label using the generation time, report number, or device identifier. The core data field is transformed as follows: the basic information field is wrapped in a basic information label, and each field is labeled with its corresponding sub-label, such as event identifier label 1025, event identifier label, and start timestamp label 2025-12-03 10:00:00.000. The event status data field is wrapped in a status data label, and is organized using table labels. The table header contains eleven columns, including time offset, longitude, and latitude, and the table body is arranged from negative to positive offset. The data is filled in the following order: row label, cell label, -2.5 seconds, cell label, cell label, 116.4510, cell label, row label; the feature coding field is wrapped with feature coding labels, and internally labeled with coding bits, with bit order = 1, feature type = time, dimension average label, and coding bit label format for each bit; the risk conclusion field is wrapped with risk conclusion labels, including risk level labels, low risk, risk level labels, risk description labels, and risk description labels; finally, the report footer information is generated: the data checksum is calculated using the CRC-32 algorithm, and the data in the report header and core data fields are verified. The text content is converted into ASCII code values, summed, and modulo 2 to the power of 32 to obtain a 32-bit checksum, which is then labeled with a checksum tag. The tail also contains a data source tag, labeled as the low-speed unmanned vehicle onboard sensor set plus deep belief network evaluation result. All tags are nested in the order of header basic information, state data feature encoding, risk conclusion, and tail, forming a complete and extensible markup language document, which is saved as a report number.xml file, such as 1025-20251203100000.xml, and simultaneously stored in the onboard storage module and cloud database to form a structured event analysis report.

[0082] This embodiment ensures that the report contains core information such as the temporal and spatial characteristics and risks of the event by extracting key elements from multi-source data; the standardized report framework and format conversion make the report structure clear and standardized, which is convenient for subsequent storage, query and analysis.

[0083] Embodiments of the present invention also provide a computing device, including: a processor and a memory storing a computer program, wherein the computer program, when executed by the processor, performs the system as described above. All implementations in the above system embodiments are applicable to this embodiment and can achieve the same technical effects.

[0084] Embodiments of the present invention also provide a computer-readable storage medium storing instructions that, when executed on a computer, cause the computer to perform the system as described above. All implementations in the above system embodiments are applicable to this embodiment and can achieve the same technical effects.

[0085] The above description represents the preferred embodiments of the present invention. It should be noted that those skilled in the art can make various improvements and modifications without departing from the principles of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.

Claims

1. A low-speed unmanned vehicle operation status data recording and analysis system, characterized in that, include: The acquisition module is used to collect multi-dimensional spatiotemporal data streams of vehicle operating status; The fusion module is used to fuse multidimensional spatiotemporal data streams, generate a multidimensional state matrix, and convert it into discrete feature topology through mapping; The partitioning module is used to fit and generate a dynamic reference manifold based on the spatial coordinates of feature nodes in the discrete feature topology; calculate and draw density contour lines based on the spatial distribution density of feature nodes on the dynamic reference manifold, and select the outermost closed density contour lines as the analysis boundary; calculate the geodesic distribution on the dynamic reference manifold based on the projection of directed edges in the discrete feature topology onto the dynamic reference manifold; divide the analysis boundary into multiple continuous boundary segments based on the intersection points of the geodesic distribution and the analysis boundary, and arrange all intersection points in order to generate a vertex sequence of the analysis boundary; calculate the first edge vector from the previous vertex to the middle vertex and the second edge vector from the middle vertex to the next vertex based on three consecutive adjacent vertices in the vertex sequence; and determine the concavity / convexity property of the middle vertex by performing a cross product of the first and second edge vectors. Based on the concavity / convexity properties, boundary segments formed by consecutive vertices with the same convex property are marked as convex segments, and boundary segments formed by consecutive vertices with the same concave property are marked as concave segments; based on all the marked convex and concave segments on the analysis boundary, a topology analysis cell is generated. The computation module is used to perform mapping operations based on topological analysis cells and discrete feature topologies to obtain the attribution relationship; based on the attribution relationship, it calculates the set of normal distances from each feature node to the boundaries of its respective topological analysis cell to obtain the manifold geometric correction; based on the manifold geometric corrections corresponding to all feature nodes within a topological analysis cell, it constructs the local decision geometry structure; and it merges the spatial adjacency relationships of all topological analysis cells and the local decision geometry structure to construct the decision geometry structure. The parsing module is used to parse the multidimensional state matrix based on the decision geometry and generate event triggering sequences; Based on the time reference in the event triggering sequence, data slices within the event time neighborhood are extracted from the multidimensional state matrix to construct a spatiotemporal data cube. The extraction module is used to extract features from the spatiotemporal data cube, obtain event feature fingerprints, and generate event feature codes. The training module is used to input event feature encodings into a pre-trained deep belief network to obtain event risk level assessment results; The generation module is used to integrate event trigger sequences, spatiotemporal data cubes, event feature codes, and event risk level assessment results to generate a structured event analysis report.

2. The low-speed unmanned vehicle operation status data recording and analysis system according to claim 1, characterized in that, The multidimensional spatiotemporal data stream is fused to generate a multidimensional state matrix, and then transformed into a discrete feature topology through mapping, including: Timestamp alignment and spatial coordinate system transformation are performed on asynchronous sensor data in a multidimensional spatiotemporal data stream to generate a fused data stream. Based on the fused data stream, a sliding window of preset time length is used to continuously extract the fused data stream, and the time series data in each extraction window are arranged and combined according to the state dimension to generate a structured multidimensional state matrix. Based on the structured multidimensional state matrix, the state vector corresponding to each time step in the multidimensional state matrix is ​​mapped to a preset multidimensional feature space. The coordinate position of each state vector in the feature space is calculated to generate a feature space trajectory. Based on the feature space trajectory, the feature distance between adjacent coordinate points in the feature space trajectory is calculated. According to the preset distance threshold, dense coordinate points with feature distance less than the threshold are aggregated into a feature node. All generated feature nodes are connected according to the time order to generate a discrete feature topology.

3. The low-speed unmanned vehicle operation status data recording and analysis system according to claim 2, characterized in that, Based on the decision geometry structure, the multidimensional state matrix is ​​analyzed to generate an event trigger sequence; Based on the time reference in the event trigger sequence, data slices within the event time neighborhood are extracted from the multidimensional state matrix to construct a spatiotemporal data cube, including: Based on the decision geometry and the multidimensional state matrix, each state vector of the multidimensional state matrix is ​​mapped to the decision geometry to obtain the geometric position coordinates corresponding to each state vector. Based on the sequence of changes in geometric position coordinates over time, the time points where abrupt changes in geometric position coordinates are detected, resulting in a set of event boundary points. Based on the set of event boundary points, the continuous timeline is divided into multiple time periods, and a unique event identifier is assigned to each time period to generate an event trigger sequence. Based on the start and end time references corresponding to each event identifier in the event triggering sequence, data rows covering the time windows before and after the event are extracted from the multidimensional state matrix to obtain event data slices; based on the event data slices, multiple data slices are aligned and integrated along the time dimension to construct a spatiotemporal data cube.

4. The low-speed unmanned vehicle operation status data recording and analysis system according to claim 3, characterized in that, Feature extraction is performed on the spatiotemporal data cube to obtain event feature fingerprints and generate event feature codes, including: Based on the spatiotemporal data cube, statistical features are calculated along the time, space, and state dimensions to obtain a multi-dimensional statistical feature set. Based on the multi-dimensional statistical feature set, key feature subsets are selected through feature importance ranking and redundancy analysis to form event feature fingerprints. Based on the event feature fingerprint, each feature value in the key feature subset is standardized and discretized to generate an event feature code.

5. The low-speed unmanned vehicle operation status data recording and analysis system according to claim 4, characterized in that, By inputting event feature encodings into a pre-trained deep belief network, the event risk level assessment results are obtained, including: Based on the event feature encoding, the numerical sequence of the event feature encoding is input into the input layer of the pre-trained deep belief network; based on the event feature encoding numerical sequence received by the input layer, the feature activation state of the first hidden layer is obtained through the operation of the first hidden layer. Based on the feature activation states of the first hidden layer, the high-level feature states of the second hidden layer are obtained through the operation of the second hidden layer; based on the high-level feature states of the second hidden layer, the probability distribution of the event corresponding to multiple risk levels is obtained through the operation of the output layer. Based on the probability distribution of an event corresponding to multiple risk levels, the risk level corresponding to the highest probability value is selected to obtain the event risk level assessment result.

6. The low-speed unmanned vehicle operation status data recording and analysis system according to claim 5, characterized in that, The event trigger sequence, spatiotemporal data cube, event feature encoding, and event risk level assessment results are integrated to generate a structured event analysis report, including: Based on the event trigger sequence, spatiotemporal data cube, event feature encoding, and event risk level assessment results, the event's time stamp, spatial range, feature vector, and risk level are extracted to obtain a set of key event elements. Based on the set of key event elements, a standardized data structure is organized to obtain an initialized event reporting framework; based on the time correspondence between the event trigger sequence and the spatiotemporal data cube, the state change data of the spatiotemporal data cube is filled into the corresponding data field of the initialized event reporting framework to obtain an event reporting framework containing state data. Based on the numerical values ​​of the event feature codes and the labels of the event risk level assessment results, the feature code values ​​and risk level labels are filled into the feature and conclusion fields of the event reporting framework containing status data to obtain a fully filled event reporting framework. Based on the fully populated event report framework, perform formatting transformation to generate a structured event analysis report.

7. A computing device, characterized in that, include: One or more processors; A storage device for storing one or more programs, which, when executed by one or more processors, cause the one or more processors to implement the system as described in any one of claims 1 to 6.

8. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a program that, when executed by a processor, implements the system as described in any one of claims 1 to 6.

Images