Over-the-horizon unmanned aerial vehicle flight control method and system

Abstract

The invention relates to the technical field of course control, in particular to a beyond-visual-range unmanned aerial vehicle flight control method and system.The method includes the steps that course included angle continuity judgment is conducted based on task preset nodes in the initial stage, connecting line segments shielded by obstacles in a path are screened out, and node primitives capable of being used for continuous steering are formed; basic support is provided for follow-up path combination stability, multiple correlation quantities such as path serial numbers, steering included angles and side length tension are read through a graph convolutional network, the spatial relation and dynamic characteristics of all path segments in a graph structure are extracted, areas with angular velocity fluctuation risks in local path connection are identified, and the stability of the local path connection is improved. Processing the angular velocity difference value sequence of the adjacent nodes through a long short-term memory network, constructing time correlation characteristics, extracting the angular velocity change trend in the path segment, screening an optimal path combination according to amplitude stability, and constructing a course channel set with low fluctuation amplitude.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of heading control technology, in particular to a beyond-visual-range unmanned aerial vehicle flight control method and system. BACKGROUND

[0002] The field of heading control technology aims to achieve accurate flight of a flight vehicle along a predetermined flight path or target point by controlling the heading angle of the flight vehicle, to ensure that the flight vehicle has path tracking accuracy, flight stability and real-time direction control in various flight environments, to solve the problems of detection and correction of heading deviation in flight, heading keeping ability in dynamic environment, and heading adjustment strategy under various task constraints.

[0003] A beyond-visual-range unmanned aerial vehicle flight control method aims to ensure that the unmanned aerial vehicle can still complete the flight task stably, safely and intelligently under non-directly visible conditions of the communication link, and to introduce a module with autonomous navigation, autonomous path correction, obstacle avoidance decision and flight state evaluation in the flight control strategy, so that the unmanned aerial vehicle can complete heading control based on task parameters, flight environment and built-in models even when it loses real-time instructions from the control station, and is suitable for performing emergency medical material transportation tasks, including transporting automatic external defibrillators and cardiovascular emergency drugs to the location of a sudden patient.

[0004] The prior art has the problem of insufficient initial inspection of path continuity and turnability, and the turnability between path nodes is not fully screened, which leads to easy occurrence of heading jump in the scene of environmental obstruction or discontinuous node turning angle, causing unstable flight attitude, and the path control relies on static planning parameters of the task model and cannot correct the control path in real time according to the dynamic change of angular velocity, so that once angular velocity mutation or path interference occurs in the flight process, high-precision angular velocity adjustment and stable recovery cannot be achieved, and when the unmanned aerial vehicle loses the instruction link, the preset path may have multiple discontinuous nodes, if the turning residual error cannot be corrected in time, and the discrimination and replacement mechanism for abnormal points is not perfect, and there is a lack of effective heading avoidance strategy support, it is easy to cause instruction failure under the conditions of sudden disturbance or data error, which easily causes path deviation, task failure and safety risk increase. SUMMARY

[0005] The purpose of the present application is to solve the problems existing in the prior art and to provide a beyond-visual-range unmanned aerial vehicle flight control method and system.

[0006] In order to achieve the above purpose, the present application adopts the following technical scheme: a beyond-visual-range unmanned aerial vehicle flight control method, comprising the following steps: S1: According to the flight task setting parameters and the task planning scheme, the path node coordinates are extracted and numbered, it is judged whether the heading angle between nodes is continuous, the obstacle obstruction connecting line is screened out, and the continuous turnable heading connection graph element is established; S2: Based on the continuous steerable heading connection graph element, the connection order number, the steering angle value and the edge length tension value of each path sequence are read by using a graph convolution network, the sum of the angular velocity difference between adjacent nodes is calculated by a long short-term memory network, and path segment comparison analysis is performed, the path combination is selected according to the continuous amplitude fluctuation, and a stable heading channel configuration set is generated; S3: Based on the stable heading channel configuration set, the difference amplitude of the current node heading trend and the configuration angular velocity suggestion value is determined, the current steering residual sequence is constructed and the valley value is screened, the channel point with angular velocity mutation is marked and replaced, the steering smooth point in the same path level is selected to reconstruct the path channel, and the offset correction angular velocity replacement sequence is obtained; S4: Based on the offset correction angular velocity replacement sequence, it is judged whether the angular velocity value falls into the stable range, and the stable control instruction is replaced for the items that do not meet the requirements, and a stable state control angular velocity instruction sequence is generated; S5: Based on the stable state control angular velocity instruction sequence, the angular velocity change critical value is compared with the historical heading error, the abnormal instruction point is identified, the low-risk path instruction is replaced, and a heading avoidance correction control path sequence is generated.

[0007] As a further scheme of the application, the specific steps for generating the continuous steerable heading connection graph element are: According to the flight task setting parameters and the task planning plan, the path information is read and the corresponding path node coordinates are extracted, the node sequence number and time stamp are used to establish a sequential index mapping, and the mapping table is sorted and segmented to generate a path numbered node set; Based on the path numbered node set, the heading angle of any adjacent node pair is calculated and compared with the standard deviation threshold value, the continuous heading difference interval is determined by using a sliding window method, and the out-of-limit node pairs are marked to obtain a continuous direction node set; Based on the continuous direction node set, the intersection of the node connection and the obstacle region coordinate group is judged, the end point coordinates and the boundary polygon intersection calculation method are used to exclude the shielding of the connected line, and the continuous steerable heading connection graph element is established.

[0008] As a further scheme of the application, the specific steps for generating the stable heading channel configuration set are: Based on the continuous steerable heading connection graph element, the connection structure between nodes is aggregated and encoded by using a graph convolution network, and the corresponding index matching in the sequential number of each connection segment is performed, the steering angle value and the edge length ranging value in the adjacent node group are extracted, the node index associated path segment is used for parameter collection, and a path structure data set is obtained; Based on the path structure data set, a long short-term memory network is used to perform time series learning and trend modeling on the structure feature sequence, to perform angle velocity change value square difference accumulation between all path segments and record the number index, to use the accumulated value to sort and judge the fluctuation trend, to filter and sort according to the path segment with the slowest angle velocity amplitude growth, and to generate a path fluctuation analysis sequence; Based on the path fluctuation analysis sequence, the continuity of the number index is searched and the consistency of the path structure is compared, the path segment group is selected according to the continuous number requirement and the fluctuation value, the corresponding number segment is extracted as the main structure integration, and a stable heading channel configuration set is generated.

[0009] As a further scheme of the application, the graph convolution network is according to the formula:

[0010] Wherein: represents the node in the unmanned aerial vehicle flight path graph The graph convolution embedding feature representation of the first layer, represents the node The embedding feature of the first layer, represents the weight parameter matrix of the graph convolution network in the first layer, represents the degree value of the node In the path graph, represents the degree value of the node In the path graph represents the adjacent node set of the node , represents a nonlinear activation function for enhancing feature expression capability, represents the turning angle normalization value of the connected edge between the node And the node , represents the Euclidean distance normalization value of the edge segment , represents the position matching weight of the connected segment in the flight path sequence.

[0011] As a further scheme of the application, the long short-term memory network is according to the formula:

[0012] Wherein: represents the weighted mean square error index of the control parameter ω , represents the total number of time steps of the training sample sequence, represents the current time step index, represents the control error weight coefficient of the first t time step, an environmental risk weight factor representing a t a state uncertainty adjustment factor representing a t a true flight control amount representing a t a flight control amount predicted by a long short-term memory network at a t

[0013] As a further scheme of the present application, the specific step of generating the offset correction angular velocity replacement sequence is: Based on the stable heading channel configuration set, current node index extraction is performed and the difference between the adjacent angular velocity suggestion value and the current heading trend value is calculated, a difference sequence is established and each segment is sequentially archived to generate a steering residual trend sequence; Based on the steering residual trend sequence, second-order difference of the difference sequence is performed and a negative change interval is identified, extreme value positioning is performed on the descending interval of the sequence and its index sequence is extracted to obtain an abnormal steering node set; Based on the abnormal steering node set, the angular velocity variation amplitude of the same path level node is sorted, the nodes that do not appear dramatic jumps are selected and position replacement is performed against the abnormal node index to obtain the offset correction angular velocity replacement sequence.

[0014] As a further scheme of the present application, the specific step of generating the offset correction angular velocity replacement sequence is: The absolute value of the first derivative of the angular velocity is sorted in the same node set of the path level to which the abnormal node belongs, stable nodes with angular velocity variation amplitude lower than a set threshold are selected as candidate replacement sources, and the angular velocity value of the abnormal node is replaced with the angular velocity value of the selected stable node at the same index position, the position index is kept unchanged, the track position and timestamp information are not changed, and the control variable is updated.

[0015] As a further scheme of the present application, the specific step of generating the offset correction angular velocity replacement sequence is: Based on the offset correction angular velocity replacement sequence, interval judgment is performed on each angular velocity value and the preset stable interval boundary, threshold upper and lower limits are set and comparison is made on whether the current value is out of range, all positions that do not meet the interval condition are identified to obtain an out-of-limit angular velocity index set;​​​​ Based on the super-limit angular velocity index set, the in-boundary value is extracted from the angular velocity variation table under the historical stable path, the stable replacement instruction sequence is generated by inserting the matching value into the index corresponding to the super-limit position and keeping the time stamp synchronous, and the stable replacement instruction sequence is generated. Based on the stable replacement instruction sequence, the time sequence splicing and the angular velocity difference continuity check of the whole segment instruction sequence are performed, the gradient smoothing processing of the change abrupt segment is performed, and the stable state control angular velocity instruction sequence is generated.

[0016] As a further scheme of the application, the specific steps for generating the heading avoidance correction control path sequence are as follows: Based on the stable state control angular velocity instruction sequence, the angular velocity difference of each instruction is calculated and compared with the angular velocity difference of the previous instruction, the change threshold is set and compared with the historical heading error curve at the corresponding time point, the angular velocity abnormal point index set is obtained; Based on the angular velocity abnormal point index set, the task number and path segment index of each abnormal point are extracted and the historical path control library is queried, the angular velocity record value of the same number and the same heading segment is matched and sorted to filter the item with the lowest jump amplitude, and the low-risk replacement instruction set is generated. Based on the low-risk replacement instruction set, each replacement instruction is inserted into the specified position of the original sequence and the instruction time sequence is re-established, the angular velocity interval between the adjacent instructions is judged and the large jump value is linearly adjusted, and the heading avoidance correction control path sequence is generated.

[0017] A super-range unmanned aerial vehicle flight control system, the super-range unmanned aerial vehicle flight control system is used for executing the super-range unmanned aerial vehicle flight control method, the system comprises: The path node generation module: based on the flight task setting parameters and the task planning plan, the path node coordinates are extracted, the sequence numbers of all nodes are set, the heading angle values between adjacent nodes are calculated, the continuity judgment operation is performed on the heading angle sequence, the node connection relationship with angle jump items is deleted, and the remaining path segments are judged for obstacle shielding and the unreachable items are removed, and the continuous node graph element set is generated. The graph structure reading module: based on the continuous node graph element set, the graph adjacency matrix structure set in the graph convolution network is called, the number sequence between each group of nodes, the path segment turning angle value and the segment length tension value are jointly read, combined into multiple groups of path segment numerical feature sequences, the multi-order aggregation mechanism in the graph convolution network is used to concentrate the connection weight between path segments, and the spatial relationship of all path segments is established into a standardized input matrix, and the graph structure weight parameter set is generated. Path sequence comparison module: based on the graph structure weight parameter set, all path segment nodes are grouped in order of numbering, the difference value of the angular velocity change value between each group of adjacent node pairs is calculated, an angular velocity sequence set is formed, the time state unit in the long short term memory network is called, the angular velocity sequence set is input in the time direction, the angular velocity difference value sum of the continuous nodes in the path segment is calculated through the memory state transmission mode, the amplitude fluctuation interval of each path segment is extracted, the angular velocity change trend in all path segments is compared, the combined paragraph with the amplitude change keeping the convergence state is screened out, and the heading stable path sequence is generated; Angular velocity control module: based on the heading stable path sequence, the current path segment angular velocity value and the heading change trend value are obtained, the difference value of the two is calculated item by item, a difference amplitude list is generated, a turning residual sequence is constructed, and the valley value node is screened out, the path segment marking operation is performed at the residual mutation place, the candidate segment with low amplitude value in the same layer path is selected, the corresponding node is inserted into the original path segment position, the path is reconstructed, the angular velocity sequence is recalculated, the stability of the sequence is judged and all angular velocity items exceeding the set stability threshold are replaced, and the heading control replacement sequence is generated; Avoidance path construction module: based on the heading control replacement sequence, the amplitude difference value of each angular velocity change point is extracted and compared with the historical heading error value, the path node corresponding to the angular velocity change jump item is identified item by item, the replacement instruction query is executed and the low-risk instruction node located in the same path layer is extracted, the angular velocity instruction value of each replacement node is combined and an updated path instruction structure is established, and an avoidance correction control channel is generated.

[0018] Compared with the prior art, the advantages and positive effects of the present application are: In the present application, the heading angle continuity is judged based on the task preset node in the initial stage, and the connection line segment blocked by the obstacle in the path is screened out, forming a node graph element that can be used for continuous turning, providing basic support for the subsequent path combination stability; In the present application, the path sequence number, turning angle and edge length tension and other related quantities are read by the graph convolution network, the spatial relationship and dynamic characteristics of each path segment in the graph structure are extracted, and the area with angular velocity fluctuation risk in the local path connection is identified; In the present application, the long short term memory network is used to process the angular velocity difference value sequence of adjacent nodes, construct time correlation characteristics, extract the angular velocity change trend in the path segment, and select the optimal path combination according to the amplitude stability, construct a set of heading channel with low fluctuation amplitude, compare the current node heading trend with the recommended angular velocity, determine the difference amplitude, construct the turning residual sequence and identify the mutation point, and realize path reconstruction through marking replacement. BRIEF DESCRIPTION OF DRAWINGS

[0019] Figure 1 The workflow schematic diagram of the present application; Figure 2 System flowchart of the present application. DETAILED DESCRIPTION

[0020] In order to make the objects, technical solutions and advantages of the present application clearer, the present application will be further described in detail below with reference to the drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present application and not to limit the present application.

[0021] In the description of the present application, it should be understood that the terms "length", "width", "upper", "lower", "front", "rear", "left", "right", "vertical", "horizontal", "top", "bottom", "inner", "outer" and the like indicate the orientation or positional relationship based on the orientation or positional relationship shown in the drawings, and are only for the convenience of describing the present application and simplifying the description, and therefore cannot be understood as indicating or implying that the device or element referred to must have a particular orientation, be constructed and operated in a particular orientation, and therefore cannot be understood as limiting the present application. In addition, in the description of the present application, the meaning of "a plurality of" is two or more, unless otherwise explicitly and specifically limited.

[0022] Embodiment one Please refer to Figure 1 The present application provides a technical solution: a method for flying control of an over-the-horizon unmanned aerial vehicle, comprising the following steps: S1: According to the flight task setting parameters and task planning scheme, the path node coordinates are extracted and numbered, it is judged whether the heading angle between nodes is continuous, the obstacle blocking connection is screened out, and the continuous steerable heading connection graph element is established; S2: Based on the continuous steerable heading connection graph element, the connection order number, steering angle value and edge length tension value of each path sequence are read by using graph convolution network, the sum of angular velocity difference between adjacent nodes is calculated by long short-term memory network and path segment comparison analysis is performed, the path combination is selected according to the continuous amplitude fluctuation, and the stable heading channel configuration set is generated; S3: Based on the stable heading channel configuration set, the difference amplitude of the current node heading trend and the configuration angular velocity suggestion value is determined, the current steering residual sequence is constructed and the valley value is screened out, the channel points with angular velocity mutation are marked and replaced, the steering smooth points within the same path level are selected to reconstruct the path channel, and the offset correction angular velocity replacement sequence is obtained; S4: Based on the offset correction angular velocity replacement sequence, it is judged whether the angular velocity value falls within the stable range, and the items that do not meet the requirements are replaced using stable control instructions to generate a stable state control angular velocity instruction sequence; S5: Based on the stable state control angular velocity instruction sequence, the angular velocity change critical value and the historical heading error are compared, the abnormal instruction points are identified and the low-risk path instruction is replaced, and the heading avoidance correction control path sequence is generated.

[0023] The specific steps for generating the continuous steerable heading connection graph element are: According to the flight task setting parameters and the task planning plan, the flight segment information is read and the corresponding path node coordinates are extracted, the node sequence number and the time stamp are used to establish a sequential index mapping, and the mapping table is sorted and segmented to generate a path number node set; Based on the path number node set, the heading angle of any adjacent node pair is calculated and compared with the standard deviation threshold value, the sliding window method is used to determine the continuous heading difference interval and mark the out-of-limit node pair, and the continuous direction node set is obtained; Based on the continuous direction node set, the intersection between the connecting line between the nodes and the obstacle region coordinate group is judged, the endpoint coordinates and the boundary polygon intersection calculation method are used to screen out the shielding condition of the connecting line, and the continuous steerable heading connection graph element is established; Based on the flight task setting parameters and the task planning plan, the geographical space information coordinate extraction method is used to match the flight segments set in the task, read the latitude and longitude coordinate group in the corresponding flight segment, perform floating point number normalization processing on the coordinate values and execute two-dimensional array organization structure conversion operation, and use the sequential index mapping to establish the corresponding relationship between the node sequence number and the flight time stamp. The function input parameters are the node number list, the time stamp list and the sorting rule identifier, the built-in key value pair construction command in the function body is called, the index table is generated from the number and the time stamp, the index table is sorted in ascending order according to the time stamp field, the quick sorting algorithm is used in the sorting process and the sorting median selection mechanism is called, after completion, the sorting result is segmented, the segmentation is based on the interval difference of the time stamp exceeding the set threshold value, the set threshold value is 3 seconds, the paragraph cutting function is called to generate the number sub-set list, and the path number node set is generated. If the task type is a medical material transportation task, including AED and first aid medicine air drop or delivery operation, the high priority path scheduling parameter is enabled in the task planning plan and the path segment is marked as a high precision track tracking mode; Based on the path number node set, the heading angle calculation method is used to perform coordinate difference calculation operation on the coordinate values of any two adjacent nodes in the set, the calculation function is called, the longitude and latitude values of the front and rear nodes are used as the embedded parameters, the coordinate difference inverse tangent function is executed to obtain the angle value, and the result is saved in the form of floating point number with four decimal places, an angle sequence array is constructed, the standard deviation calculation function is called to extract the mean and standard deviation of the angle sequence, the deviation threshold is set to 3.5 degrees, the absolute value of the difference between each group of node angle values and the mean value is calculated, and the difference ratio is judged with the standard deviation threshold, the judgment formula is: if the absolute difference value divided by the standard deviation is greater than 1, it is an abnormal angle pair, the corresponding node pair is marked as abnormal, and then the sliding window detection method is used to perform a continuous heading difference interval judgment operation with a window width of 5, if there are more than two abnormal angle marker pairs in the sliding window, all node pairs in the whole window are set as direction abnormal, finally the unmarked node pairs are extracted to form a continuous direction node set, if the detection task is medical material transportation path, the system will enable the flight buffer adjustment mechanism for the paragraph containing the abnormal angle, and weight the heading stability score value in the sliding window to improve the stability and load protection capability in the flight process; Based on the continuous direction node set, a polygon occlusion detection method is used to perform occlusion judgment on the coordinates of the connecting lines between each group of nodes in the set and the obstacle region, a line segment and polygon boundary intersection function is called, the input parameters are the starting point and ending point longitude and latitude of the node connecting line and the polygon vertex array of the obstacle region, an intersection point calculation command is executed, each boundary edge segment and the current connecting line are traversed to judge the cross product direction, if there is an intersection point, the connecting line is marked as an occlusion, then an endpoint coordinate attribution judgment function is executed to judge whether the endpoints are within the polygon boundary, if true, it is marked as an occlusion line segment at the same time, all unmarked connecting line node pairs are extracted as passable connections, a node connection index graph structure is established, and a continuous turnable heading connection graph element is generated, wherein if the task identification is to transport medical equipment or emergency medicine, the occlusion detection module will preferentially exclude risk zones and prohibit the planning of paths that pass through building-intensive areas or complex occlusion areas, and automatically adjust the heading reconstruction strategy in the connection graph element to enhance the continuity, safety and timeliness reliability of the delivery path.

[0024] The specific steps for generating a stable heading passage configuration set are: Based on the continuous turnable heading connection graph element, a graph convolution network is used to aggregate and encode the connection structure between nodes, and perform corresponding index matching of each connection segment in the order number, extract the turning angle value and edge length ranging value in the adjacent node group, use the node index associated path segment to collect parameters, and obtain a path structure data set; Based on the path structure data set, the long short-term memory network is used for time series learning and trend modeling of the structure feature sequence, the square difference of the angular velocity change value between all path segments is accumulated and the number index is recorded, the accumulated value is sorted to judge the fluctuation trend, the path segment with the slowest increase in angular velocity amplitude is selected and sorted, and the path fluctuation analysis sequence is generated; Based on the path fluctuation analysis sequence, the continuity of the number index is searched and the consistency of the path structure is compared, the path segment group is selected according to the continuous number requirement and the fluctuation value, the corresponding number segment is extracted as the main structure integration, and the stable heading channel configuration set is generated; Based on the continuous steerable heading connection graph element, the graph convolution network is used to aggregate and encode the connection structure between nodes, the node number matrix and the connection relationship matrix are called as the initial input by calling the graph structure input layer, the number matrix is a number vector of each node, the connection matrix is in the form of an adjacency matrix, the propagation function used by the graph convolution propagation layer is a linear transformation after ReLU activation plus adjacency matrix multiplication, the weight matrix is initialized as a Gaussian distribution with mean 0 and variance 0.01, the input feature dimension is 128, the output feature dimension is 64, two layers of propagation structure are used, L2 regularization operation is performed on the output of each layer, the node embedding result is index rearranged, the graph convolution result is output to the path segment number index table, the node index of each group of connected segments is matched according to the path segment sequence number, the steering angle value and the edge length ranging value in the adjacent node group are extracted, the steering angle value is converted to float type with five decimal places after being calculated in radians, the edge length ranging value is calculated in Euclidean distance and is kept in meters with three decimal places, the obtained node index, angle value and edge length ranging value are grouped into a structured record according to the path segment number, and the path structure data set is generated; Based on the path structure data set, the long short-term memory network is used for time series learning and trend modeling of the structure feature sequence, the input layer of the long short-term memory network structure is called to organize each group of path segment structure feature vectors into time input sequences in order of number, each input vector includes the steering angle value, the edge length ranging value and the connection weight coding of the corresponding segment, the time step of the long short-term memory network unit is set to 10, the dimension of the hidden layer is set to 64, the activation function uses tanh, all sequences are forward propagated, the predicted angular velocity value trajectory of each path segment in time order is output, the predicted angular velocity value between adjacent path segments is processed by item difference and square, the angular velocity change value square difference of all path segments is recorded according to the number index, the number index of all path segments and the corresponding accumulated square difference value form a binary tuple set, the set is sorted in ascending order and the path segment index sequence with the smallest angular velocity fluctuation value is extracted, and the path fluctuation analysis sequence is generated; Based on the path fluctuation analysis sequence, the numbered index continuity search and path structure consistency comparison are performed, the path segment numbers in the fluctuation sequence are sequentially scanned, the step difference of any continuous number group is judged, whether the continuous step length is 1 is judged, the number group meeting the condition is recorded, and the structure characteristics of all path segments in the continuous number group are checked for consistency, the change amplitude of the turning angle between nodes and the change ratio of the edge length ranging value are compared, the change amplitude threshold of the turning angle is set to 0.02 radian, the ranging ratio change threshold is 0.1, the path segment group meeting the double threshold constraints is screened and extracted, the corresponding numbered segment is taken as the main structure segment to construct the integrated path channel, and the stable heading channel configuration set is generated.

[0025] The graph convolution network is according to the formula:

[0026] Wherein: represents the node in the unmanned aerial vehicle flight path graph The graph convolution embedding feature representation of the layer represents the node The embedding feature of the layer represents the weight parameter matrix of the graph convolution network at the layer represents the degree value of the node in the path graph represents the degree value of the node in the path graph represents the adjacent node set of the node represents a nonlinear activation function for enhancing feature expression ability represents the turning angle normalized value of the connected edge between the node and the node represents the Euclidean distance normalized value of the edge segment represents the position matching weight of the connected segment in the flight path sequence Execution process: first, based on the adjacent node set of any node in the flight path graph, the embedding feature of each adjacent node is extracted, then the turning angle of the connected edge segment between the node and the node is calculated, the standardization processing is performed to obtain the parameter , which is used to reflect the influence of turning intensity on flight path structure, the Euclidean distance between nodes is calculated and normalized to the parameter ​​​​​This describes the effect of path segment length on control accuracy and generates sequence matching weights based on the index position of the edge segment in the overall path sequence. Attenuation factor The standard deviation of the total path index difference is dynamically set to enhance the role of path sequence information in feature propagation, and then the node degree value is calculated. and Used to normalize adjacency relationships, thereby adjusting the aforementioned weight parameters. With weight matrix and embedded features Substitute these into the formula to complete the adjacency feature aggregation and use a nonlinear function. Output the first Layer embedding results The node encoding values, which integrate flight direction, path length, and sequence structure information, are obtained and used for decision support and trajectory optimization of subsequent UAV control paths.

[0027] Long Short-Term Memory (LSTM) networks, according to the formula:

[0028] in: Indicates control parameters ω The weighted mean square error index, This represents the total number of time steps in the training sample sequence. Indicates the index of the current time step. Indicates the first t The control error weighting coefficient at each time step Indicates the first t Environmental risk weighting factors at each time step Indicates the first t Adjustment factor for state uncertainty at each time step This represents the actual flight control parameters at time step t. This indicates that the Long Short-Term Memory network is in the first... t The flight control parameters predicted at each time step; Execution process: First, using the historical state sequence as input, predict the control commands for each future moment. and compared with the actual control commands recorded during the actual flight. To enhance the environmental sensitivity and task adaptability of error evaluation, time-by-time comparisons were performed, and three dynamic weighting factors were introduced. 、 、 ,in, Used to enhance error weights at critical flight nodes. Weights are assigned based on risk levels such as terrain complexity and communication blind spots. Then, considering the uncertainties caused by state estimation errors or channel instability, at each time step... Calculate the weighted error term and throughout the flight time The average value is calculated to obtain the global WMSE index, which not only reflects the long short-term memory network's ability to fit control commands, but also provides a quantitative basis for subsequent model structure optimization and flight strategy updates.

[0029] The specific steps for generating the offset correction angular velocity replacement sequence are as follows: Based on the stable heading channel configuration set, the current node index is extracted and the difference between the adjacent angular velocity suggested value and the current heading trend value is calculated. By establishing the difference sequence and performing sequential archiving of each segment, the steering residual trend sequence is generated. Based on the trend sequence of turning residuals, the second difference of the difference sequence is performed and the negative change interval is identified. The extreme value location is performed on the descending interval of the sequence and its index sequence is extracted to obtain the set of abnormal turning nodes. Based on the set of abnormal turning nodes, the nodes at the same path level are sorted by the magnitude of angular velocity change. By selecting nodes that do not show drastic changes and performing position replacements according to the abnormal node index, the offset correction angular velocity replacement sequence is obtained. Based on a stable heading channel configuration set, an interpolation sliding calculation method is used to extract the current node index and calculate the difference between the heading change trend value corresponding to the current node and the angular velocity suggestion value in the adjacent path segment. The interpolation generation function is called with the node index, trend value array, and suggestion value array as input parameters. The interpolation operation is performed to subtract the trend value and suggestion value under the corresponding index. The interpolation result is retained to six decimal places and stored in the interpolation sequence. After the sequence is established, it is segmented and archived in ascending order of node number. The archiving operation calls the sequence segment division function to slice the entire interpolation sequence according to the set segmentation step size of 5. The slicing result is organized into a structured time series matrix to generate the steering residual trend sequence. Based on the trend sequence of turning residuals, a second-order difference identification method is used to analyze the fluctuation trend of the difference sequence. The difference function is executed on each segment of the sequence. The difference function is executed by subtracting the previous term from the current term and then subtracting the difference between the two previous terms to form a new difference value array. The direction judgment function is called to scan the difference value array and mark the continuous segments with values ​​less than zero as negative change intervals. Then, the extreme value location operation is performed on all difference values ​​in the marked interval. The extreme value search function is called to perform the minimum value search operation on each continuous interval, extract the original index position of the difference sequence where the corresponding minimum value is located, and construct an index set to generate an abnormal turning node set. Based on the abnormal turning node set, the jump amplitude sorting method is adopted, the path level grouping function is called to divide the current all path segment nodes according to the path level, the angular velocity value set of all nodes in each path layer is extracted, the jump amplitude calculation function is called to calculate the absolute difference value of any adjacent two values in the angular velocity sequence, the difference value array is sorted from small to large to generate the jump amplitude list, and the node index with the difference value greater than the set threshold is selected from the list, the angular velocity jump threshold is set to 0.3 rad / s, the selected stable node index is replaced with the index in the abnormal turning node set, and the angular velocity array is reorganized after the replacement is completed to generate the offset correction angular velocity replacement sequence.

[0030] By establishing the difference sequence and sequentially archiving each segment, a fixed length window is set to slide and segment the difference sequence, each segment represents the angular velocity deviation evolution interval in the continuous heading adjustment process, and in each segment, the difference change amount, average change rate and change direction identifier of the starting and ending points are calculated to form a standardized structure body for archiving; By selecting the node without sharp jump and performing position replacement by comparing the abnormal node index, in the same node set of the abnormal node belonging to the path level, the angular velocity first derivative absolute value is sorted, the stable node with angular velocity change amplitude lower than the set threshold is selected as the candidate replacement source, and the angular velocity value of the abnormal node is replaced with the angular velocity value of the selected stable node at the same index position, the position index is kept unchanged, the track position and timestamp information are not changed, and the control variable is updated.

[0031] The specific steps of generating the steady-state control angular velocity instruction sequence are as follows: Based on the offset correction angular velocity replacement sequence, interval judgment is performed on each angular velocity value and the preset stable interval boundary, whether the current value is out of range is compared by setting the upper and lower limits of the threshold, all positions not meeting the interval condition are identified, and the out-of-limit angular velocity index set is obtained; Based on the out-of-limit angular velocity index set, the in-boundary value is extracted by matching the value in the angular velocity variation table of the historical stable path, the matching value is inserted into the index corresponding to the out-of-limit position, the timestamp is kept synchronous, and the stable replacement instruction sequence is generated; Based on the stable replacement instruction sequence, the time sequence splicing of the full segment instruction sequence and the angular velocity difference continuity check between instructions are performed, the gradient smoothing processing is performed on the change abrupt segment, and the steady-state control angular velocity instruction sequence is generated; Based on the offset correction angular velocity replacement sequence, interval judgment method is adopted for boundary comparison operation on each angular velocity value, the numerical interval judgment function is called, the input parameters are angular velocity value sequence, lower limit value, upper limit value and logical judgment condition, the lower limit of the angular velocity stable interval is set to 0.4 rad / s, the upper limit is 0.4 rad / s, the call judgment condition is that the current value is less than the lower limit or greater than the upper limit, the angular velocity value is scanned item by item when the judgment operation is executed, and the range verification is performed by applying the logical comparison operator, the index extraction operation is performed on all positions that do not meet the interval condition, the index number corresponding to the out-of-range item is reserved as an integer type list, and an out-of-range angular velocity index set is generated; Based on the out-of-range angular velocity index set, the interval replacement method is used to check the angular velocity value table recorded in the historical path instruction library, the stable path matching function is called with the out-of-range index number and the historical angular velocity data matrix as input parameters, each row of the matrix records the historical task number, timestamp, angular velocity value and path number, the matching function uses the path number and time adjacent condition as a filter to filter out all records with angular velocity values falling within the set interval, the filtered angular velocity values are sorted in ascending order by timestamp and the value sequence is extracted, each matching value is inserted into the corresponding out-of-range angular velocity position in order of index number, and the timestamp synchronization function is called to keep the inserted value time field consistent with the original angular velocity value format, generating a stable replacement instruction sequence; Based on the stable replacement instruction sequence, the instruction timing splicing and gradient processing method is used to splice all instructions, the splicing function input parameters are stable instruction array, original sequence timestamp list and splicing mode identifier, the execution mode is to splice in ascending order of timestamp field, the difference calculation function is called for the spliced angular velocity value sequence to extract the angular velocity difference of adjacent two instruction items and construct a difference array, the gradient smoothing processing is performed on the paragraphs in the difference array with a change amplitude greater than 0.2 rad / s, the smoothing processing uses local linear interpolation, the input is start value, end value and step length, the changed segment is reconstructed as a linear gradient segment to replace the original mutation segment, generating a stable control angular velocity instruction sequence.

[0032] The specific steps of generating the heading avoidance correction control path sequence are as follows: Based on the stable control angular velocity instruction sequence, the angular velocity difference of each instruction is calculated and compared with the angular velocity of the previous instruction, the change threshold is set and compared with the historical heading error curve at the corresponding time point to obtain an angular velocity abnormal point index set; Based on the angular velocity abnormal point index set, the task number and path segment index of each abnormal point are extracted and the historical path control library is queried, the angular velocity record value of the same number and the same heading segment is matched and sorted to filter out the item with the lowest jump amplitude, and a low-risk replacement instruction set is generated; Based on the low-risk replacement instruction set, each replacement instruction is inserted into the specified position of the original sequence and a new instruction time sequence is established, the angular velocity interval between the adjacent instructions is judged and the linear adjustment processing is performed on the large jump value to generate a heading avoidance correction control path sequence; Based on the steady-state control angular velocity instruction sequence, the difference comparison and error retrieval method is used to calculate the difference value of each angular velocity instruction and the previous instruction in the sequence, and the difference value generation function is called with the current angular velocity value, the previous angular velocity value and the current timestamp as input parameters. The difference calculation result is output as a floating point type with four decimal places. The difference amplitude sequence is constructed. All difference values in the sequence are compared with the set change threshold value. The threshold value is set to 0.35 rad / s. At the same time, the historical error curve comparison function is called to locate the corresponding value point in the error curve at the current timestamp. The two-way difference synchronization judgment is executed. The position index that meets the angular velocity change threshold value and has high fluctuation record in the error curve is filtered out. The angular velocity abnormal point index set is generated. Based on the angular velocity abnormal point index set, the number path screening method is used to extract the task number and path segment index associated with each abnormal point. The historical path library task matching function is called with the abnormal task number and path segment direction label as input parameters. The angular velocity records of the same number task and consistent direction vector in the record table are screened. For medical material transportation tasks, including AED equipment and various cardiovascular emergency medicine air delivery tasks, the path risk level judgment identifier is added, and the path segment where such tasks are located is preferentially marked as a high steady-state demand path segment. The angular velocity field in all records is extracted and the jump amplitude sorting function is called. The function uses the absolute value of the angular velocity difference between adjacent records as the sorting indicator. The screening results are sorted from small to large according to the jump amplitude. The top 10% of angular velocity records are extracted as the candidate set. The conservative screening strategy is enhanced for records belonging to medical material task numbers. The angular velocity segments with high historical frequency fluctuations are removed to ensure the stability of the replacement angular velocity. The lowest amplitude records corresponding to each abnormal point are combined to generate a low-risk replacement instruction set. Based on the low-risk replacement instruction set, the insertion replacement and gradient processing method is used. Each replacement instruction is inserted into the original sequence at the abnormal index position through the instruction insertion function. The timestamp reordering function is called to sort the updated sequence in ascending order of timestamp. The time sequence list is reconstructed and the complete angular velocity instruction set is generated. The angular velocity interval judgment function is used to calculate the angular velocity difference between any adjacent instructions. The jump section greater than 0.3 rad / s is extracted. The linear interpolation processing function is called with the start value, end value and intermediate index position list of the jump section to reconstruct and replace the original value of the jump section. When processing medical transportation task tracks, the interpolation function introduces a smoothing factor weight to automatically reduce the jump amplitude at the end of the interpolation section to avoid the interference of heading mutation on delivery stability and ensure the attitude and trajectory stability of AED and emergency medicine before reaching the target location. The heading avoidance correction control path sequence is generated.

[0033] Please refer to Figure 2The application discloses an over-the-horizon unmanned aerial vehicle flight control system, and relates to the field of over-the-horizon unmanned aerial vehicle flight control. The path node generation module is used for extracting path node coordinates, setting sequential numbers for all nodes, calculating the heading angle values between adjacent nodes, performing a continuity judgment operation on the heading angle sequence, deleting the node connection relationship with angle jump items, performing obstacle shielding judgment and eliminating unreachable items on the remaining path segments, and generating a continuous node graph element set based on the flight task setting parameters and the task planning plan. The graph structure reading module is used for performing joint reading actions on the numbered sequence between each group of nodes, the path segment turning angle values and the segment length tension values, combining the path segment numerical feature sequences, centrally reading the connection weights between the path segments by using the multi-order aggregation mechanism in the graph convolution network, establishing the spatial relationship of all path segments into a standardized input matrix, and generating a graph structure weight parameter set based on the graph structure weight parameter set. The path sequence comparison module is used for grouping all path segment nodes according to the numbering sequence, performing difference calculation on the angular velocity change values between each group of adjacent node pairs to form an angular velocity sequence set, inputting the angular velocity sequence set into the time state unit in the long short-term memory network in the time direction, calculating the angular velocity difference sum of the continuous nodes in the path segment by using the memory state transmission mode, extracting the amplitude fluctuation interval of each path segment, comparing the angular velocity change trend in all path segments, screening out the combined paragraphs with the amplitude change keeping in a convergent state, and generating a heading stable path sequence. The angular velocity control module is used for obtaining the current path segment angular velocity value and the heading change trend value based on the heading stable path sequence, performing item-by-item calculation on the difference between the two values, generating a difference amplitude list, screening the valley nodes from the turning residual sequence, performing path segment marking operation at the residual mutation position, selecting the candidate segments with low amplitude values in the same layer path, inserting the corresponding nodes into the original path segment position, recalculating the angular velocity sequence after reconstructing the path, judging the stability of the sequence and replacing all angular velocity items exceeding the set stable threshold, and generating a heading control replacement sequence. The avoidance path construction module is used for extracting the amplitude difference values of the angular velocity change points and performing comparison operation with the historical heading error values, identifying the path nodes corresponding to the angular velocity change jump items item by item based on the heading control replacement sequence, performing replacement instruction query and extracting low-risk instruction nodes in the same path layer, combining the angular velocity instruction values of the replacement nodes and establishing an updated path instruction structure, and generating an avoidance correction control channel.

[0034] The above merely describes the preferred embodiments of the present application, and is not intended to limit the present application in other forms. Any skilled person in the art can modify or change the disclosed technical content into equivalent embodiments with equivalent changes, and apply them to other fields. However, any simple modification, equivalent change and modification made to the above embodiments according to the technical essence of the present application, without departing from the technical solution content of the present application, still falls within the protection scope of the present application.

Claims

1. A method for controlling the flight of a beyond-visual-range unmanned aerial vehicle (UAV), characterized in that, Includes the following steps: S1: Based on the flight mission setting parameters and planning scheme, extract the coordinates of the path nodes and number them, determine the continuity of the heading angle between nodes, remove the connecting lines that are blocked by obstacles, and construct a continuous and steerable heading connection primitive. S2: Based on the connected primitives, a graph convolutional network is used to extract the connection order, turning angle and side length tension value of the path sequence. The sum of the angular velocity difference between adjacent nodes is calculated through a long short-term memory network. The stability of the path segment is analyzed. Based on the fluctuation amplitude, a suitable path combination is selected to generate a stable heading channel configuration set. S3: Based on the configuration set, determine the difference between the current node's heading trend and the suggested angular velocity, construct a steering residual sequence and filter out valley values, mark angular velocity mutation points, select steering stability points at the same level for path channel reconstruction, and generate an offset correction angular velocity replacement sequence. S4: Determine whether the angular velocity values ​​in the replacement sequence are within a stable range, and replace the out-of-limit values ​​with control commands to form a steady-state control angular velocity command sequence; S5: Compare the critical change points of angular velocity in the instruction sequence with the historical heading errors, identify abnormal points and replace them with low-risk path instructions, and generate a heading avoidance correction control path sequence.

2. The beyond-visual-range UAV flight control method according to claim 1, characterized in that, The specific steps for generating the continuous steerable heading connection elements are as follows: Based on the flight mission settings and mission planning plan, the flight segment information is read and the corresponding path node coordinates are extracted. A sequential index mapping is established using node number and timestamp, and the mapping table is sorted and segmented to generate a set of path number nodes. Based on the path number node set, the heading angle between any two adjacent node pairs is calculated and compared with the standard deviation threshold. The continuous heading difference interval is determined by a sliding window method and the node pairs exceeding the limit are marked to obtain the continuous direction node set. Based on the continuous directional node set, the intersection of the connection between the nodes and the coordinate set of the established obstacle area is determined. The occlusion of the connecting line is screened out by using the intersection calculation method of endpoint coordinates and boundary polygons, and a continuous steerable heading connection primitive is established.

3. The beyond-visual-range UAV flight control method according to claim 1, characterized in that, The specific steps for generating the stable heading channel configuration set are as follows: Based on the continuous steerable heading connection primitives, a graph convolutional network is used to aggregate and encode the connection structure between nodes, and to match the corresponding indexes of each connection segment in the sequential numbering. The turning angle value and side length measurement value in the adjacent node group are extracted, and the path segments are associated with the node index to collect parameters and obtain the path structure data set. Based on the path structure data set, a long short-term memory network is used to perform time series learning and trend modeling on the structural feature sequence. The squared differences of angular velocity changes between all path segments are accumulated and recorded with index numbers. The fluctuation trend is judged by sorting the accumulated values. The path segments with the slowest increase in angular velocity amplitude are filtered and sorted to generate a path fluctuation analysis sequence. Based on the path fluctuation analysis sequence, the continuity of the number index and the consistency of the path structure are compared. According to the continuous numbering requirements and fluctuation values, path segment groups are selected, and the corresponding numbered segments are extracted as the main structure integration to generate a stable heading channel configuration set.

4. The beyond-visual-range UAV flight control method according to claim 3, characterized in that, The graph convolutional network is defined according to the formula: in: Nodes in the drone flight path graph In the Graph convolutional embedding feature representation of layers Represents a node In the Layer embedding features The graph convolutional network represents the first... The layer's weight parameter matrix, Represents a node Degree value in the path graph. Represents a node Degree value in the path graph Represents a node The set of adjacent nodes, This represents a non-linear activation function used to enhance feature representation capabilities. Represents a node With nodes The normalized value of the turning angle of the connected sides. Representing edge segment The normalized value of the Euclidean distance, This indicates the position matching weight of the connecting segment in the flight path sequence.

5. The beyond-visual-range UAV flight control method according to claim 3, characterized in that, The Long Short-Term Memory (LSTM) network is configured according to the formula: in: Indicates control parameters ω The weighted mean square error index, This represents the total number of time steps in the training sample sequence. Indicates the index of the current time step. Indicates the first t The control error weighting coefficient at each time step Indicates the first t Environmental risk weighting factors at each time step Indicates the first t Adjustment factor for state uncertainty at each time step Indicates the first t The actual flight control parameters at each time step. This indicates that the Long Short-Term Memory network is in the first... t The flight control parameters predicted at each time step.

6. The beyond-visual-range UAV flight control method according to claim 1, characterized in that, The specific steps for generating the offset correction angular velocity replacement sequence are as follows: Based on the stable heading channel configuration set, the current node index is extracted and the difference between the adjacent angular velocity suggested value and the current heading trend value is calculated. By establishing a difference sequence and performing sequential archiving of each segment, a steering residual trend sequence is generated. Based on the turning residual trend sequence, second-order difference of the difference sequence is performed and negative change intervals are identified. Extreme value localization is performed on the descending interval of the sequence and index sequence is extracted to obtain a set of abnormal turning nodes. Based on the set of abnormal turning nodes, the angular velocity variation amplitude of nodes at the same path level is sorted. By selecting nodes that do not show drastic changes and performing position replacements according to the abnormal node index, the offset correction angular velocity replacement sequence is obtained.

7. The beyond-visual-range UAV flight control method according to claim 5, characterized in that, The process involves establishing a difference sequence and sequentially archiving each segment. A fixed-length window is set to slide the difference sequence into segments, with each segment representing the evolution interval of angular velocity deviation during continuous heading adjustment. Within each segment, the difference change, average rate of change, and direction of change between the start and end points are calculated to form a standardized structure for archiving. The process involves selecting nodes that do not exhibit drastic changes and performing position replacements by comparing them with the index of abnormal nodes. Within the set of nodes at the same path level as the abnormal nodes, the nodes are sorted by the absolute value of the first derivative of their angular velocity. Stable nodes with angular velocity changes below a set threshold are selected as candidate replacement sources. The angular velocity values ​​of the abnormal nodes are then replaced with the angular velocity values ​​of the selected stable nodes at the same index position, keeping the position index unchanged and not altering the track position and timestamp information, while updating the control variables.

8. The beyond-visual-range UAV flight control method according to claim 1, characterized in that, The specific steps for generating the steady-state control angular velocity command sequence are as follows: Based on the offset correction angular velocity replacement sequence, each angular velocity value is judged against the preset stable interval boundary. By setting upper and lower limits of thresholds and comparing whether the current value exceeds the limit, all positions that do not meet the interval conditions are identified, and an over-limit angular velocity index set is obtained. Based on the set of over-limit angular velocity indexes, the boundary values ​​are extracted by referring to the table of angular velocity changes under historical stable paths. By inserting the matching values ​​into the index corresponding to the over-limit position and keeping the timestamp synchronized, a stable replacement instruction sequence is generated. Based on the stable replacement instruction sequence, the timing of the entire instruction sequence is spliced ​​and the continuity of the angular velocity difference between instructions is checked. Gradient smoothing is performed on the change segments to generate a steady-state control angular velocity instruction sequence.

9. The beyond-visual-range UAV flight control method according to claim 1, characterized in that, The specific steps for generating the aforementioned heading avoidance correction control path sequence are as follows: Based on the steady-state control angular velocity command sequence, the angular velocity difference of each command is calculated and compared with the angular velocity of the previous command. By setting a change critical threshold and comparing it with the corresponding time point of the historical heading error curve, an index set of angular velocity anomaly points is obtained. Based on the set of angular velocity anomaly points, the task number and path segment index of each anomaly point are extracted and the historical path control database is queried. By matching the angular velocity record values ​​of the same number and the same direction segment and sorting and filtering the items with the lowest jump amplitude, a low-risk alternative instruction set is generated. Based on the low-risk alternative instruction set, each alternative instruction is inserted into the specified position of the original sequence and the instruction time sequence is re-established. By judging the angular velocity interval between adjacent instructions and performing linear adjustment on the large jump value, a heading avoidance correction control path sequence is generated.

10. A beyond-visual-range unmanned aerial vehicle (UAV) flight control system, characterized in that, The beyond-visual-range UAV flight control method according to any one of claims 1-9, wherein the system comprises: Path node generation module: Based on flight mission setting parameters and mission planning plan, extract path node coordinates, set sequential numbers for all nodes, calculate the heading angle between adjacent nodes, perform continuity judgment operation on the heading angle sequence, delete the node connection relationship with angle jump item, and then perform obstacle occlusion judgment on the remaining path segment and remove unreachable items to generate a continuous node primitive set. Graph structure reading module: Based on the continuous node primitive set, it calls the graph adjacency matrix structure set in the graph convolutional network to perform joint reading actions on the number sequence between each group of nodes, the turning angle value of the path segment, and the segment length tension value, and combines them into multiple groups of path segment numerical feature sequences. It uses the multi-order aggregation mechanism in the graph convolutional network to centrally read the connection weights between path segments and establish the spatial relationship of all path segments into a standardized input matrix to generate a graph structure weight parameter set. Path sequence comparison module: Based on the graph structure weight parameter set, all path segment nodes are grouped in numerical order. The difference between the angular velocity changes of adjacent node pairs in each group is calculated to form an angular velocity sequence set. The time state unit in the long short-term memory network is called to input the time direction into the angular velocity sequence set. The sum of the angular velocity differences of consecutive nodes in the path segment is calculated through the memory state transfer method. The amplitude fluctuation range of each path segment is extracted. The angular velocity change trend in all path segments is compared. Combined segments with amplitude changes that remain in a convergent state are selected to generate a heading-stable path sequence. Angular velocity control module: Based on the heading stable path sequence, obtain the current path segment angular velocity value and heading change trend value, calculate the difference between the two item by item, generate a difference amplitude list, construct the steering residual sequence and filter out valley nodes, perform path segment marking operation at the residual abrupt change, select candidate segments with low amplitude values ​​in the same layer path, insert the corresponding nodes into the original path segment position, reconstruct the path and recalculate the angular velocity sequence, perform stability judgment on the sequence and replace all angular velocity items that exceed the set stability threshold, and generate heading control replacement sequence; The evasion path construction module: Based on the heading control replacement sequence, it extracts the magnitude difference of each angular velocity change point and compares it with the historical heading error value. It identifies the path node corresponding to each angular velocity change jump item, executes the replacement command query and extracts the low-risk command node located in the same path layer, combines the angular velocity command values ​​of each replacement node and establishes the updated path command structure to generate an evasion correction control channel.

Images