Unmanned aerial vehicle operational error dynamic threshold optimization method and system based on adaptive weighting

Abstract

This invention relates to the field of UAV operation error detection technology, and discloses a method and system for dynamic threshold optimization of UAV operation errors based on adaptive weighting. The method includes: parsing formation configuration data to generate role history data; retrieving the set of role operation characteristic parameters and load characteristic parameters; performing role baseline normalization on the operation time sequence feature set; generating a role-specific basic threshold template and performing cognitive adaptation transition processing; comparing operation deviations with thresholds to output the judgment result. This invention generates differentiated threshold standards through role adaptation transformation, eliminates feature jumps caused by role switching through role baseline normalization, and captures role-specific degradation patterns through multi-dimensional load weighting, effectively solving the problem of threshold mismatch in multi-UAV formation role dynamic switching scenarios.

Description

Technical Field

[0001] This invention relates to the field of unmanned aerial vehicle (UAV) operation error detection technology, and more specifically, to a method and system for optimizing dynamic thresholds for UAV operation errors based on adaptive weighting. Background Technology

[0002] In multi-UAV swarm collaborative flight missions, different UAVs assume different roles, and roles need to be dynamically switched due to factors such as battery rotation and target switching. The system needs to detect operator errors in real time based on dynamic threshold comparison of operational deviations. Existing technologies perform time-series weighting and cumulative risk calculation on historical sequences of operational deviations, retrieve basic threshold templates from a stage-threshold association database by combining task stage identifiers, and generate dynamic thresholds by dynamically scaling stage execution quality scores. Real-time operational deviations are then compared with dynamic thresholds to determine operational errors. However, existing dynamic threshold methods use a uniform threshold standard for all UAVs within the same task stage, failing to distinguish the significant differences in operational precision, control frequency, and response timeliness among different roles, leading to inappropriate threshold adaptation. When roles switch, the control load changes abruptly, and structural jumps in operational response delay and correction frequency are misjudged as operational degradation by cumulative risk calculations based on direct accumulation of deviations. Different roles consume different dimensions of operator attention and cognitive resources, and a uniform weighting method cannot capture role-specific operational degradation patterns. Summary of the Invention

[0003] This invention provides a method and system for optimizing the dynamic threshold of unmanned aerial vehicle (UAV) operation errors based on adaptive weighting. It solves the problem in related technologies that it is difficult to effectively handle the misjudgment of operation errors caused by the differences in role operation characteristics when operators frequently switch roles in multi-UAV formation tasks, as well as the technical problem that it is difficult to accurately distinguish between normal role switching operation fluctuations and real operation errors.

[0004] This invention discloses a dynamic threshold optimization method for UAV operation errors based on adaptive weighting, comprising the following steps: acquiring formation configuration data for multi-UAV collaborative tasks, parsing role assignment identifiers and role switching event logs of each UAV, extracting the role switching time sequence and corresponding role identifier sequence of the current operator during task execution, and generating role history data; Retrieve the set of character operation characteristic parameters corresponding to each character identifier in the character resume data from the character characteristic database. The set of character operation characteristic parameters includes position accuracy weight, mobility weight, and response speed weight. Obtain the operation data stream within the current role segment, extract the operation response delay sequence, operation correction frequency sequence, and operation amplitude variation coefficient sequence, generate an operation timing feature set, and perform role baseline normalization processing on the operation timing feature set based on the current role's role operation characteristic parameter set to eliminate the structural jumps in operation timing features caused by role switching, and generate role-normalized operation timing features. Obtain the current task stage identifier, retrieve the corresponding general basic threshold template from the stage-threshold association database, and perform role-adaptive transformation on the general basic threshold template based on the current role's role operation characteristic parameter set to generate a role-specific basic threshold template. The real-time operation deviation is compared with the role-specific basic threshold template. When the real-time operation deviation exceeds the corresponding threshold, the operation error judgment result is output.

[0005] Furthermore, the role baseline normalization process adopts the Z-score standardization method, which uses the mean and standard deviation of the operation time sequence characteristics of the current role under normal operation state as a benchmark to standardize the current operation time sequence characteristic value to obtain the normalized operation time sequence characteristics of the role; wherein, the mean and standard deviation are parameters obtained in advance based on the normal operation samples of each role in historical task data, and the normal operation samples are operation data segments of operators in historical tasks under conditions of no fatigue and no operation errors.

[0006] Furthermore, the role adaptation transformation is a process of differentially scaling the position deviation threshold component, maneuver deviation threshold component, and response deviation threshold component in the general basic threshold template based on the position accuracy weight, mobility weight, and response speed weight in the role operation characteristic parameter set; the threshold component of each dimension in the role-specific basic threshold template is the quotient obtained by dividing the threshold component of the corresponding dimension in the general basic threshold template by the weight of the current role in that dimension.

[0007] Furthermore, after generating the role-specific basic threshold template, it also includes: detecting whether the current moment is in the cognitive adaptation period after the role switch, wherein the cognitive adaptation period is the time required for the operator to transition from the previous role control mode to the current role control mode after the role switch occurs; During the cognitive adaptation period, a progressive convergence modulation process is applied to the character-specific basic threshold template. Based on the convergence coefficient that increases with the adaptation period, the character-specific basic threshold template corresponding to the previous character and the character-specific basic threshold template corresponding to the current character are linearly interpolated to generate the character transition threshold. The convergence coefficient is the ratio of the elapsed time from the moment of character switching to the preset adaptation period duration, and the value of the convergence coefficient is limited to between zero and one. During the cognitive adaptation period, the real-time operational deviation is compared with the role transition threshold; after the cognitive adaptation period ends, the real-time operational deviation is compared with the role-specific basic threshold template.

[0008] Furthermore, it also includes: extracting the attention load coefficient, control precision load coefficient and monotonicity load coefficient corresponding to each character identifier from the character characteristic database, associating the attention load coefficient, control precision load coefficient and monotonicity load coefficient with the corresponding character segment, and generating a character load feature vector sequence. For each role segment in the role load feature vector sequence, the flight time of the role segment is weighted according to the corresponding load coefficient, and the equivalent cumulative value of attention dimension, control dimension and cognition dimension are calculated by summing them to generate a multi-dimensional cumulative risk vector; wherein, the equivalent cumulative value of each dimension is the sum of the product of the flight time of each segment from the first role segment to the current role segment and the load coefficient of the corresponding role in that segment in that dimension.

[0009] Furthermore, it also includes: calculating the slope of change of the normalized operation time sequence features of the role within the short-term time window, the medium-term time window and the long-term time window respectively using a linear fitting algorithm to generate a multi-scale trend vector; The multi-scale trend vector and the multi-dimensional cumulative risk vector are concatenated and input into a multi-dimensional fatigue pattern recognition network, which outputs independent fatigue probability scores for each dimension and a comprehensive fatigue development stage label. The multi-dimensional fatigue pattern recognition network includes a temporal feature encoding layer, a multi-dimensional attention layer, and a classification output layer. The temporal feature encoding layer encodes the input sequence using an LSTM network and outputs a temporal encoded representation. The multi-dimensional attention layer performs weighted attention processing on the temporal encoded representation based on the cumulative risk weights of the attention dimension, manipulation dimension, and cognition dimension, and outputs attention representations for each dimension. The classification output layer includes independent probability output heads for each dimension and a comprehensive fatigue stage classification head. The independent probability output heads for each dimension output the independent fatigue probability score for that dimension using a sigmoid activation function, and the comprehensive fatigue stage classification head outputs the comprehensive fatigue development stage label using a softmax activation function.

[0010] Furthermore, it also includes: retrieving the fatigue sensitivity dimension weights corresponding to the current role from the role characteristic database based on the current role identifier, performing role sensitivity weighted aggregation on the independent fatigue probability scores of the attention dimension, manipulation dimension, and cognition dimension, and generating a comprehensive role sensitivity fatigue score; the comprehensive role sensitivity fatigue score is the sum of the product of the independent fatigue probability scores of each dimension and the fatigue sensitivity dimension weights of the current role in that dimension.

[0011] Furthermore, it also includes: calculating a threshold pre-adjustment coefficient based on the character's sensitivity fatigue comprehensive score, wherein the threshold pre-adjustment coefficient is the difference between a preset fatigue sensitivity adjustment coefficient and the product of the character's sensitivity fatigue comprehensive score, and applying the threshold pre-adjustment coefficient to the character's exclusive basic threshold template or the character's transition threshold for pre-tightening processing; The topology of inter-role collaboration relationships in the formation configuration data is analyzed, real-time status data of associated UAVs that have collaborative constraints with the current UAV are obtained, a collaborative constraint satisfaction index is calculated, a collaborative state adjustment coefficient is calculated based on the collaborative constraint satisfaction index, and the collaborative state adjustment coefficient is applied to the pre-tightened threshold to generate an adaptive weighted dynamic threshold. The real-time operation deviation is compared with the adaptive weighted dynamic threshold. When the real-time operation deviation exceeds the corresponding threshold, the operation error judgment result is output.

[0012] Furthermore, the calculation method of the collaborative constraint satisfaction index is as follows: For each associated UAV in the collaborative relationship topology that has a collaborative constraint relationship with the current UAV, calculate the normalized relative position deviation and the normalized relative speed deviation respectively. Then, sum the difference between the position constraint weight and speed constraint weight of each collaborative constraint relationship and the corresponding normalized deviation. Finally, take the average of the weighted summation results of all associated UAVs to obtain the collaborative constraint satisfaction index. The coordination state adjustment coefficient is determined as follows: when the coordination constraint satisfaction index is greater than or equal to the preset coordination normal threshold, the coordination state adjustment coefficient is one; when the coordination constraint satisfaction index is less than the coordination normal threshold, the coordination state adjustment coefficient is one minus the product of the preset coordination sensitivity adjustment coefficient and the difference between the coordination normal threshold and the coordination constraint satisfaction index. When the overall fatigue score of a character exceeds the preset character rotation trigger threshold, a character rotation suggestion is output; a fatigue warning level is output based on the overall fatigue development stage label.

[0013] This invention provides a dynamic threshold optimization system for unmanned aerial vehicle (UAV) operational errors based on adaptive weighting, comprising: The role history generation module is used to obtain the formation configuration data of multi-drone collaborative tasks, parse the role assignment identifier and role switching event log of each drone, extract the role switching time sequence and corresponding role identifier sequence of the current operator during the task execution process, and generate role history data. The role characteristic retrieval module is used to retrieve the set of role operation characteristic parameters corresponding to each role identifier in the role history data from the role characteristic database. The role normalization processing module is used to acquire the operation data stream within the current role segment, extract the operation response delay sequence, operation correction frequency sequence, and operation amplitude variation coefficient sequence, generate an operation timing feature set, and perform role baseline normalization processing on the operation timing feature set based on the current role's role operation characteristic parameter set to generate role normalized operation timing features. The role threshold generation module is used to obtain the current task stage identifier, retrieve the corresponding general basic threshold template from the stage-threshold association database, and perform role adaptation transformation on the general basic threshold template based on the current role's role operation characteristic parameter set to generate a role-specific basic threshold template. The operation error determination module is used to compare the real-time operation deviation with the role-specific basic threshold template. When the real-time operation deviation exceeds the corresponding threshold, the operation error determination result is output.

[0014] This invention generates role-specific basic threshold templates by performing role-adaptive transformation on a general basic threshold template using a set of role operation characteristic parameters. It provides differentiated threshold standards for different roles based on differences in operational precision, control frequency, and response timeliness, solving the adaptation mismatch problem caused by using a uniform threshold for all roles. Through role baseline normalization, it maps the operational timing features of different roles to a unified metric space, eliminating structural jumps in operational timing features caused by role switching and solving the problem of misjudging role switching feature jumps as operational degradation. By calculating a multi-dimensional cumulative risk vector using multi-dimensional load coefficient weighting and performing role-sensitive weighted aggregation using role fatigue-sensitive dimension weights, fatigue detection can focus on the dimensions most susceptible to fatigue under the current role, solving the problem that a uniform weighting method cannot capture role-specific degradation patterns. By applying progressive convergence modulation processing to the role-specific basic threshold template during the cognitive adaptation period, the threshold change rhythm is kept consistent with the operator's cognitive adaptation process, avoiding misjudgments caused by threshold jumps. Attached Figure Description

[0015] Figure 1 This is a flowchart of the dynamic threshold optimization method for unmanned aerial vehicle (UAV) operation errors based on adaptive weighting provided in the embodiments of the present invention; Figure 2 This is a schematic diagram of the timeline of operator role resume provided in an embodiment of the present invention; Figure 3 This is a schematic diagram illustrating the comparison of role operation characteristic weights provided in an embodiment of the present invention; Figure 4 This is a schematic diagram illustrating the comparison of role load characteristic coefficients provided in an embodiment of the present invention; Figure 5 This is a schematic diagram illustrating the evolution of operation timing features before and after role switching, provided in an embodiment of the present invention. Figure 6 This is a schematic diagram illustrating the trend of operational characteristics after role normalization provided in an embodiment of the present invention; Figure 7 This is a schematic diagram of the gradual convergence process of the cognitive adaptation period threshold provided in an embodiment of the present invention; Figure 8 This is a schematic diagram of the evolution of the multidimensional cumulative risk vector provided in an embodiment of the present invention. Detailed Implementation

[0016] In multi-UAV swarm collaborative flight missions, the system needs to detect operator errors in real time based on dynamic threshold comparison of operational deviations. Existing technologies perform time-series weighting and cumulative risk calculation on historical sequences of operational deviations, retrieve basic threshold templates from a stage-threshold association database in conjunction with mission stage identifiers, generate dynamic thresholds through dynamic scaling of stage execution quality scores, and compare real-time operational deviations with dynamic thresholds to determine operational errors.

[0017] However, in multi-UAV swarm collaborative missions, different UAVs assume different roles (such as navigator, follower, reconnaissance, etc.), and need to dynamically switch roles due to battery rotation, target switching, etc. Existing dynamic threshold methods have the following technical problems: First, applying a uniform threshold standard to all UAVs within the same mission phase fails to distinguish the significant differences in operational precision, control frequency, and response timeliness among different roles, resulting in an overly wide operational deviation threshold for the navigator and an overly strict operational deviation threshold for the follower, leading to inappropriate threshold adaptation; Second, the operational load changes abruptly during role switching, and the structural jumps in operational response delay and correction frequency are misjudged as operational degradation by the cumulative risk calculation based on the direct accumulation of deviations, failing to distinguish between changes in operational characteristics caused by role switching and gradual operational degradation caused by fatigue; Third, different roles consume different dimensions of operator attention and cognitive resources, and a uniform weighting method cannot capture role-specific operational degradation patterns, making early degradation signals more easily masked by role switching interference.

[0018] According to an embodiment of this implementation, a method for optimizing the dynamic threshold of unmanned aerial vehicle (UAV) operational errors based on adaptive weighting is provided. It should be understood that the computer system executing the method is communicatively connected to a ground control station of a multi-UAV formation, enabling it to acquire formation configuration data, flight status data of each UAV, and operator operation data streams in real time. A role characteristic database and a stage-threshold association database are pre-deployed in the computer system.

[0019] At least one embodiment of the present invention discloses a method for optimizing the dynamic threshold of unmanned aerial vehicle (UAV) operational errors based on adaptive weighting, such as... Figure 1 As shown, it includes the following steps: Step 1: Parse the formation configuration data and generate character resume data; Acquire formation configuration data for multi-drone collaborative tasks, parse the role assignment identifiers and role switching event logs of each drone, extract the role switching time sequence and corresponding role identifier sequence of the current operator during task execution, and generate role history data. The role history data is organized into a role segment sequence in chronological order, with each role segment containing a role identifier, start time, and end time.

[0020] It should be noted that the role assignment identifier is the identification information that marks the functional role undertaken by each UAV in the formation, including role type identifiers such as navigator, follower, and reconnaissance aircraft. The role switch event log is a structured log data that records the timestamps, the role identifier before the switch, and the role identifier after the switch for each UAV during mission execution. In other words, the role history data completely records all role segments and their time span experienced by the current operator from the start of the mission to the present moment.

[0021] In this embodiment of the application, in order to obtain the cooperative dependency information between each UAV in the formation to support subsequent cooperative state assessment, the role cooperative relationship topology in the formation configuration data is also parsed. The role cooperative relationship topology describes the data dependencies, command transmission and formation constraint relationships between each role, and is represented by a graph structure, where nodes represent role instances and edges represent cooperative constraint relationships and their constraint parameters (including relative position constraints, relative speed constraints and communication delay constraints, etc.).

[0022] Step 2: Retrieve the set of role operation characteristic parameters and load characteristic parameters; Retrieve the set of character operation characteristic parameters corresponding to each character identifier from the character profile data in the character characteristic database. The set of character operation characteristic parameters includes position accuracy weight, mobility weight, and response speed weight.

[0023] It should be noted that the positional accuracy weight is a parameter value that measures the degree of positional accuracy required for the role corresponding to the role identifier. For example, the navigator needs to accurately maintain the preset route, so its positional accuracy weight is relatively high. The maneuverability weight is a parameter value that measures the degree of frequency and magnitude of maneuver operations required for the role corresponding to the role identifier. For example, the follower needs to frequently adjust its attitude to maintain formation, so its maneuverability weight is relatively high. The response speed weight is a parameter value that measures the degree of timeliness required for the response of the action corresponding to the role identifier. The role characteristic database is a set of parameters pre-calibrated according to the task characteristics and operational requirements of each role.

[0024] In this embodiment, to support subsequent multi-dimensional fatigue accumulation risk assessment, attention load coefficient, control precision load coefficient, and monotonicity load coefficient corresponding to each role identifier are extracted from the role characteristic database. These coefficients are then associated with corresponding role segments to generate a role load feature vector sequence. Specifically, the attention load coefficient represents the intensity of the operator's sustained attention consumption by the role corresponding to the role identifier; the control precision load coefficient represents the intensity of the consumption of fine control ability by the role corresponding to the role identifier; and the monotonicity load coefficient represents the intensity of the consumption of the operator's cognitive resources by the repetitive nature of the task corresponding to the role identifier.

[0025] Step 3: Perform role baseline normalization on the operation time sequence feature set; The system acquires the operation data stream within the current role segment, extracts the operation response delay sequence, operation correction frequency sequence, and operation amplitude variation coefficient sequence, and generates an operation timing feature set. Based on the current role's operation characteristic parameter set, the system performs role baseline normalization processing on the operation timing feature set. Using the feature distribution of the current role under normal operation as a benchmark, it eliminates structural jumps in operation timing features caused by role switching, generating normalized operation timing features for the role.

[0026] It should be noted that the character baseline normalization process uses Z-score standardization, which standardizes the current operation time-series features based on the mean and standard deviation of features for each character under normal operating conditions. Specifically, for the current character... A certain operation timing feature Role normalization operation timing characteristics The calculation method is as follows: in, This is the current operation timing characteristic value. For the role The mean of the timing characteristics under normal operating conditions. For the role The standard deviation of the timing characteristics under normal operating conditions.

[0027] Furthermore, the mean of the above-mentioned normal operating characteristic distribution and standard deviation The parameters are obtained in advance based on the normal operation samples of each role in the historical task data. The normal operation samples are operation data segments of the operator in the historical task when the operator is not fatigued and no operation error has occurred.

[0028] Furthermore, through role baseline normalization, the operational timing features under different roles are mapped to a unified metric space, so that the feature value changes before and after role switching can reflect the changes in the operator's own operational state, rather than the baseline shift caused by role differences.

[0029] Step 4: Generate a character-specific basic threshold template and perform cognitive adaptation transition processing; Obtain the current task stage identifier, retrieve the corresponding general basic threshold template from the stage-threshold association database, and perform role-adaptive transformation on the general basic threshold template based on the current role's role operation characteristic parameter set to generate a role-specific basic threshold template.

[0030] It should be noted that the character adaptation transformation is a process of differentially scaling the threshold components of each dimension in the general basic threshold template based on the position accuracy weight, mobility weight, and response speed weight in the character's operational characteristic parameter set. For the position deviation threshold component in the general basic threshold template... , Maneuver deviation threshold component and response deviation threshold component The calculation method for the corresponding component in the character-specific basic threshold template is as follows: in, This refers to the position deviation threshold component in the general basic threshold template. For the maneuver deviation threshold component in the general basic threshold template, For the response bias threshold component in the general basic threshold template, For the role The character-specific positional deviation threshold component, For the role Character-specific maneuver deviation threshold component For the role The role-specific response deviation threshold component, For the role Position accuracy weights, For the role Mobility weight, For the role The response speed weight.

[0031] Furthermore, when the weight of a certain dimension is high, the threshold component of the corresponding dimension is reduced, thereby enabling more rigorous detection of operational deviations in that dimension.

[0032] In this embodiment, to avoid misjudgments caused by the mismatch between the threshold change of the role switching time and the operator's cognitive adaptation rhythm, it is detected whether the current time is within the cognitive adaptation period after the role switch. The cognitive adaptation period is the time required for the operator to transition from the previous role control mode to the current role control mode after the role switch occurs. The length of the cognitive adaptation period is determined based on the switching time in the role switching event log and a preset adaptation period duration parameter.

[0033] Furthermore, during the cognitive adaptation period, a progressive convergence modulation process is applied to the role-specific basic threshold template, gradually transitioning from the role-specific basic threshold template corresponding to the previous role to the role-specific basic threshold template corresponding to the current role, thus generating a role transition threshold. : in, For at any time Role transition threshold, This is a base threshold template specific to the previous character. This is a character-specific base threshold template for the current role. The convergence coefficient increases with the adaptation period and , This represents the elapsed time since the character switch.

[0034] Furthermore, the convergence coefficient The value and adaptation period Related, its calculation method is as follows ,in This is the preset adaptation period duration. At the beginning of the cognitive adaptation period, Approaching 0, the character transition threshold is close to the previous character's unique base threshold template level; as the adaptation period progresses, Gradually approaching 1, the role transition threshold gradually converges to the level of the current role's specific basic threshold template. After the cognitive adaptation period ends, the role's specific basic threshold template is directly adopted.

[0035] Step 5: Compare the operational deviation with the character-specific basic threshold template and output the judgment result; The real-time operation deviation is compared with the character-specific basic threshold template (or with the character transition threshold during the cognitive adaptation period). When the real-time operation deviation exceeds the corresponding threshold, the operation error judgment result is output.

[0036] It should be noted that the comparison process involves comparing the values ​​of each dimension component of the real-time operation deviation with the corresponding threshold component in the character-specific basic threshold template. When the operation deviation component in any dimension exceeds the threshold component corresponding to that dimension, an operation error is determined to exist in that dimension, and an operation error determination result containing the error dimension identifier and the deviation exceeding amount is output.

[0037] In this embodiment of the application, based on obtaining the role load feature vector sequence in step 2, the following steps are also included to achieve fatigue perception threshold adjustment based on multidimensional cumulative risk: Step 501: Calculate the multidimensional cumulative risk vector; For each role segment in the role load feature vector sequence, the flight duration of the role segment is weighted according to the corresponding load coefficient. The equivalent cumulative values ​​for the attention dimension, control dimension, and cognition dimension are then calculated and summed to generate a multi-dimensional cumulative risk vector. Specifically, for the segment up to the [number missing]th ... Equivalent cumulative value of attention dimension for each role segment The calculation method is as follows: in, As of the date Equivalent cumulative value of attention dimension for each role segment The segment number for the current character. For summation index and Traverse from 1 to , For the first The attention load coefficient for each role segment. For the first Flight duration for each character segment. Equivalent cumulative value for control dimensions. Equivalent cumulative value of cognitive dimension Using control precision load factor respectively and monotonic load factor The calculation is performed in the same manner. The multidimensional cumulative risk vector is represented as follows: ,in As of the date The equivalent cumulative value of the control dimension for each character segment. As of the date The equivalent cumulative value of the cognitive dimensions of each role segment.

[0038] Step 502: Perform multi-scale trend analysis and fatigue pattern identification; Multi-scale trend analysis was performed on the time series characteristics of normalized operations for roles, specifically within short-term time windows. Mid-term time window and long-term time window Within the network, a linear fitting algorithm is used to calculate the slope of change for each feature, generating a multi-scale trend vector. This multi-scale trend vector is then concatenated with a multi-dimensional cumulative risk vector and input into a multi-dimensional fatigue pattern recognition network, outputting independent fatigue probability scores for each dimension and a comprehensive fatigue development stage label.

[0039] It should be noted that the slope of change is the slope value obtained by linearly fitting the feature time series data within the corresponding time window, reflecting the direction and rate of change of the feature at that time scale. Short-term time window. Used to capture short-term operational status fluctuations, medium-term time window Long-term time window used to capture transitional trends in operational degradation Used to capture persistent operational degradation signals.

[0040] Furthermore, short-term time window The duration ranges from 3 to 5 minutes, with a mid-term time window. The duration ranges from 10 to 15 minutes, with a long time window. The duration ranges from 30 to 45 minutes.

[0041] Furthermore, the multidimensional fatigue pattern recognition network is a neural network structure that includes a temporal feature encoding layer, a multidimensional attention layer, and a classification output layer.

[0042] Furthermore, the temporal feature encoding layer receives the concatenated multi-scale trend vector and multi-dimensional cumulative risk vector as input, encodes the input sequence through an LSTM network, extracts temporal dependent features, and outputs a temporal encoded representation.

[0043] Furthermore, the multi-dimensional attention layer receives temporal encoded representations as input, and performs weighted attention processing on the temporal encoded representations based on the cumulative risk weights of the attention dimension, manipulation dimension, and cognition dimension, outputting attention representations for each dimension.

[0044] Furthermore, the classification output layer includes independent probability output heads for each dimension and a comprehensive fatigue stage classification head. Each independent probability output head receives the attention representation for its corresponding dimension as input, is mapped through a fully connected layer, and then outputs an independent fatigue probability score for that dimension via a sigmoid activation function. The range of values ​​for the independent fatigue probability score is... Specifically, this includes fatigue probability scores for the attention dimension, the manipulation dimension, and the cognition dimension. The comprehensive fatigue stage classification head receives the concatenated vectors of attention representations from each dimension as input. After being mapped through a fully connected layer, it outputs the probability distribution of the comprehensive fatigue development stage label through a softmax activation function. The comprehensive fatigue development stage label includes stages such as normal, mild fatigue, moderate fatigue, and severe fatigue. The final output comprehensive fatigue development stage label is the category with the highest probability value in the probability distribution.

[0045] Furthermore, the multidimensional fatigue pattern recognition network is trained using supervised learning. The training data comes from operator fatigue state data labeled in historical tasks, including fatigue labels for each dimension and comprehensive fatigue stage labels. The loss function is a weighted combination of the binary cross-entropy loss of independent fatigue probability scores for each dimension and the multi-class cross-entropy loss for comprehensive fatigue stage classification.

[0046] Step 503: Generate the character's overall sensitivity and fatigue score; Based on the current role identifier, the fatigue sensitivity dimension weights of the corresponding role are retrieved from the role characteristic database. The independent fatigue probability scores of each dimension are then aggregated using role sensitivity weighting to generate a comprehensive role-sensitive fatigue score. : in, The overall score for character sensitivity and fatigue. For dimension index and Taken from set , Represents the attention dimension. Indicates the dimension of manipulation. Represents cognitive dimensions. For the current role In dimensions The weight of fatigue sensitivity dimension on the surface. For dimension Independent fatigue probability fractions.

[0047] Furthermore, the fatigue sensitivity dimension weight reflects the degree to which the degradation of each dimension affects operational safety under the current role. For example, the navigator role is more sensitive to fatigue in the attention dimension, so its fatigue sensitivity dimension weight in the attention dimension is... The weight of fatigue sensitivity in the control dimension is relatively high; since follower characters are more sensitive to fatigue in the control dimension, the weight of fatigue sensitivity in the control dimension is relatively high. Relatively high.

[0048] Step 504: Generate an adaptive weighted dynamic threshold based on the comprehensive score of role-sensitive fatigue and collaborative state; Based on the overall score of role-sensitive fatigue Calculate the threshold pre-adjustment coefficient , threshold pre-adjustment coefficient Pre-tightening is applied to the role transition threshold (or the role-specific baseline threshold template outside the cognitive adaptation period). Threshold pre-adjustment coefficient. Combined score with character sensitivity and fatigue They are inversely correlated, that is The higher, The smaller the value, the greater the tightening of the threshold.

[0049] It should be noted that the threshold pre-adjustment coefficient The calculation method is as follows: in, This is the threshold pre-adjustment coefficient. The preset fatigue sensitivity adjustment coefficient has a value range of [value missing]. , This represents the overall score for character sensitivity and fatigue. When... When it increases, Decrease, so that the threshold is multiplied by It was later tightened.

[0050] Furthermore, real-time status data of associated drones that have collaborative constraints with the current drone in the role-to-role collaborative relationship topology are obtained, and collaborative constraint satisfaction index is calculated. Based on the collaborative constraint satisfaction index The pre-tightened threshold is dynamically adjusted to generate an adaptive weighted dynamic threshold.

[0051] Furthermore, the collaborative constraint satisfaction index To measure the degree to which the cooperative constraints in the current formation are satisfied, a comprehensive index is obtained by weighted calculation based on the relative position deviation, relative speed deviation, and formation maintenance degree between associated UAVs.

[0052] Furthermore, the collaborative constraint satisfaction index The calculation method is as follows: in, As a collaborative constraint satisfaction index, This represents the number of associated drones in the inter-role collaboration relationship topology that have collaborative constraints with the current drone. For summation index and Traverse from 1 to , In order to be with the first Normalized relative position deviations between the associated drones In order to be with the first Normalized relative speed deviation between the associated drones For the first The positional constraint weights of the collaborative constraint relationships. For the first The speed constraint weights of the collaborative constraint relationships.

[0053] Furthermore, when the collaborative constraint satisfaction index A lower threshold indicates that the formation coordination state deviates from expectations, and the operational deviation threshold should be further tightened to improve detection sensitivity.

[0054] Furthermore, adaptive weighted dynamic threshold The calculation method is as follows: in, For adaptive weighted dynamic threshold, This is a template for character transition thresholds or character-specific basic thresholds. This is the threshold pre-adjustment coefficient. This is the collaborative state adjustment coefficient based on the collaborative constraint satisfaction index.

[0055] Furthermore, the cooperative state adjustment coefficient The calculation method is as follows: when Greater than or equal to the preset collaborative normal threshold hour, ;when Less than the normal threshold for coordination hour, ,in The preset threshold for normal coordination. The preset collaborative sensitivity adjustment coefficient has a value range of [value range missing]. .when When within the normal range, When the threshold is close to 1, no additional threshold adjustment is performed; when When it is below the normal range, If the value is less than 1, the threshold will be further tightened.

[0056] Step 505: Output a comprehensive judgment result based on adaptive weighted dynamic threshold; The system compares real-time operational deviations with adaptive weighted dynamic thresholds to output operational error judgment results. Furthermore, it outputs fatigue warning levels based on comprehensive fatigue development stage labels; it outputs role switching adaptation status assessments based on the convergence status of the cognitive adaptation period's time progress and role-normalized operational timing characteristics within the cognitive adaptation period; and it outputs a comprehensive fatigue score for each role. When the preset character rotation trigger threshold is exceeded, output a character rotation suggestion.

[0057] It should be noted that the role rotation trigger threshold is a pre-set fatigue comprehensive score threshold. When the comprehensive fatigue score of a role exceeds the role rotation trigger threshold, it indicates that the operator's fatigue level in the current role has reached a level that requires role rotation or rest.

[0058] The operational error detection method provided in this embodiment addresses the inadequacy of existing unified threshold schemes in scenarios involving dynamic role switching in multi-drone formations. It achieves adaptive optimization of operational error detection through the following technical means: First, because the set of character operation characteristic parameters corresponding to each character was retrieved from the character characteristic database, and the general basic threshold template was adapted to the character based on position accuracy weight, mobility weight and response speed weight, the generated character-specific basic threshold template can provide differentiated threshold standards for different characters in terms of operation accuracy, control frequency and response timeliness, thereby overcoming the problem of misfitting caused by using a uniform threshold for all characters.

[0059] Second, because the operation timing feature set is normalized based on the feature distribution of the current role under normal operation, the operation timing features under different roles are mapped to a unified metric space. Therefore, the structural jumps in operation response delay and correction frequency during role switching are eliminated as role baseline differences. The change signals extracted by subsequent trend analysis can reflect the gradual changes in the operator's own operation status, thus overcoming the problem that the jumps in role switching features are misjudged as operation degradation.

[0060] Third, because the flight duration of each role segment is weighted in multiple dimensions according to the attention load coefficient, control precision load coefficient and monotonicity load coefficient corresponding to the role, a multidimensional cumulative risk vector is generated that distinguishes different cognitive resource consumption dimensions. Furthermore, the fatigue sensitivity dimension weight is used in the aggregation stage of fatigue probability score to perform role-sensitive weighting. Therefore, fatigue detection can focus on the dimension most susceptible to fatigue under the current role, thus overcoming the problem that the unified weighting method cannot capture the role-specific degradation pattern.

[0061] Fourth, because a gradual convergence modulation process is applied to the role-specific basic threshold template during the cognitive adaptation period after role switching, the threshold gradually transitions from the previous role level to the current role level. Therefore, the pace of threshold change is consistent with the operator's cognitive adaptation process, avoiding misjudgments caused by threshold jumps. Furthermore, because a collaborative constraint satisfaction index based on topological calculation of inter-role collaborative relationships is introduced to dynamically adjust the threshold, the threshold can respond to changes in the formation collaborative state, improving detection sensitivity when the formation collaborative state deviates from expectations.

[0062] In a border patrol mission, a coordinated formation of five UAVs performed a long-endurance surveillance mission. The formation configuration was as follows: one lead UAV (UAV-L01) was responsible for maintaining the flight path and formation navigation; two follower UAVs (UAV-F02 and UAV-F03) were responsible for maintaining formation; one reconnaissance UAV (UAV-R04) was responsible for scanning the target area; and one communications relay UAV (UAV-C05) was responsible for data transmission. Operator A began the mission at 08:30 on March 15, 20XX, initially operating UAV-L01 as the lead aircraft. At 10:15, due to UAV-L01's battery level dropping to 45%, the system triggered a role rotation, and Operator A switched from the lead aircraft to take over the follower role on UAV-F02. The ground control station's error detection system needed to monitor the operator's operational status in real time before and after the role switch, identify fatigue accumulation and operational deviations, and prevent operational errors caused by role switching.

[0063] The system retrieves the formation configuration data packet from the ground control station and parses out the role assignment change records of Operator A during the mission execution. From the start of the mission at 08:30 to the role switch at 10:15, Operator A operated UAV-L01 as the navigator, with a flight time of 105 minutes; from 10:15 onwards, Operator A took over UAV-F02 as the follower. The system extracts the role switch timestamp and corresponding role identifier to generate role history data.

[0064] Table 1 Operator Role Resume Data: At the same time, the system analyzes the collaborative relationship topology in the formation configuration data and identifies the collaborative constraint relationships between each UAV.

[0065] Table 2 Topology data of collaboration relationships between roles: The role history data records the complete switching process of operator A from the navigator role to the follower role in chronological order. The collaborative relationship topology data describes the dependencies and constraint parameters between the drones in the formation, providing a basis for subsequent collaborative status assessment.

[0066] Based on the role identifier in the role resume data, the system retrieves the set of role operation characteristic parameters corresponding to the navigator and follower aircraft from the role characteristic database.

[0067] Table 3: Character Operation Characteristic Parameter Set The positional accuracy weight of the navigator is 1.2, higher than the follower's 0.9, indicating that the navigator requires higher precision in maintaining the flight path. The maneuverability weight of the follower is 1.3, higher than the navigator's 0.8, reflecting the need for frequent attitude adjustments to maintain formation. Regarding load characteristics, the navigator's attention load coefficient is 0.85, higher than the follower's 0.75, because the navigator needs to continuously monitor flight path deviations and obstacle avoidance. The follower's control precision load coefficient is 0.90, higher than the navigator's 0.70, because the follower requires more precise control to maintain formation.

[0068] The system associates load characteristic parameters with role segments to generate a sequence of role load characteristic vectors.

[0069] Table 4: Role Load Feature Vector Sequence The role-based load feature vector sequence provides a basis for subsequent multidimensional cumulative risk calculations, with load weights categorized by role and dimension.

[0070] The system acquires the operation data stream of operator A within the navigation role segment (08:30-10:15) and extracts the operation timing features. Within the time period of 10:10-10:15, the system extracts the operation response delay sequence, operation correction frequency sequence, and operation amplitude variation coefficient sequence.

[0071] Table 5. Raw data on the timing characteristics of the navigation role's operations (10:10-10:15 time period): The role switch occurred at 10:15. After operator A took over the follower UAV-F02, the operation timing features extracted during the period from 10:15 to 10:20 showed a structural jump.

[0072] Table 6. Raw data on the timing characteristics of the following character's actions (10:15-10:20 time period): The response delay for actions under the following role jumped from 2.7 seconds to 1.5 seconds, and the frequency of action corrections jumped from 4.3 times / minute to 5.8 times / minute. This jump is due to the baseline shift caused by the difference in roles, rather than changes in the operator's own state.

[0073] The system retrieves the characteristic distribution parameters of the navigator and follower aircraft under normal operating conditions from the role characteristic database.

[0074] Table 7 Distribution parameters of normal character operation characteristics: The system performs role baseline normalization on the operational timing characteristics of the navigator role during different time periods. Taking the operational response delay at 10:14:30 as an example: Similarly, the timing characteristics of operations during the following character's time period are normalized. Taking the operation response delay at 10:15:30 as an example: Table 8. Normalized Operation Timing Characteristics of Roles: By normalizing the role baseline, the feature jumps at the time of role switching are eliminated, and the normalized feature values ​​can reflect the gradual change trend of the operator's own operating status. As can be seen from the normalized data, the normalized feature values ​​following the role time period show a continuous upward trend, indicating that the operator's operating deviation after role switching gradually increases relative to the normal baseline of that role.

[0075] The system identifies the current task phase as "cruise monitoring phase" and retrieves the corresponding general basic threshold template from the phase-threshold association database.

[0076] Table 9 General Basic Threshold Template (Cruise Monitoring Phase): The system adapts a general basic threshold template to the navigator's role by applying the navigator's role-specific operational characteristic parameter set, generating a navigator-specific basic threshold template. For example, the navigator's positional accuracy weight is 1.2 for the position deviation threshold. Similarly, the computer dynamic deviation threshold and response deviation threshold are: For follower machines, generate a base threshold template specific to the follower machine role: The position deviation threshold for the navigator was tightened to 12.5 meters because of its higher weighting for position accuracy; the maneuver deviation threshold for the follower was tightened to 6.2 degrees / second because of its higher weighting for maneuverability.

[0077] The system detected a role switch at 10:15, entering a cognitive adaptation period. The preset adaptation period duration is 8 minutes. The system applies progressive convergence modulation processing to the role-specific basic threshold template to generate a role transition threshold.

[0078] Taking the position deviation threshold as an example, at 10:17 (2 minutes after the role switch), the convergence coefficient is calculated as follows: The threshold for character transition position deviation is: At 10:23 (8 minutes after the role switch), the cognitive adaptation period ended. The character transition threshold is completely converged to the basic threshold template specific to the random character.

[0079] Similarly, calculate the role transition thresholds for other time periods and dimensions to generate a complete role transition threshold evolution sequence.

[0080] The role transition threshold gradually shifts from the navigator level to the follower level, keeping pace with the operator's cognitive adaptation process and avoiding misjudgments caused by threshold jumps.

[0081] The system acquires operator A's real-time operational deviation data at 10:20, which is within the cognitive adaptation period (5 minutes after role switching).

[0082] The real-time operational deviations are: position deviation 14.2 meters, maneuver deviation 7.8 degrees / second, and response deviation 2.8 seconds.

[0083] Based on the convergence coefficient Calculate the role transition threshold at the corresponding time point: Position deviation threshold: rice Maneuver deviation threshold: degrees / second Response deviation threshold: Second The system performs the following comparisons: - Position deviation 14.2 meters < Position deviation threshold 15.1 meters, not exceeded - Maneuver deviation 7.8 degrees / second > Maneuver deviation threshold 7.6 degrees / second, exceeding by 0.2 degrees / second - Response deviation 2.8 seconds < Response deviation threshold 2.82 seconds, not exceeded The system outputs the operation error judgment result: An operation error in the dimension of the maneuver deviation was detected, exceeding the limit by 0.2 degrees / second, triggering a level one warning.

[0084] The system calculates the multidimensional cumulative risk vector up to 10:20 based on the role load feature vector sequence.

[0085] The flight time for the Navigator segment (Segment 1) is 105 minutes, while the flight time for the Follower segment (Segment 2) is 5 minutes from 10:15 to 10:20.

[0086] Calculation of equivalent cumulative value of attention dimension: Calculation of equivalent cumulative value of manipulation dimension: Cognitive dimension equivalent cumulative value calculation: The multidimensional cumulative risk vector is .

[0087] The system performs multi-scale trend analysis on the time-series characteristics of normalized operations for each role. During the time period of following the role (10:15-10:20), a short-term window of 3 minutes and a medium-term window of 5 minutes are set.

[0088] Linear fitting was performed on the data points (1.33, 1.67, 2.00) within the short-term window (10:17-10:20) of the normalized response delay feature. The slope of the change was calculated using the least squares method and was 0.67, indicating that the normalized response delay increased at a rate of 0.67 / minute.

[0089] Linear fitting was performed on the data points (0.38, 0.88, 1.25, 1.63, 1.88) within the mid-term window (10:15-10:20) of the normalized corrected frequency features, with a slope of 0.30.

[0090] A linear fit was performed on the data points (0.50, 0.83, 1.17, 1.50, 1.67) within the medium-term window, with a slope of 0.23.

[0091] The system concatenates the multi-scale trend vector [0.67, 0.30, 0.23] with the multi-dimensional cumulative risk vector [93.0, 78.0, 71.0] and inputs the result into a multi-dimensional fatigue pattern recognition network. The network outputs independent fatigue probability scores for each dimension: 0.62 for the attention dimension, 0.58 for the manipulation dimension, and 0.48 for the cognitive dimension. The overall fatigue development stage is labeled as "mild fatigue".

[0092] The system retrieves the fatigue sensitivity dimensions weights of the following characters from the character characteristic database: attention dimension weight 0.3, control dimension weight 0.5, and cognition dimension weight 0.2.

[0093] Calculate the character's overall sensitivity and fatigue score: Follower characters are more sensitive to fatigue in the control dimension. The fatigue probability of the control dimension (0.58) contributes the most to the overall score after being weighted by a high weight of 0.5.

[0094] Step 504: Generate adaptive weighted dynamic thresholds The system is based on a comprehensive score of role-sensitive fatigue. Set the fatigue sensitivity adjustment coefficient to 0.4, and calculate the threshold pre-adjustment coefficient: The system acquires real-time status data of associated UAVs in the formation that have cooperative constraints with UAV-F02. The relative positional deviation between UAV-F02 and UAV-L01 is 3.8 meters (preset constraint 50±5 meters, allowable deviation range 5 meters, normalized deviation). The relative velocity deviation is 1.2 m / s (preset constraint ≤ 2 m / s, normalized deviation). The lateral spacing deviation between UAV-F02 and UAV-F03 is 2.1 meters (preset constraint 30±3 meters, allowable deviation range 3 meters, normalized deviation). The relative velocity deviation is 0.5 m / s (preset constraint ≤ 1 m / s, normalized deviation). ).

[0095] Calculate the collaborative constraint satisfaction index: The coordination constraint satisfaction index of 0.314 is lower than the preset normal coordination threshold of 0.75, indicating that the formation coordination state deviates from expectations. The coordination sensitivity adjustment coefficient is set to 0.5, and the coordination state adjustment coefficient is calculated as follows: Generate adaptive weighted dynamic thresholds. Taking the positional deviation threshold as an example, the role transition threshold is 15.1 meters at this time: Similarly, calculate the adaptive weighted dynamic thresholds for other dimensions: Maneuver deviation threshold: degrees / second Response deviation threshold: Second At 10:20, the system compares the real-time operational deviation with the adaptive weighted dynamic threshold.

[0096] Real-time operational deviations: position deviation 14.2 meters, maneuver deviation 7.8 degrees / second, response deviation 2.8 seconds.

[0097] Adaptive weighted dynamic thresholds: position deviation threshold 9.1 meters, maneuver deviation threshold 4.6 degrees / second, response deviation threshold 1.7 seconds.

[0098] Comparison results: - Position deviation 14.2 meters > Position deviation threshold 9.1 meters, exceeding by 5.1 meters - Maneuver deviation 7.8 degrees / second > Maneuver deviation threshold 4.6 degrees / second, exceeding by 3.2 degrees / second - Response deviation 2.8 seconds > Response deviation threshold 1.7 seconds, exceeding by 1.1 seconds The system outputs the operation error judgment result: the position deviation, maneuver deviation and response deviation have all exceeded the adaptive weighted dynamic threshold, triggering a level 3 severe warning.

[0099] The system outputs a fatigue warning level of Level II based on the comprehensive fatigue development stage label "mild fatigue".

[0100] The system assesses the adaptation status of role switching: It is currently in the 5th minute of the cognitive adaptation period. The normalized operation timing characteristics show a continuous upward trend, and the convergence is poor. The output adaptation status assessment is "adaptation progress is lagging".

[0101] The system detected that the character's sensitivity fatigue score was 0.572, which exceeded the preset character rotation trigger threshold of 0.55. The system output a character rotation suggestion: It is recommended to rotate operators or arrange a short rest after completing the current cruise segment.

[0102] Throughout the implementation process, data begins with the role history data of the formation configuration. After retrieving role characteristic parameters, the operational characteristic weights and load coefficients of different roles are obtained. The original operational time-series features are normalized using role baselines to eliminate structural jumps caused by role switching, and mapped to a unified metric space. A general basic threshold template is transformed into role-specific thresholds through role adaptation, and a smooth transition is achieved through progressive convergence modulation during the cognitive adaptation period. A multi-dimensional cumulative risk vector, combined with multi-scale trend analysis, is input into the fatigue pattern recognition network, outputting independent fatigue probabilities for each dimension and a comprehensive fatigue stage. The role-sensitive fatigue comprehensive score is converted into a threshold pre-adjustment coefficient through a fatigue sensitivity adjustment coefficient, and the collaborative constraint satisfaction index further generates a collaborative state adjustment coefficient. Both of these factors work together to form an adaptive weighted dynamic threshold for the role transition. Finally, real-time operational deviations are compared with the dynamic threshold to output operational error judgments, fatigue warnings, adaptation state assessments, and role rotation suggestions, forming a complete data flow chain and ensuring the comprehensive adaptability of the detection process to role differences, role switching, and fatigue accumulation.

[0103] The embodiments of the present invention have been described above. However, the embodiments are not limited to the specific implementation methods described above. The specific implementation methods described above are merely illustrative and not restrictive. Those skilled in the art can make more equivalent embodiments under the guidance of the present embodiments, and all of them are within the protection scope of the present embodiments.

Claims

1. A method for optimizing the dynamic threshold of unmanned aerial vehicle (UAV) operational errors based on adaptive weighting, characterized in that, Includes the following steps: Acquire formation configuration data for multi-drone collaborative tasks, parse role assignment identifiers and role switching event logs for each drone, extract the role switching time sequence and corresponding role identifier sequence of the current operator during task execution, and generate role resume data; Retrieve the set of character operation characteristic parameters corresponding to each character identifier in the character resume data from the character characteristic database. The set of character operation characteristic parameters includes position accuracy weight, mobility weight, and response speed weight. Obtain the operation data stream within the current role segment, extract the operation response delay sequence, operation correction frequency sequence, and operation amplitude variation coefficient sequence, generate an operation timing feature set, and perform role baseline normalization processing on the operation timing feature set based on the current role's role operation characteristic parameter set to eliminate the structural jumps in operation timing features caused by role switching, and generate role-normalized operation timing features. Obtain the current task stage identifier, retrieve the corresponding general basic threshold template from the stage-threshold association database, and perform role-adaptive transformation on the general basic threshold template based on the current role's role operation characteristic parameter set to generate a role-specific basic threshold template. The real-time operation deviation is compared with the role-specific basic threshold template. When the real-time operation deviation exceeds the corresponding threshold, the operation error judgment result is output.

2. The method for optimizing the dynamic threshold of UAV operation errors based on adaptive weighting according to claim 1, characterized in that, The baseline normalization process for the role adopts Z-score standardization. Based on the mean and standard deviation of the operation time sequence features of the current role under normal operation conditions, the current operation time sequence feature values ​​are standardized to obtain the normalized operation time sequence features of the role. The mean and standard deviation are parameters obtained in advance based on the normal operation samples of each role in historical task data. The normal operation samples are operation data segments of operators in historical tasks when they are not fatigued and have not made any operation errors.

3. The method for optimizing the dynamic threshold of UAV operation errors based on adaptive weighting according to claim 1, characterized in that, The role adaptation transformation is a process of differentially scaling the position deviation threshold component, maneuver deviation threshold component, and response deviation threshold component in the general basic threshold template based on the position accuracy weight, maneuverability weight, and response speed weight in the role operation characteristic parameter set. The threshold components of each dimension in the character-specific basic threshold template are the quotients obtained by dividing the threshold components of the corresponding dimensions in the general basic threshold template by the weight of the current character in that dimension.

4. The method for optimizing the dynamic threshold of UAV operation errors based on adaptive weighting according to claim 1, characterized in that, After generating the character-specific basic threshold template, it also includes: Detect whether the current moment is in the cognitive adaptation period after a role switch, where the cognitive adaptation period is the time required for the operator to transition from the previous role control mode to the current role control mode after a role switch occurs; During the cognitive adaptation period, a progressive convergence modulation process is applied to the character-specific basic threshold template. Based on the convergence coefficient that increases with the adaptation period, the character-specific basic threshold template corresponding to the previous character and the character-specific basic threshold template corresponding to the current character are linearly interpolated to generate the character transition threshold. The convergence coefficient is the ratio of the elapsed time from the moment of character switching to the preset adaptation period duration, and the value of the convergence coefficient is limited to between zero and one. During the cognitive adaptation period, the real-time operational deviation is compared with the role transition threshold; after the cognitive adaptation period ends, the real-time operational deviation is compared with the role-specific basic threshold template.

5. The method for optimizing the dynamic threshold of unmanned aerial vehicle (UAV) operation errors based on adaptive weighting according to claim 1, characterized in that, Also includes: Extract the attention load coefficient, control precision load coefficient, and monotonicity load coefficient corresponding to each role identifier from the role characteristic database, and associate the attention load coefficient, control precision load coefficient, and monotonicity load coefficient with the corresponding role segments to generate a role load feature vector sequence. For each role segment in the role load feature vector sequence, the flight time of the role segment is weighted according to the corresponding load coefficient, and the equivalent cumulative value of attention dimension, control dimension and cognition dimension are calculated by summing them to generate a multi-dimensional cumulative risk vector; wherein, the equivalent cumulative value of each dimension is the sum of the product of the flight time of each segment from the first role segment to the current role segment and the load coefficient of the corresponding role in that segment in that dimension.

6. The method for optimizing the dynamic threshold of unmanned aerial vehicle (UAV) operation errors based on adaptive weighting according to claim 5, characterized in that, Also includes: For the time-series features of the normalized operation of the role, the slope of change is calculated using a linear fitting algorithm within the short-term, medium-term, and long-term time windows, respectively, to generate a multi-scale trend vector. The multi-scale trend vector and the multi-dimensional cumulative risk vector are concatenated and input into a multi-dimensional fatigue pattern recognition network, which outputs independent fatigue probability scores for each dimension and a comprehensive fatigue development stage label. The multi-dimensional fatigue pattern recognition network includes a temporal feature encoding layer, a multi-dimensional attention layer, and a classification output layer. The temporal feature encoding layer encodes the input sequence using an LSTM network and outputs a temporal encoded representation. The multi-dimensional attention layer performs weighted attention processing on the temporal encoded representation based on the cumulative risk weights of the attention dimension, manipulation dimension, and cognition dimension, and outputs attention representations for each dimension. The classification output layer includes independent probability output heads for each dimension and a comprehensive fatigue stage classification head. The independent probability output heads for each dimension output the independent fatigue probability score for that dimension using a sigmoid activation function, and the comprehensive fatigue stage classification head outputs the comprehensive fatigue development stage label using a softmax activation function.

7. The method for optimizing the dynamic threshold of unmanned aerial vehicle (UAV) operation errors based on adaptive weighting according to claim 6, characterized in that, Also includes: Based on the current role identifier, the fatigue sensitivity dimension weights corresponding to the current role are retrieved from the role characteristic database. The independent fatigue probability scores of the attention dimension, manipulation dimension, and cognition dimension are aggregated by role sensitivity weighting to generate a comprehensive role sensitivity fatigue score. The comprehensive role sensitivity fatigue score is the sum of the product of the independent fatigue probability scores of each dimension and the fatigue sensitivity dimension weights of the current role in that dimension.

8. The method for optimizing the dynamic threshold of unmanned aerial vehicle (UAV) operation errors based on adaptive weighting according to claim 7, characterized in that, Also includes: The threshold pre-adjustment coefficient is calculated based on the character's sensitivity fatigue comprehensive score. The threshold pre-adjustment coefficient is the difference between the preset fatigue sensitivity adjustment coefficient and the product of the character's sensitivity fatigue comprehensive score. The threshold pre-adjustment coefficient is applied to the character's exclusive basic threshold template or the character's transition threshold for pre-tightening processing. The topology of inter-role collaboration relationships in the formation configuration data is analyzed, real-time status data of associated UAVs that have collaborative constraints with the current UAV are obtained, a collaborative constraint satisfaction index is calculated, a collaborative state adjustment coefficient is calculated based on the collaborative constraint satisfaction index, and the collaborative state adjustment coefficient is applied to the pre-tightened threshold to generate an adaptive weighted dynamic threshold. The real-time operation deviation is compared with the adaptive weighted dynamic threshold. When the real-time operation deviation exceeds the corresponding threshold, the operation error judgment result is output.

9. The method for optimizing the dynamic threshold of unmanned aerial vehicle (UAV) operation errors based on adaptive weighting according to claim 8, characterized in that, The calculation method of the collaborative constraint satisfaction index is as follows: For each associated UAV in the collaborative relationship topology that has a collaborative constraint relationship with the current UAV, calculate the normalized relative position deviation and the normalized relative speed deviation respectively. Then, sum the difference between the position constraint weight and speed constraint weight of each collaborative constraint relationship and the corresponding normalized deviation. Finally, take the average of the weighted summation results of all associated UAVs to obtain the collaborative constraint satisfaction index. The coordination state adjustment coefficient is determined as follows: when the coordination constraint satisfaction index is greater than or equal to the preset coordination normal threshold, the coordination state adjustment coefficient is one; when the coordination constraint satisfaction index is less than the coordination normal threshold, the coordination state adjustment coefficient is one minus the product of the preset coordination sensitivity adjustment coefficient and the difference between the coordination normal threshold and the coordination constraint satisfaction index. When the overall score of a character's sensitivity fatigue exceeds a preset character rotation trigger threshold, a character rotation suggestion is output. The fatigue warning level is output based on the comprehensive fatigue development stage label.

10. A dynamic threshold optimization system for unmanned aerial vehicle (UAV) operation errors based on adaptive weighting, used to execute the dynamic threshold optimization method for UAV operation errors based on adaptive weighting as described in any one of claims 1 to 9, characterized in that, include: The role history generation module is used to obtain the formation configuration data of multi-drone collaborative tasks, parse the role assignment identifier and role switching event log of each drone, extract the role switching time sequence and corresponding role identifier sequence of the current operator during the task execution process, and generate role history data. The role characteristic retrieval module is used to retrieve the set of role operation characteristic parameters corresponding to each role identifier in the role history data from the role characteristic database. The role normalization processing module is used to acquire the operation data stream within the current role segment, extract the operation response delay sequence, operation correction frequency sequence, and operation amplitude variation coefficient sequence, generate an operation timing feature set, and perform role baseline normalization processing on the operation timing feature set based on the current role's role operation characteristic parameter set to generate role normalized operation timing features. The role threshold generation module is used to obtain the current task stage identifier, retrieve the corresponding general basic threshold template from the stage-threshold association database, and perform role adaptation transformation on the general basic threshold template based on the current role's role operation characteristic parameter set to generate a role-specific basic threshold template. The operation error determination module is used to compare the real-time operation deviation with the role-specific basic threshold template. When the real-time operation deviation exceeds the corresponding threshold, the operation error determination result is output.

Images