An adversarial design training method

Abstract

The present application relates to the technical field of simulation training, in particular to a kind of countermeasure design training method, comprising the following steps: monitoring training state obtains time series data input hidden Markov generates behavior sequence, detects coordinate offset angle deviation consumption ratio identifies weak link, calls identification table and matches factor library to generate interference scheme, according to coefficient threshold control activation timing particle swarm optimization generates dynamic sequence, according to frequency coefficient injection adjustment rhythm monitors response delay switching parameter group generates evaluation result, in the present application, behavior sequence dynamic tracking task performance is generated by collecting training state, weak link is calculated and identified in combination with coordinate offset trajectory angle difference and resource consumption rate, hierarchical selection is realized in interference link according to risk parameter combination, in triggering, periodically match and combine optimization algorithm dynamic adjustment, so that disturbance intensity trigger rhythm adaptive continuous optimization, improve pressure performance and flexible response ability.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of simulation training, in particular to an adversarial design training method. BACKGROUND

[0002] The technical field of simulation training involves virtual training and simulation exercises of target objects in specific environments. The core matters include simulating the behavior of training objects by establishing virtual environments or models, reproducing complex scenarios, and providing controllable variables for training and verification under different conditions. This technical field covers scene modeling, environment reconstruction, data generation, and action interaction, and overall forms a systematic training method based on virtual environment construction and focusing on controllable training processes.

[0003] Among them, the traditional adversarial design training method refers to designing specific adversarial samples or adversarial scenarios for training objects in the simulation training process to force them to respond under non-ideal conditions. The technical matter it aims at is how to strengthen the training of training objects through artificially constructed interference conditions. The traditional adversarial design training method achieves this by generating adversarial data based on sample disturbance or by artificially setting non-equilibrium task conditions, thereby forming an adversarial environment and task input that can be used to test and train objects in the simulation training scenario.

[0004] The existing technology relies on fixed sample disturbance or non-equilibrium conditions to set up adversarial scenarios. The state changes of training objects during execution are not captured in detail, resulting in a lack of targeted recognition of task phase characteristics and weak links. The risk level is not distinguished and the timing is not flexibly adjusted in the interference application, leading to a single disturbance method that is difficult to adapt to changes in the task process. When training objects perform complex tasks, the interference effect may not match the actual task requirements, resulting in insufficient dynamic adaptability and targeting of training results, which affects the stability of the overall training effect. SUMMARY

[0005] To solve the technical problem that the existing technology relies on fixed sample disturbance or non-equilibrium conditions to set up adversarial scenarios, the state changes of training objects during execution are not captured in detail, resulting in a lack of targeted recognition of task phase characteristics and weak links. The risk level is not distinguished and the timing is not flexibly adjusted in the interference application, leading to a single disturbance method that is difficult to adapt to changes in the task process. When training objects perform complex tasks, the interference effect may not match the actual task requirements, resulting in insufficient dynamic adaptability and targeting of training results, which affects the stability of the overall training effect, the present application embodiment provides an adversarial design training method, comprising the following steps:

[0006] In order to achieve the above object, the present application adopts the following technical scheme: An anti-design training method, comprising the following steps:

[0007] S1: Obtain task node timing data by monitoring the execution state of the training subject, collect node start time, node end time and node interdependence relationship identifier, input the data into a hidden Markov model for state transition calculation, and generate a training subject behavior state sequence;

[0008] S2: Call the training subject behavior state sequence, obtain the spatial coordinate offset, the action trajectory deviation angle and the resource consumption rate by detection and weight distribution, respectively perform the Euclidean distance, the angle difference and the consumption ratio operation, and form a training subject weak link identification table;

[0009] S3: Call the training subject weak link identification table, obtain the interference parameter group in the countermeasure factor library, select the reinforced interference parameter group when the identification table component is greater than the risk trigger threshold, select the maintenance interference parameter group when the component is less than the stable operation threshold, and generate a targeted countermeasure interference scheme;

[0010] S4: Call the targeted countermeasure interference scheme, based on the matching time of the interference duration coefficient and the task execution cycle parameter in the scheme, activate the countermeasure factor when the coefficient is greater than the implantation opportunity threshold, and delay activation when it is less than the delay implantation threshold, input the scheme into the particle swarm optimization algorithm for optimization, and generate a dynamic countermeasure training sequence.

[0011] As a further scheme of the present application, the Euclidean distance and angle difference calculation step is combined with threshold determination and particle swarm optimization algorithm, which can adapt to the interference characteristics in real time in dynamic training, and improve the anti-interference ability of the training subject;

[0012] The training subject behavior state sequence includes state transition path, state switching frequency and state duration, the training subject weak link identification table includes spatial offset, action trajectory deviation and resource consumption rate, the targeted countermeasure interference scheme includes interference start time, interference duration and interference intensity configuration, and the dynamic countermeasure training sequence includes interference injection timing, interference intensity curve and interference adjustment interval.

[0013] As a further scheme of the present application, the specific steps of S1 are:

[0014] S101: Monitor the task node execution state of the training subject, collect the node start time and end time, and record the node interdependence relationship identifier, format the node start and end time for each node respectively, sort them according to time sequence according to the node interdependence relationship, and generate a node timing data set;

[0015] S102: based on the node timing data set, adjacent node time difference value is calculated and matching dependence relation mark is adopted, state transition probability matrix of hidden Markov model is used to calculate state transition probability, and timing transition probability value is obtained;

[0016] The timing transition probability value refers to the value of state transition probability of the training subject between adjacent nodes, which is calculated based on the node timing data set and the dependence relation mark, and the state transition probability matrix of the hidden Markov model, and the range is [0, 1] and the sum of transition probabilities in the same state is 1;

[0017] S103: according to the timing transition probability value, the hidden Markov model is used to infer the behavior state sequence of the training subject, the transition probability between nodes is calculated, the node path is gradually deduced combining the state space information, and the behavior state sequence of the training subject is generated.

[0018] As a further scheme of the application, the specific steps of S2 are:

[0019] S201: calling the training subject behavior state sequence, detecting the spatial coordinate data, comparing the actual coordinates of adjacent action points with the reference coordinates one by one, calculating the offset between the two by using the Euclidean distance formula, and merging the offset results of the action points to generate the coordinate offset interval;

[0020] S202: according to the coordinate offset interval, calling the action trajectory data of the training subject, analyzing the included angle between the trajectory vector and the reference trajectory vector, calculating the angle difference of each trajectory segment by using the angle difference method, and obtaining the action trajectory deviation angle degree by weighted average of the trajectory segment angle difference;

[0021] S203: based on the action trajectory deviation angle degree, calling the energy and resource consumption rate data of the training subject, calculating the weighted ratio by weight distribution, combining the angle difference, coordinate offset and consumption rate of each data, and sorting to generate the training subject weak link identification table.

[0022] As a further scheme of the application, the specific steps of S3 are:

[0023] S301: obtaining the component data in the training subject weak link identification table, corresponding the component data with the interference parameter groups of the counter factor library one by one, comparing the component values with the risk trigger threshold and the stable operation threshold, recording the corresponding threshold state, and generating the threshold comparison value;

[0024] S302: Call the threshold comparison value, compare the components that are within the risk trigger threshold range and select the enhanced interference parameter group, compare the components that are within the stable operation threshold range and select the maintenance interference parameter group, filter out the components that do not fall into any threshold range, and obtain the number of parameter groups to be filtered.

[0025] S303: The parameter group selection quantity is called to perform merging in the interference parameter group set, the enhanced interference parameter group and the maintenance interference parameter group are classified and integrated, and the parameters are mapped according to the integrated structure to generate a targeted anti-interference scheme.

[0026] As a further aspect of the present invention, the threshold comparison value is the corresponding value obtained by comparing the component data with the preset threshold item by item.

[0027] The risk trigger threshold is a threshold set based on the boundary value of the component corresponding to the weak link of the training subject.

[0028] The stable operation threshold is a threshold set based on the normal range value of the component corresponding to the weak link of the training subject.

[0029] The number of parameter groups to be screened is the number of interference parameter groups corresponding to components that fall within the range of the risk trigger threshold or stable operation threshold after threshold comparison.

[0030] As a further aspect of the present invention, the specific steps of S4 are as follows:

[0031] S401: Invoke the targeted anti-interference scheme. Based on the interference duration coefficient and the task execution cycle parameter, if the coefficient is greater than the implantation timing threshold, trigger the interference factor activation command. If it is less than the delayed implantation threshold, trigger the delayed activation command, record the mark, and generate the interference matching judgment value.

[0032] The implantation timing threshold refers to the time domain boundary set based on the interference duration coefficient and the task execution cycle parameter, which is used to determine whether the interference factor should be activated immediately or delayed.

[0033] The interference matching judgment value refers to a quantitative index calculated based on the interference duration coefficient, task cycle parameters, and trigger / delay markers, which is used to characterize the degree of matching between interference features and task cycle.

[0034] S402: Based on the interference matching judgment value, call the trigger flag and the delay flag, map them to the activation instruction and the delay instruction, and generate a factor activation instruction set according to the task cycle analysis and adjustment instruction adaptation.

[0035] S403: Based on the factor activation instruction set, input it into the particle swarm optimization algorithm, adjust the instruction weights through algorithm iteration, optimize the weight coefficients according to the task cycle, perform dynamic adaptation of instructions and cycle, and generate dynamic adversarial training sequences.

[0036] As a further aspect of the present invention, the method further includes step S5:

[0037] S5: Call the dynamic adversarial training sequence, calculate the injection interval time according to the interference frequency coefficient and the remaining time, adjust the transmission rhythm of the interference signal, monitor the response delay time deviation index of the training subject, and when the deviation index exceeds the strategy adjustment threshold, switch the backup parameter group of the adversarial scheme to perform injection, and generate the adversarial training enhancement effect evaluation result.

[0038] The evaluation results of the adversarial training enhancement effect include response delay time, adaptive change rate, and risk assessment index.

[0039] As a further aspect of the present invention, the specific steps of S5 are as follows:

[0040] S501: Call the dynamic adversarial training sequence, perform weighted processing based on the interference frequency coefficient and the remaining time value, allocate the weighted value to the signal injection period and set the interval, and generate the injection interval time value.

[0041] S502: Based on the injection interval time value, the rhythm of the interference signal transmission is adjusted, the response delay time of the training subject is monitored during the training operation, the difference between the delay time and the set response benchmark time is calculated and converted into a deviation index to obtain the delay time deviation index.

[0042] S503: Compare the delay time deviation index with the strategy adjustment threshold. When the deviation index exceeds the threshold, switch the backup parameter group and re-execute the signal injection. After numerically reorganizing the interference signal features, calculate the training performance and obtain the evaluation result of the adversarial training enhancement effect.

[0043] As a further aspect of the present invention, the injection interval time value is a time parameter obtained by weighting the interference frequency coefficient and the remaining time value and allocating it to the signal injection period;

[0044] The delay time deviation index is a numerical parameter calculated from the difference between the actual response delay time and the set reference response time.

[0045] The strategy adjustment threshold is a preset numerical limit parameter compared with the delay time deviation index.

[0046] Compared with the prior art, the advantages and positive effects of the present invention are as follows:

[0047] In this invention, by collecting the execution status of the training process in a time sequence and generating a behavioral state sequence, the performance of the task at different stages is dynamically tracked. By combining the comprehensive calculation of coordinate offset, trajectory angle difference and resource consumption rate, multi-dimensional identification of weak links is achieved. In the interference application stage, parameter combinations are distinguished according to the degree of risk, so that the interference effect has stratification and selectivity. At the same time, in the triggering process, timing is matched according to the task cycle and dynamic adjustment is achieved by combining optimization algorithms. Thus, the training process is continuously optimized in terms of disturbance intensity, triggering rhythm and adaptability, so that the training object's stress resistance and flexible response ability under complex conditions are steadily improved. Attached Figure Description

[0048] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0049] Figure 1 This is a schematic diagram of the steps of the present invention;

[0050] Figure 2 This is a detailed schematic diagram of S1 of the present invention;

[0051] Figure 3 This is a detailed schematic diagram of S2 of the present invention;

[0052] Figure 4 This is a detailed schematic diagram of S3 of the present invention;

[0053] Figure 5 This is a detailed schematic diagram of S4 of the present invention;

[0054] Figure 6 This is a detailed schematic diagram of S5 of the present invention. Detailed Implementation

[0055] The technical solution of the present invention will now be described with reference to the accompanying drawings.

[0056] In embodiments of the present invention, words such as "exemplarily," "for example," etc., are used to indicate that something is an example, illustration, or description. Any embodiment or design described as "exemplary" in the present invention should not be construed as being more preferred or advantageous than other embodiments or designs. Specifically, the use of the word "exemplary" is intended to present the concept in a concrete manner. Furthermore, in embodiments of the present invention, the meaning expressed by "and / or" can be both, or either one.

[0057] In the embodiments of this invention, the terms "image" and "picture" may sometimes be used interchangeably. It should be noted that, without emphasizing the distinction between them, they convey the same meaning. Similarly, the terms "of," "corresponding (relevant)," and "corresponding" may sometimes be used interchangeably. It should be noted that, without emphasizing the distinction between them, they convey the same meaning.

[0058] In this embodiment of the invention, sometimes a subscript such as W1 may be written in a non-subscript form such as W1. When the difference is not emphasized, the meaning they express is the same.

[0059] To make the technical problems, technical solutions and advantages of the present invention clearer, a detailed description will be given below in conjunction with the accompanying drawings and specific embodiments.

[0060] Please see Figure 1 This invention provides an adversarial design training method, comprising the following steps:

[0061] S1: By monitoring the execution state of the training subject, the timing data of task nodes is obtained, the start time of nodes, the end time of nodes, and the dependency relationship identifiers between nodes are collected. The data is input into the Hidden Markov Model for state transition calculation to generate the behavioral state sequence of the training subject.

[0062] S2: Call the training subject's behavior state sequence, obtain spatial coordinate offset, action trajectory deviation angle, and resource consumption rate by detection and weight allocation, and perform Euclidean distance, angle difference and consumption ratio calculations respectively to form a training subject's weak link identification table;

[0063] S3: Call the weak link identification table of the training subject, obtain the interference parameter group in the adversarial factor library, select the enhanced interference parameter group when the component of the identification table is greater than the risk trigger threshold, select the maintenance interference parameter group when the component is less than the stable operation threshold, and generate a targeted adversarial interference scheme.

[0064] S4: Invoke the targeted adversarial interference scheme. Based on the timing of matching the interference duration coefficient and the task execution cycle parameter in the scheme, activate the adversarial factor when the coefficient is greater than the implantation timing threshold, and activate it after the delay when it is less than the delay implantation threshold. Input the scheme into the particle swarm optimization algorithm for optimization and generate a dynamic adversarial training sequence.

[0065] S5: Call the dynamic adversarial training sequence, calculate the injection interval based on the interference frequency coefficient and the remaining time, adjust the transmission rhythm of the interference signal, monitor the response delay time deviation index of the training subject, and switch the backup parameter group of the adversarial scheme to perform injection when the deviation index exceeds the policy adjustment threshold, and generate the evaluation result of the adversarial training enhancement effect.

[0066] The training subject's behavioral state sequence includes state transition path, state switching frequency, and state duration; the training subject's weak link identification table includes spatial offset, action trajectory deviation, and resource consumption rate; the targeted adversarial interference scheme includes interference initiation timing, interference duration, and interference intensity configuration; the dynamic adversarial training sequence includes interference injection timing, interference intensity curve, and interference adjustment interval; and the adversarial training enhancement effect evaluation results include response delay time, adaptive change rate, and risk assessment index.

[0067] Please see Figure 2 The specific steps of S1 are as follows:

[0068] S101: Monitor the execution status of the main training task nodes, collect the start and end times of the nodes, record the dependency relationship identifiers between the nodes, perform formatting processing on the start and end times of each node, sort them in chronological order according to the node dependencies, and generate a node time series dataset.

[0069] The system monitors the execution status of the main training task nodes. The start and end times of each node are sampled at the millisecond level using the system's built-in high-precision time acquisition module. The generated raw timestamps are stored in a sequence with the structure YYYY-MM-DD-HH:MM:SS.mmm. For example, when monitoring node A, the timestamp for its start signal is recorded as 2025-08-25-09:00:00.125, and the timestamp for its end signal is recorded as 2025-08-25-09:05:30.430. This information is collected during node acquisition. Afterwards, a unified time zone correction is performed, converting the time values ​​to Coordinated Universal Time (UTC). Based on the collected start and end times, the execution duration of each node is calculated. The execution duration is calculated by subtracting the start time from the end time. For example, the execution duration of node A = 09:05:30.430 - 09:00:00.125 = 5 minutes 30.305 seconds, or 330.305 seconds. Then, the start and end times of the nodes are sorted in ascending order according to their timestamp values. For any two nodes whose times need to be compared during the sorting process... To retrieve the start time value of node 1 (e.g., node A), a direct comparison is made with the start time value of node 2 (e.g., node B). If the start time value of node 1 is less than that of node 2, then node 1 is prioritized over node 2; otherwise, their order is swapped. During the comparison process, nodes with dependencies are marked, for example, "D1→D2" indicates that node D2 depends on node D1 to start. This dependency marker is obtained by analyzing the task scheduling log. For the dependency description field in each log entry, the predecessor node number and successor node number are extracted and formatted... Create a dependency array, such as {[A, B], [B, C], [A, D]}. In the sorting result, first check if A is already sorted and precedes B in the first group [A, B]. If not, force the order to satisfy the dependency relationship. Repeat this step for the dependency pairs until the sorting is stable. Generate a structured table dataset by mapping the processed node time order to the dependency relationship identifier. Each row of the table contains the node number, start time (UTC), end time (UTC), execution duration (seconds), and dependency identifier. Its data format is shown in Table 1.

[0070] Table 1: Node Timing Acquisition and Dependency Relationship Table

[0071]

[0072] As shown in Table 1, the timestamp data collected is first formatted at the millisecond level, then uniformly converted to UTC time base. The precise execution duration is then obtained by subtracting the start time from the end time. Nodes are continuously sorted in ascending order and their dependencies adjusted. The resulting node time-series dataset provides accurate input for subsequent time difference calculations and state inference. The actual execution steps of this dataset formation process are analogous to monitoring the start and stop times of each device on a production line. For example, device 1 runs for 5 minutes and 30 seconds before triggering device 2 to run, and device 3 must start only after device 2 has finished. The start and end times and dependency order in this process are highly consistent with the monitoring of the main task nodes during training.

[0073] S102: Based on the node time series dataset, calculate the time difference between adjacent nodes and match the dependency identifier. Use the state transition probability matrix of the hidden Markov model to calculate the state transition probability and obtain the time series transition probability value.

[0074] Based on the node time-series dataset, the sorted node list is read sequentially. The end time of the previous node is subtracted from the start time of the next node. The result is converted to seconds and rounded to three decimal places. For example, the end time of node A is... The start time of node B is The difference between the two is The end time of node B is The start time of node C is The difference is The end time of node A is The start time of node D is The difference is The time difference array formed by the difference is matched one by one with the dependency identifiers of the records. When the predecessor node in the dependency identifier is the same as the source node of the time difference value, and the successor node is the same as the next node, it is marked as a real dependency delay; otherwise, it is marked as a non-dependency interval, such as a time difference. Corresponding dependency Marked as a true dependency delay, but not present in the dependency table. If the time difference is not specified, it is classified as an independent interval. After labeling, it is necessary to calculate the ratio of the number of transitions of each type of node in the training set to the total number of transitions of the predecessor node to obtain the probability value. For example, in the record... The transfer action occurred Node A's subsequent startup events totaled [number] times. Next, its transition probability is , The transfer action occurred Node B's subsequent startup events totaled [number] times. Next, its transition probability is , The transfer action occurred Node A's subsequent startup events totaled [number] times. Next, its transition probability is Then, the probability values ​​are classified and judged according to preset intervals. When the probability value is within a certain range... Interval time is defined as low probability, in Interval time is defined as medium probability, in Intervals are defined as high probability, for example and This falls within the high probability range. It belongs to the medium probability range. Finally, the node transition pairs and their probability values ​​are generated into a probability matrix data structure. Each row corresponds to the predecessor node and each column corresponds to the successor node. The value in the cell of the matrix is ​​the statistical probability of the transition pair.

[0075] S103: Based on the temporal transition probability value, a hidden Markov model is used to infer the behavioral state sequence of the training subject. By calculating the transition probability between nodes and combining the state space information, the node path is gradually deduced to generate the behavioral state sequence of the training subject.

[0076] Based on the temporal transition probability values, nodes are treated as independent state points in the state space. Starting states are sequentially retrieved from the sorted node list, and the next state is selected based on the probability value of the corresponding row in the matrix. For example, if the node set in the state space is... In the calculation results , , The derivation uses node A as the initial state and calls its corresponding set of transition probabilities. With successor node labels To compare, first sort the probability values ​​by size. Therefore, it is determined that the primary transition target of node A is B. After entering node B, the set of transition probabilities for the corresponding row of B in the matrix is ​​read, for example... Compared with the probability values ​​of other successor nodes, because Greater than the set upper limit of medium probability If a state transition is determined to be high-probability, then C is set as the next state. The process then proceeds to node C, and based on its probability set and dependency constraints, the next state is determined sequentially. If the transition probability values ​​of any current node are all below the low-probability upper limit... If the state chain terminates, it is considered terminated. Before each state transition, a Boolean check is needed to determine whether the target node has already appeared in the path sequence. If it has, the avoidance loop is skipped; otherwise, it is included in the sequence. Throughout the path derivation process, for transitions with multiple probability values, a secondary sorting rule needs to be invoked. This rule prioritizes nodes based on the absolute value of their time differences. For example, if two candidate targets have the same probability value... When comparing their time differences, the node with the smaller time difference is selected first. Assume the time difference from node A to D is... The time difference from A to E is If the transition from A to D is preferred, then the state sequence can be obtained by starting from A according to the aforementioned rules. If the condition for a direct transition from A to D in the test set data satisfies the probability interval judgment and the path does not repeat, then another path will be formed. The effective paths derived multiple times form a set of behavioral state sequences, providing a clear time and state mapping relationship for the dynamic changes of the training subject in the entire task node network.

[0077] Please see Figure 3 The specific steps of S2 are as follows:

[0078] S201: Call the training subject's behavior state sequence, detect the spatial coordinate data, compare the actual coordinates of adjacent action points with the reference coordinates one by one, calculate the offset between the two using the Euclidean distance formula, and merge the offset results of the action points to generate a coordinate offset range.

[0079] The training subject's behavioral state sequence is invoked, and the spatial coordinate values ​​of each action state point are read. These coordinates are represented in a three-dimensional structure. The unit is meters. The sensor module detects and records the actual coordinates of each action point point by point, while simultaneously using the reference coordinates of the corresponding action point in the reference database as a comparison basis. The actual coordinates of each action point are then subjected to a dimension-by-dimensional difference calculation with the reference coordinates to obtain the result. , , The displacement components in three dimensions are summed, and the square root of the sum of the squares of the three-dimensional components is taken to obtain the offset value. For example, the coordinates of the first action point actually acquired are... The reference coordinates are The three-dimensional differences are respectively , , The squared differences are respectively Summing them together yields The square root yields the offset of that point as approximately For example, the actual coordinates of the second action point are... The reference coordinates are The three-dimensional differences are respectively The square value is also Three, summation , take the square root to get Thus, the offset sequence is calculated for the action points. Then, a merging process is performed on the sequence. This involves first setting a segmentation threshold for the offset interval, and then dividing the offset into segments based on the actual scenario. Defined as a low offset region Defined as the middle offset region, Defined as a high offset zone, the judgment rule is to directly compare the offset with the upper and lower limits of the interval. Those that meet the range are assigned to the corresponding group. For example, the offsets of the two action points mentioned above... Belonging to the middle offset zone, after repeatedly merging the action points, we obtain a set of middle offset, low offset, and high offset, forming coordinate offset range data, which provides input data for subsequent motion trajectory deviation calculation.

[0080] Table 2: Calculation Table of Spatial Motion Point Coordinates and Offsets

[0081]

[0082] As shown in Table 2, by comparing the actual three-dimensional coordinates and reference coordinates of each action point, the three-dimensional difference, square, sum, and square root are calculated to obtain the offset. Based on the preset interval, the offset is directly judged and classified, thus forming the statistical segmentation result of the coordinate offset.

[0083] S202: Based on the coordinate offset range, call the motion trajectory data of the training subject, analyze the angle between the trajectory vector and the reference trajectory vector, calculate the angle difference of each trajectory segment using the angle difference method, and obtain the motion trajectory deviation angle by weighted averaging of the angle differences of the trajectory segments.

[0084] Obtain the coordinate offset range, retrieve the training subject's motion trajectory data, and sequentially read the trajectory segment vectors composed of consecutive motion points. Each trajectory segment vector is formed by the 3D coordinate difference between adjacent motion points. For example, starting from motion point 1... To action point 2 The corresponding trajectory vector components are The vectors of the reference trajectory within the same interval are calculated from the reference coordinates, starting from reference point 1. To reference point 2 Corresponding components are When calculating the included angle, first find the lengths of the two vectors separately, square the components, sum them, and then take the square root to obtain the vector magnitude. For example, the actual vector magnitude is:

[0085] Right now Approximately ;

[0086] The reference vector magnitude is Right now Approximately ;

[0087] Then, the dot product of the two vectors is calculated, which is obtained by multiplying the corresponding components and then summing them:

[0088] and Then divide the dot product value by the product of the magnitudes of the two vectors. Approximately equal to Taking the inverse cosine, the included angle is approximately This process is repeated for each subsequent trajectory segment to form a set containing all included angle values. Then, a weighted average angle deviation is calculated, with the weights set based on the interval of the coordinate offset corresponding to each segment. For example, low offset regions are assigned higher weights. Weights are assigned to the offset region. Weighting of high offset regions Taking the trajectory segment from action point 1 to action point 2 as an example, its offset is... It belongs to the middle offset region, corresponding to the weight. If the angle difference of this segment is The weighted value is The weighted angle values ​​of the trajectory segments are summed and then divided by the sum of the corresponding weights to obtain the overall weighted average deviation angle of the motion trajectory.

[0089] S203: Based on the angle of the motion trajectory deviation, call the energy and resource consumption rate data of the training subject, calculate the weighted ratio by weighting the angle difference, coordinate offset and consumption rate of each data item, sort them and generate a weak link identification table of the training subject.

[0090] Based on the weighted average deviation angle of the action trajectory, the energy consumption rate and resource consumption rate data corresponding to each action node in the training subject's behavioral state sequence are sequentially retrieved. The energy consumption rate is obtained through real-time monitoring equipment, such as the number of kilojoules consumed per unit time, while the resource consumption rate is statistically analyzed from work records, such as the number of kilograms of material consumed per unit time. Each data point is directly correlated with the angle deviation value and coordinate offset of that node, and then weighted. The weighting follows the principle of importance allocation for three types of indicators. Assuming the angle deviation weight is set to... The coordinate offset weight is set to Consumption rate weight is set to Taking action node 1 as an example, its angular deviation is The coordinate offset is The energy consumption rate is The resource consumption rate is During the calculation, the three values ​​are first normalized to make them normal. The interval, for example, the maximum value of the angle deviation among all nodes is Then the normalized angle deviation of node 1 is The maximum value of the coordinate offset among all nodes is Then the normalized coordinate offset is The maximum energy consumption rate is Then the normalized energy consumption rate is The maximum resource consumption rate is Then the normalized resource consumption rate is After obtaining the normalized values, multiply them by their corresponding weights and sum them up. The weighted ratio for node 1 is... ,Right now After calculating the weighted ratio of the nodes using the same steps, the values ​​are sorted from highest to lowest. The nodes at the top are identified as weak links. For example, if the weighted ratio of node 3 is... Node 2 is Node 1 is If the sorting result is node3>node2>node1, then this sorting constitutes the data content of the training subject's weak link identification table.

[0091] Please see Figure 4 The specific steps of S3 are as follows:

[0092] S301: Obtain component data from the weak link identification table of the training subject, match the component data with the interference parameter group of the adversarial factor library item by item, compare the component values ​​with the risk trigger threshold and stable operation threshold, record the corresponding threshold status, and generate threshold comparison values.

[0093] To obtain the weak link identification table for the training subject, first read the value of each component in the table sequentially. Here, "component" refers to the weighted ratio of the node. For example, based on the sorting result, the weighted ratio of node 3 is [value missing]. Node 2 is Node 1 is Then, the pre-defined interference parameter groups in the adversarial factor library are called, ensuring that each component has a corresponding interference parameter group number in the table. For example, node 3 corresponds to parameter group ID P7, node 2 corresponds to P4, and node 1 corresponds to P2. Then, the component values ​​are directly compared one by one with two types of thresholds: the risk trigger threshold and the stable operation threshold. In practical scenarios, it can be set to (This was obtained through statistical analysis of raw data showing that component indicators in previous tasks led to task instability, and is currently in a state of...) (Above and above can be considered high risk probability) Stable operation threshold Can be set to (Under actual testing conditions, values ​​at or above this level up to below the risk trigger threshold can be considered as stable operation.) When performing comparisons, for example, the weighted ratio of node 3... and If the result is greater than this value, it is marked as "triggered risk threshold" and the comparison value status code is recorded as 1; Node 2 and The ratio is less than, and If the ratio is greater than 2, it is marked as "at a stable threshold" and the status code is recorded as 2; Node 1 At the same time less than and If the value is not within any threshold range, it is marked as "not within any threshold range" and the status code is recorded as 0. During this process, each comparison involves directly comparing the component value with the upper and lower threshold limits (greater than or less than), and writing the corresponding result to the status flag field. Simultaneously, a threshold comparison value sequence is generated and stored as a numerical array. Each element consists of the component ID, the actual component value, and the status code. For example... .

[0094] S302: Call the threshold comparison value, compare the components that are within the risk trigger threshold range and select the enhanced interference parameter group, compare the components that are within the stable operation threshold range and select the maintenance interference parameter group, filter out the components that do not fall into any threshold range, and obtain the number of parameter groups to be filtered.

[0095] The threshold comparison numerical sequence is invoked, and the status code field of each element is read sequentially. First, it is determined whether the status code is equal to 1. If so, the corresponding interference parameter group is added to the "enhanced interference parameter group" set. For example, the record with component ID 3 corresponds to status code 1, and its parameter group is P7. P7 is added to the enhanced set. Then, it is determined whether the status code is equal to 2. If so, its corresponding parameter group is added to the "maintained interference parameter group" set. For example, the record with component ID 2 corresponds to status code 2, and its parameter group is P4. P4 is added to the maintenance set. When the status code is equal to 0, a filtering operation is performed, that is, the component is skipped directly, and its corresponding parameter group is not recorded. For example, the status code of component ID 1 is 0, and its parameter group P2 is excluded. During the filtering process, the integer value of the status code is compared sequentially with... The logical judgment of specific values ​​1 or 2 is all done by traversing each element one by one, checking each item, and storing the parameter groups that meet the conditions into the corresponding sets. The sets are stored in memory as arrays and automatically deduplicated to ensure that the same parameter group appears only once in the set. For example, after a complete traversal, the set of enhanced interference parameter groups is {P7}, and the set of maintained interference parameter groups is {P4}. The total number of selections is obtained by directly summing the lengths of the two sets. The length of set {P7} is 1, and the length of set {P4} is 1. The sum gives a total of 2. This value is used as the input condition for the subsequent interference parameter group merging process. In the storage process, the original association information of the component ID is retained so that the corresponding node of the original weak link identification table can be accurately traced when the parameter mapping is executed in the next step.

[0096] S303: The number of parameter groups to be filtered is called and merged in the set of interference parameter groups. The enhanced interference parameter groups and the maintenance interference parameter groups are classified and integrated. The parameters are mapped according to the integrated structure to generate a targeted anti-interference scheme.

[0097] To filter the number and content of parameter sets, first establish two target sets: one to store the enhanced interference parameter set and the other to store the maintenance interference parameter set. Then, the enhanced set... Copying directly to the reinforcement set will maintain the set The data is copied directly to the maintenance set, and then a merge operation is performed. The two sets are concatenated into a unified set of parameter groups as arrays. During the merge process, the category label of each parameter group is recorded. For example, P7 is recorded as "enhancement" and P4 is recorded as "maintenance". Then, the data is classified and integrated according to the labels to ensure that different categories have clear partitions in the set. Subsequently, parameter mapping is performed to map each parameter group to its corresponding component ID and the node number in the original weak link identification table. For example, P7 comes from component ID 3 and corresponds to node 3, with a weighted ratio of 1. This is a risk-triggered threshold state; P4 originates from component ID2, corresponding to node 2, and the weighted ratio is... If the system is in a stable operating threshold state, the mapping information is then paired with specific parameter values ​​stored in the interference factor library. For example, the interference parameter for P7 is the force load coefficient. The interference parameter for P4 is the motion rhythm coefficient. This results in a merged and integrated mapping table, as shown in Table 3.

[0098] Table 3: Merging and Mapping Table for Interference Parameter Groups

[0099]

[0100] As shown in Table 3, after merging and classification, each parameter group has a clear mapping relationship with the original component and the corresponding node, and the specific numerical parameters to be targeted are given in the interference parameter values. This structure is the data entity of the targeted anti-interference scheme.

[0101] Please see Figure 5 The specific steps of S4 are as follows:

[0102] S401: Invoke the targeted anti-interference scheme. Based on the interference duration coefficient and the task execution cycle parameter, if the coefficient is greater than the implantation timing threshold, the interference factor activation command is triggered. If it is less than the delayed implantation threshold, the delayed activation command is triggered. Record the mark and generate the interference matching judgment value.

[0103] To invoke a targeted anti-interference scheme, first, the associated parameter group information and the execution data of the corresponding task node are read one by one from the scheme list. Then, the interference duration coefficient bound to that parameter group and the task execution cycle parameter of that node are retrieved from the interference factor library. The interference duration coefficient is expressed in seconds and is obtained by averaging the interference duration intervals for that parameter group. For example, for parameter group P7, its original duration record is... The interference duration coefficient is obtained by averaging the seconds. The task execution cycle parameter is the average time taken for that task node from start to finish in a single task, for example, the execution time of node 3 over five executions:

[0104] Second;

[0105] Take the average Then, a preset implantation timing threshold is compared with a delayed implantation threshold. Set as 25% of the task execution cycle, delay implantation threshold Assuming it accounts for 15% of the task execution cycle, and calculating using a task execution cycle of 30.08 seconds, we can obtain... , During the judgment process, the interference duration coefficient is compared with two thresholds one by one. If it is greater than the implantation timing threshold, a trigger interference factor activation command flag is recorded (trigger flag value is set to 1). If it is less than the delayed implantation threshold, a delayed activation command flag is recorded (delay flag value is set to 2). Otherwise, it is recorded as a normal implantation flag (flag value is set to 0). For example, the interference duration coefficient of P7 is 8.066 seconds, which is greater than the implantation timing threshold of 7.520 seconds, so the flag is assigned as 1. Another example is P4, assuming its interference duration coefficient is taken from... The average duration is 4.033 seconds, which is less than the delayed implantation threshold of 4.512 seconds. Therefore, it is marked as 2. Through such calculation and comparison, a set of interference matching judgment data sequence is generated for the parameter group, in the form of {(parameter group ID, duration coefficient, execution cycle, mark value)}.

[0106] Table 4: Comparison Results of Interference Duration and Threshold

[0107]

[0108] As shown in Table 4, the actual interference duration coefficient is obtained by sequentially calculating the average value, and then directly compared with the implantation and delay thresholds to generate a clear judgment mark, providing a quantitative basis for the mapping of activation or delay instructions in the next step.

[0109] S402: Based on the interference matching judgment value, call the trigger flag and the delay flag, map them to the activation command and the delay command, analyze and adjust the command adaptation according to the task cycle, and generate the factor activation command set;

[0110] The interference matching judgment value is invoked, and the judgment flag value of each parameter group is read one by one and a mapping operation is performed. When the judgment flag value is equal to 1, the parameter group is directly mapped to the activation instruction set, and the parameter group number and the corresponding task node number are recorded in the set. For example, if the P7 flag value is 1, the mapping result is activation instruction (P7, node 3). When the judgment flag value is equal to 2, the parameter group is mapped to the delay instruction set, and its associated task node number is also recorded. For example, if the P4 flag value is 2, it is mapped to delay instruction (P4, node 2). During the mapping process, two empty sets are first created and named activation... The system maps data to a set and a deferred set, and determines the set based on a flag value during data traversal. If the flag value is 0, the data is ignored and not added to any set. After mapping, the elements of the two sets are adapted according to the task execution cycle. This involves comparing the task cycle parameters of a parameter group in a set with the average cycle of all cycles in that set. If the task cycle of a parameter group is greater than 1.10 times the average cycle of the set, the instruction is moved to the end of the set for delayed execution; conversely, if it is less than 0.90 times the average cycle of the set, it is moved to the beginning of the set for execution. For example, the task cycle of P7 in the activation set is... Since the set contains only this element, the average value is also 1. The ratio is exactly 1.00, requiring no adjustment; the P4 task cycle in the delayed set is... The set also contains only this element, and its ratio is 1.00. It also does not need to be adjusted. The final generated factor activation instruction set is composed of the activation set and the deferred set, and retains the original order of the elements and the adjustment identifier.

[0111] S403: Based on the factor activation instruction set, input it into the particle swarm optimization algorithm, adjust the instruction weights through algorithm iteration, optimize the weight coefficients according to the task cycle, perform dynamic adaptation of instructions and cycle, and generate dynamic adversarial training sequences.

[0112] Obtain the factor activation instruction set, input the elements of the activation set and the deferred set sequentially into the weight adjustment processing unit, and initialize the weight value for each instruction. Then, the initial weight adjustment coefficient is calculated based on the task cycle parameter. The calculation rule is as follows: if the task cycle is higher than the current set average, the weight is reduced proportionally, and the reduction ratio is calculated as the percentage of the cycle difference to the average cycle. If the period is lower than the average, the weight is increased proportionally, and the increase ratio is calculated in the same way. For example, if the average period of the activation set is... The task cycle of P7 is also... Since the period difference is 0, the weight adjustment coefficient is 0, and the initial weights remain unchanged. Unchanged; the average period of the delayed set is P4 is also The period difference is 0, and the initial weights are also maintained. In the first iteration, the current weight of each instruction is multiplied by the disturbance duration coefficient, and the resulting value is used as the iteration evaluation value. For example, P7 is... P4 is The average of the iterative evaluation values ​​for the activation set and the delayed set is used as the group evaluation for this round. The average of the activation set is... The average value of the delayed set is In the second iteration, the weights of instructions whose iterative evaluation values ​​are higher than the set average are reduced. Instructions with weights below the set average have their weights increased. Since both the activation set and the deferred set have only one instruction in this round, their evaluation values ​​are exactly equal to the set average, so the weights remain unchanged. To demonstrate dynamic adjustment, let's assume that a new activation instruction P9 is added in the third round of simulation, and its evaluation value is... Below the average of the activation set Then P9's weight will be increased in the next round. Conversely, the weight of P7 decreases to This process is repeated multiple times until the change in weights is less than... The dynamic adversarial training sequence that converges over time and is finally output is denoted as:

[0113] {(P7, w=0.950, node 3), (P9, w=1.050, node X), (P4, w=1.000, node 2)};

[0114] This sequence binds weights, task nodes, and instruction types to form a dynamic adversarial training output that can be directly invoked.

[0115] Please see Figure 6 The specific steps of S5 are as follows:

[0116] S501: Call the dynamic adversarial training sequence, perform weighted processing based on the interference frequency coefficient and the remaining time value, allocate the weighted value to the signal injection period and set the interval, and generate the injection interval time value.

[0117] The dynamic adversarial training sequence is invoked. First, the corresponding interference frequency coefficient and remaining time value are read for each sequence element. The interference frequency coefficient is obtained by counting the number of times the interference signal recurs per unit time. For example, P7 experienced 35 interference occurrences in a past task cycle, and the total cycle duration was... Then its interference frequency coefficient is P9 experiences 31 interferences within a period, with a frequency coefficient of [missing value]. P4 is the 28th power, with a corresponding coefficient of The remaining time value represents the available injection window time for the corresponding parameter group before the task ends. It is obtained by subtracting the executed time from the task cycle. For example, if P7 has executed for 15.0 seconds in a 30.08-second cycle, then the remaining time is 15.08 seconds; if P9 has executed for 12.0 seconds, then the remaining time is 18.08 seconds; and if P4 has executed for 20.0 seconds, then the remaining time is 10.08 seconds. Then, a weighted processing is performed: the interference frequency coefficient of each parameter group is multiplied by its corresponding dynamic weight, and the result is multiplied by the remaining time value. The resulting value is the weighted value. For example, the weighted result for P7 is... P9 is P4 is Then, the weighted values ​​are allocated according to the actual available injection signal time periods. Each cycle, totaling 30.08 seconds, is divided into time periods equal to the number of sequence elements. Here, there are 3 elements, therefore the baseline length of a single time period is... The weighted result is mapped proportionally to the length of the time period to obtain the time each parameter group occupies in that time period. Then, the difference between this and the remaining time period length is used to calculate the injection interval time value. For example, the weighted value of P7, 14.76, corresponds to a proportional... The time taken is The corresponding interval value is The process of generating the injection interval time value is completed by performing the same calculation on all elements.

[0118] Table 5: Calculation Table of Weighted Processing and Injection Interval

[0119]

[0120] As shown in Table 5, the weighted value is obtained by calculating the frequency coefficients one by one, multiplying them by the dynamic weights and the remaining time, and then mapping them proportionally to generate the occupancy and interval time of each signal injection period, forming an injection interval time value dataset.

[0121] S502: Based on the injection interval time value, the rhythm of the interference signal transmission is adjusted. During the training operation, the response delay time of the training subject is monitored. The difference between the delay time and the set response benchmark time is calculated and converted into a deviation index to obtain the delay time deviation index.

[0122] The injection interval time value data is retrieved. First, the interval time of each parameter group is read one by one, and this is used as the reference injection period for the interference signal. Then, the start and end times of signal transmission are set sequentially. For example, the injection interval of P7 is... Its response window starts at 0 seconds of each cycle, with a delay after the first signal injection. Execute the next injection; the interval for P9 is... The interval of P4 is During the continuous transmission of interference signals, a periodic injection rhythm is formed according to the interval value. The time difference between the moment the interference signal is emitted and the moment the training subject completes the prescribed action feedback is monitored in real time during operation and used as the response delay time. This time is recorded to the millisecond by a high-precision timing module. For example, in one cycle, the recorded response delay after the P7 trigger signal is... P9 is P4 is Then, a preset baseline response time is called for comparison. This baseline time is measured by the average of the task responses under multiple interference-free conditions. For example, multiple rounds of experiments show that the baseline time for task Node 3 (P7) is... Node X (P9) is Node 2 (P4) is When performing the difference calculation, the actual delay time is subtracted from the corresponding reference time to obtain the delay time difference. For example, P7 is... P9 is P4 is The difference is converted into a dimensionless deviation index, calculated by dividing the difference by the reference time. For example, the deviation index for P7 is... P9 is P4 is The final set of deviation indices .

[0123] S503: Compare the delay time deviation index with the strategy adjustment threshold. When the deviation index exceeds the threshold, switch the backup parameter group and re-execute the signal injection. After numerically reorganizing the characteristics of the interference signal, calculate the training performance and obtain the evaluation results of the adversarial training enhancement effect.

[0124] The deviation index set is invoked. First, each record is read sequentially, and the deviation index is directly compared with a preset policy adjustment threshold. This threshold is set based on the task's tolerance error level and actual training requirements. In this embodiment, the threshold is set to... As an adjustment threshold (i.e., the allowable deviation percentage exceeding the baseline time is 12%), during comparison, if the deviation index is greater than the threshold, the system immediately switches to the backup parameter group for that parameter group and re-executes signal injection within the task cycle; if the deviation index is not greater than the threshold, the current parameter group and injection rhythm are maintained. For example, the deviation index for P7 is... Therefore, a switch is triggered. Assuming the spare parameter group for P7 is P7′, the injection plan is rebuilt within the remaining task time after the replacement; P9 is... Maintain the original parameter set and rhythm; the deviation index of P4 is This triggers a switch to the backup parameter group P4′ and rebuilds the injection plan. During the rebuilding process, the characteristic parameters of the interference signal from the switched parameter group are called for numerical reconstruction. This process involves reading three basic elements: signal amplitude (in V), signal duration (in s), and pulse interval (in ms), and performing linear scaling and time-domain offset processing. For example, if the amplitude of P7′ increases by 10% compared to P7, the reconstructed amplitude value will be the original value. Increase to Duration is determined by Change to The pulse interval is from Change to The amplitude of P4′ is determined by Increase to Duration is determined by Increase to The pulse interval is from shorten to The new feature values ​​are then recombined into an injection signal data packet and sent to the execution module for injection. After the training run is complete, the sensors are called to record the key performance indicators of the training subject under all interference rounds. For example, in this case, three core data points are taken: average reaction time (in seconds), error rate (%), and key task completion rate (%). These are weighted and summarized according to the training performance calculation rules. For example, let the reaction time weight be 1. The error rate weight is The task completion rate is weighted as follows: The final evaluation value of the enhanced adversarial training effect is obtained by summing the indicators according to their weights.

[0125] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. An adversarial design training method, characterized in that, Includes the following steps: S1: By monitoring the execution state of the training subject, the timing data of task nodes is obtained, the start time of nodes, the end time of nodes, and the dependency relationship identifiers between nodes are collected. The data is input into the Hidden Markov Model for state transition calculation to generate the behavioral state sequence of the training subject. S2: Call the training subject behavior state sequence, obtain the spatial coordinate offset, action trajectory deviation angle, and resource consumption rate of each node in the training subject behavior state sequence, and perform weight allocation and normalization. After obtaining each normalized value, multiply it by the corresponding weight and sum them to obtain the weighted ratio of the node. After calculating the weighted ratio of each node in the state sequence, sort the weighted ratios from high to low, and determine the nodes in the top column as weak links to form a training subject weak link identification table. S301: Obtain the weighted ratio value in the training subject's weak link identification table, match the weighted ratio value with the interference parameter group number in the adversarial factor library item by item, compare the weighted ratio value with the risk trigger threshold and the stable operation threshold, record the corresponding threshold status, and generate threshold comparison values. S302: Call the threshold comparison value, select the enhanced interference parameter group for the weighted ratio value that is within the risk trigger threshold range, select the maintenance interference parameter group for the weighted ratio value that is within the stable operation threshold range, and filter out the weighted ratio values ​​that do not fall into any threshold range to obtain the number of parameter groups to be filtered. S303: The parameter group selection quantity is called to perform merging in the interference parameter group set, the enhanced interference parameter group and the maintenance interference parameter group are classified and integrated, and the parameters are mapped according to the integrated structure to generate a targeted anti-interference scheme. S401: Invoke the targeted anti-interference scheme. Based on the interference duration coefficient and the task execution cycle parameter, if the interference duration coefficient is greater than the implantation timing threshold, trigger the anti-factor activation command; if it is less than the delayed implantation threshold, trigger the delayed activation command. Record the marker and generate an interference matching judgment value. The implantation timing threshold is set to 25% of the task execution cycle, and the delayed implantation threshold is set to 15% of the task execution cycle. S402: Based on the interference matching judgment value, call the trigger flag and the delay flag, map them to the activation command and the delay command, and generate the adversarial factor activation command set according to the task cycle analysis and adjustment command adaptation. S403: Obtain the adversarial factor activation instruction set, input it into the particle swarm optimization algorithm, iteratively adjust the instruction weights, optimize the instruction weight coefficients according to the task execution cycle, dynamically adapt the instructions to the task execution cycle, and generate a dynamic adversarial training sequence.

2. The adversarial design training method according to claim 1, characterized in that, The training subject's behavioral state sequence includes state transition path, state switching frequency, and state duration; the targeted adversarial interference scheme includes interference initiation timing, interference duration, and interference intensity configuration; and the dynamic adversarial training sequence includes interference injection timing, interference intensity curve, and interference adjustment interval.

3. The adversarial design training method according to claim 1, characterized in that, The specific steps of S1 are as follows: S101: Monitor the execution status of the main training task nodes, collect the start and end times of the nodes, record the dependency relationship identifiers between the nodes, perform formatting processing on the start and end times of each node, sort them in chronological order according to the node dependencies, and generate a node time series dataset. S102: Based on the node time series dataset, calculate the time difference between adjacent nodes and match the dependency identifier, use the state transition probability matrix of the hidden Markov model to calculate the state transition probability, and obtain the time series transition probability value. S103: Based on the temporal transition probability value, a hidden Markov model is used to infer the training subject's behavioral state sequence. By calculating the transition probability between nodes and combining the state space information, the node path is gradually deduced to generate the training subject's behavioral state sequence.

4. The adversarial design training method according to claim 1, characterized in that, The method also includes step S5: S5: Call the dynamic adversarial training sequence, calculate the injection interval time according to the interference frequency coefficient and the remaining time, adjust the transmission rhythm of the interference signal, monitor the response delay time deviation index of the training subject, and when the deviation index exceeds the strategy adjustment threshold, switch the backup parameter group of the adversarial scheme to perform injection, and generate the adversarial training enhancement effect evaluation result. The evaluation results of the adversarial training enhancement effect include response delay time, adaptive change rate, and risk assessment index.

Images