Banking area fire early warning method and system based on unmanned aerial vehicle

Abstract

This invention discloses a method and system for early warning of fires in bank areas based on unmanned aerial vehicles (UAVs), relating to the field of fire early warning technology. The method includes: acquiring multi-dimensional monitoring data based on pre-programmed monitoring by a UAV swarm; mapping the multi-dimensional monitoring data to a high-dimensional order parameter tensor space using an order parameter transformation component, extracting critical behavioral features, and determining the approach to phase transition critical points to generate risk control instructions; activating a disaster management module based on the risk control instructions to determine a first vulnerability heatmap, a first worst-case development path, and a first intervention measure, thereby conducting proactive monitoring and collaborative risk control management of the bank area based on the UAV swarm; wherein the disaster management module includes a first predictor and a second predictor based on first-order adversarial training and second-order game training. This invention solves the technical problem of difficulty in timely and accurate identification of early fire risks in bank areas in existing technologies, achieving the technical effect of high-sensitivity early warning and early intervention before a fire occurs.

Description

Technical Field

[0001] This invention relates to the field of fire early warning technology, specifically to a method and system for early warning of fires in bank areas based on unmanned aerial vehicles (UAVs). Background Technology

[0002] Within banking areas, due to the complex spatial structure, dense personnel and equipment, and high requirements for business continuity, fire hazards often exhibit characteristics such as unclear early signs and rapid evolution. Traditional fire monitoring methods mainly rely on fixed sensing equipment or single parameter thresholds for judgment, which have limited monitoring dimensions and are difficult to comprehensively reflect changes in environmental conditions and the process of risk accumulation. When a fire is still in its incubation stage, it is often impossible to form an effective identification and early warning, which can easily lead to delayed discovery of fire risks. Summary of the Invention

[0003] This application provides a method and system for early warning of fires in bank areas based on drones, which is used to address the technical problem of difficulty in timely and accurate identification of early fire risks in bank areas in the existing technology.

[0004] In view of the above problems, this application provides a method and system for early warning of fires in bank areas based on drones.

[0005] The first aspect of this application provides a method for early warning of fires in bank areas based on unmanned aerial vehicles (UAVs), the method comprising: Based on pre-programmed monitoring by the drone swarm, multi-dimensional monitoring data is acquired; through an ordinal parameter transformation component, the multi-dimensional monitoring data is mapped to a high-dimensional ordinal parameter tensor space, critical behavioral features are extracted, and the approach determination of the phase transition critical point is performed to generate risk control instructions; based on the risk control instructions, the disaster management module is activated to determine the first vulnerability heatmap, the first worst-case development path, and the first intervention measures, and to carry out proactive monitoring and collaborative risk control management of the bank area based on the drone swarm; wherein, the disaster management module includes a first predictor and a second predictor based on first-order adversarial training and second-order game training.

[0006] A second aspect of this application provides an early warning system for fires in bank areas based on unmanned aerial vehicles (UAVs), the system comprising: The monitoring module is used to acquire multi-dimensional monitoring data based on pre-programmed monitoring of the drone swarm; the proximity judgment module is used to map the multi-dimensional monitoring data to a high-dimensional order parameter tensor space through an order parameter transformation component, extract critical behavioral features, and perform proximity judgment of phase transition critical points to generate risk control instructions; the risk control management module is used to activate the disaster management module according to the risk control instructions, determine the first vulnerability heatmap, the first worst-case development path, and the first intervention measures, and conduct proactive monitoring and collaborative risk control management of the bank area based on the drone swarm; wherein, the disaster management module includes a first predictor and a second predictor based on first-order adversarial training and second-order game training.

[0007] One or more technical solutions provided in this application have at least the following technical effects or advantages: This application acquires multi-dimensional monitoring data based on pre-programmed monitoring by a drone swarm; through an ordinal parameter conversion component, the multi-dimensional monitoring data is mapped to a high-dimensional ordinal parameter tensor space, critical behavioral features are extracted, and a phase transition critical point is approximated to generate risk control instructions; based on the risk control instructions, a disaster management module is activated to determine a first vulnerability heatmap, a first worst-case development path, and a first intervention measure, enabling proactive monitoring and collaborative risk control management of the bank area based on the drone swarm; wherein, the disaster management module includes a first predictor and a second predictor based on first-order adversarial training and second-order game training. This invention solves the technical problem of difficulty in timely and accurate identification of early fire risks in bank areas in existing technologies. By performing ordinal parameter conversion on the multi-dimensional monitoring data acquired by the drone swarm and approximating the phase transition critical point, it achieves the technical effect of high-sensitivity early warning and early intervention before a fire occurs. Attached Figure Description

[0008] 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.

[0009] Figure 1 A schematic flowchart of an early warning method for fires in bank areas based on unmanned aerial vehicles (UAVs) provided in an embodiment of this application. Figure 2 A schematic diagram of the structure of an early warning system for fires in a bank area based on unmanned aerial vehicles (UAVs) provided in this application embodiment.

[0010] Explanation of reference numerals in the attached diagram: Monitoring module 11, Approach determination module 12, Risk control management module 13. Detailed Implementation

[0011] This application provides a method and system for early warning of fires in bank areas based on drones. It addresses the technical problem of difficulty in timely and accurate identification of early fire risks in bank areas in existing technologies. By performing order parameter conversion on multi-dimensional monitoring data acquired by drone swarms and determining the approach of phase transition critical points, it achieves the technical effect of high-sensitivity early warning and early intervention before a fire occurs.

[0012] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of this application, and not all of them. All other embodiments obtained by those skilled in the art based on the embodiments of this application without creative effort are within the scope of protection of this application.

[0013] It should be noted that any variation of the terms "comprising" and "having" is intended to cover non-exclusive inclusion, for example, a process, method, system, product, or server that includes a series of steps or units is not necessarily limited to those steps or units that are explicitly listed, but may include other steps or modules that are not explicitly listed or that are inherent to such processes, methods, products, or devices.

[0014] Example 1, as Figure 1 As shown, this application provides an early warning method for fires in bank areas based on unmanned aerial vehicles (UAVs), the method comprising: Step S100: Obtain multi-dimensional monitoring data based on the pre-programmed monitoring of the drone swarm.

[0015] In this embodiment, a pre-programmed monitoring method based on a drone swarm is used. The flight paths, inspection sequences, and monitoring tasks of each drone in the swarm are pre-set uniformly, enabling the drone swarm to continuously monitor the bank area according to the preset monitoring plan. During the pre-programmed monitoring, each drone simultaneously collects environmental information related to fire risk during its patrol. The acquired multi-dimensional monitoring data is a multi-dimensional data set used to characterize the fire incubation status of the bank area, including information on changes in ambient temperature, smoke concentration, combustible gas concentration, and air flow. Subsequently, the multi-dimensional monitoring data collected by each drone is collected and integrated according to the corresponding time markers and spatial locations to form multi-dimensional monitoring data reflecting the overall environmental change status of the bank area.

[0016] Step S200: Through the order parameter conversion component, the multivariate monitoring data is mapped to a high-dimensional order parameter tensor space, critical behavior features are extracted, and the approach determination of the phase transition critical point is performed to generate risk control instructions.

[0017] In this embodiment, a sequence parameter conversion component is used to process multivariate monitoring data. This component converts the multivariate monitoring data characterizing the environmental state of the bank area into sequence parameters describing the overall evolutionary state. Specifically, the sequence parameters include thermodynamic entropy gradient, information structure complexity, environmental resonance mode, and material phase transition barrier. The thermodynamic entropy gradient characterizes the changing trend of environmental energy disorder; information structure complexity reflects the degree of evolution of the internal correlation structure of the monitoring data; environmental resonance mode describes the cooperative change characteristics between different physical quantities in the environment; and material phase transition barrier characterizes the energy threshold that must be overcome for the environmental state to evolve from stable to abrupt. By performing sequence parameter conversion on the multivariate monitoring data, a high-dimensional sequence parameter tensor space is constructed and dynamically updated. This high-dimensional sequence parameter tensor space is used to comprehensively express the state of multiple sequence parameters changing with time and space within a unified mathematical space.

[0018] Next, the high-dimensional order parameter tensor space is analyzed to extract critical behavioral features. Critical behavioral features include at least one of the following: correlation length divergence, exponential growth of relaxation time, and intensified order parameter fluctuations. When a critical behavioral feature is detected to meet a preset condition, it is determined that the bank's regional environmental state is approaching the phase transition critical point. After completing the approach determination of the phase transition critical point, a risk control instruction corresponding to the current risk state is generated. This risk control instruction is used to trigger the prediction and control process of the subsequent disaster management module.

[0019] Furthermore, the method provided in the application embodiments also includes: The order parameters include thermodynamic entropy gradient, information structure complexity, environmental resonance mode, and material phase transition barrier. By performing order parameter transformation on the multivariate monitoring data, the high-dimensional order parameter tensor space is dynamically updated, and critical behavioral features are extracted. If the phase transition point is approached, the risk control instruction is generated.

[0020] In the embodiments of this application, the order parameters include thermodynamic entropy gradient, information structure complexity, environmental resonance mode, and material phase transition barrier. The thermodynamic entropy gradient is used to reflect the degree of acceleration of the change from order to disorder in time and space. The information structure complexity is used to reflect the complexity after the internal regularity of multivariate monitoring data is destroyed. The environmental resonance mode is used to reflect the periodic characteristics of the synchronous fluctuation of multiple monitoring quantities on the same time scale. The material phase transition barrier is used to reflect the remaining distance that the current state still needs to overcome before it can reach a sudden change state.

[0021] When performing ordinal parameter transformation on multivariate monitoring data, the continuous multivariate monitoring data is first divided into continuous time windows with a fixed time length. Within each time window, the temperature, smoke concentration, combustible gas concentration, and air flow state information are processed into sequences under the same time reference, so that each monitoring quantity forms a one-to-one corresponding sampling sequence within the time window. Subsequently, within the time window, the thermodynamic entropy gradient, information structure complexity, environmental resonance mode, and material phase transition barrier are calculated respectively. The four types of ordinal parameters obtained are continuously written into the high-dimensional ordinal parameter tensor space according to the time window order, so that the high-dimensional ordinal parameter tensor space is dynamically updated as the time window rolls.

[0022] In calculating the thermodynamic entropy gradient, within a certain time window, the temperature sequence, smoke concentration sequence, and combustible gas concentration sequence are divided into several numerical intervals according to their value ranges. The number of sampling points falling into each interval is counted, and this number is divided by the total number of sampling points in the time window to obtain the probability of the interval appearing. The natural logarithm of the probability of each interval is taken, and then the probability is multiplied by its logarithm. The product of all intervals is added together and the negative value is taken to obtain the entropy value of the monitored quantity in the time window. After obtaining the entropy values ​​of temperature, smoke concentration, and combustible gas concentration separately, the three are added together to obtain the comprehensive entropy value of the time window. The difference between the comprehensive entropy values ​​of adjacent time windows is then used to obtain the entropy change, and the entropy change rate is obtained by dividing the entropy change by the time interval between adjacent time windows. If the multivariate monitoring data corresponds to different spatial locations, the entropy change rate of different spatial locations is differentiated and divided by the spatial distance to obtain the thermodynamic entropy gradient, which is used to characterize whether the change in disorder exhibits faster diffusion and stronger change in space.

[0023] When calculating information structure complexity, within a certain time window, temperature, smoke concentration, combustible gas concentration, and airflow status information are arranged into a multidimensional sequence in the same chronological order, and fixed-length data segments are extracted sequentially from this sequence. For any two data segments, the corresponding sample values ​​are compared point by point, the difference at each sample point is calculated, and the maximum difference is taken as the distance between the two segments. When this distance is less than a preset threshold, the two segments are determined to be similar, and the proportion of all similar segment pairs among all comparable segment pairs is calculated to obtain the similarity ratio. Subsequently, the segment length is increased by one sample point, and the same similarity determination and statistical steps are repeated to obtain another similarity ratio. The logarithmic ratio of the two similarity ratios is calculated to obtain the information structure complexity, which reflects the rate of decrease in the repeatability of the sequence pattern. The faster the repeatability decreases, the higher the complexity.

[0024] When calculating the environmental resonance mode, within a certain time window, the temperature sequence, smoke concentration sequence, combustible gas concentration sequence, and air flow state sequence are periodically decomposed. Specifically, a set of candidate periods is selected, and corresponding sine and cosine wave sequences are constructed for each candidate period. The monitored quantity sequence and the sine wave sequence are multiplied point by point and summed to obtain the cosine projection coefficient. Then, the monitored quantity sequence and the cosine wave sequence are multiplied point by point and summed to obtain the sine projection coefficient. The two projection coefficients are divided by the number of sampling points within the time window to obtain the normalized coefficients. Finally, the normalized cosine projection coefficient and the sine projection coefficient are squared. The square root of the sum is used to obtain the amplitude corresponding to the candidate period, which is the intensity of the periodic component. The above calculation is repeated for all candidate periods, and the candidate period with the largest amplitude is selected as the main period of the monitoring quantity, and the amplitude of the main period is recorded. After obtaining the main periods for the four types of monitoring quantities, the main periods are compared to see if they are the same or close. When the main periods of at least two types of monitoring quantities are the same or close, and their main period amplitudes show an increasing trend within a continuous time window, the common main period and its corresponding amplitude enhancement state are determined as the environmental resonance mode, which is used to characterize the periodic changes of multiple monitoring quantities that show synchronous enhancement on the same time scale.

[0025] When calculating the phase transition barrier, the average values ​​of the temperature sequence, smoke concentration sequence, and combustible gas concentration sequence are calculated separately within a certain time window. The average value is obtained by summing the sampled values ​​within the time window and dividing by the total number of sampling points. Then, the three average values ​​are weighted and summed according to preset weights to obtain the comprehensive state quantity. A critical state quantity is preset as the phase transition threshold. The comprehensive state quantity is compared with the critical state quantity. When the comprehensive state quantity is less than the critical state quantity, the difference between the critical state quantity and the comprehensive state quantity is obtained. This difference is the phase transition barrier. The smaller the difference, the closer it is to the abrupt change boundary.

[0026] After obtaining the thermodynamic entropy gradient, information structure complexity, environmental resonance mode, and material phase transition barrier, the four types of order parameters corresponding to each time window and each spatial location are written into the high-dimensional order parameter tensor space in a fixed order. This calculation is repeated and the data at the corresponding positions is updated as the time window rolls, thus achieving dynamic updating of the high-dimensional order parameter tensor space. Based on the dynamic updating of the high-dimensional order parameter tensor space, critical behavior features are extracted from the changing trends of the four types of order parameters within continuous time windows. This is achieved by analyzing the thermodynamic entropy gradient, information structure complexity, environmental resonance mode, and material phase transition barrier. The distribution characteristics of the potential barrier in the spatial and temporal dimensions are used to determine whether at least one of the following preset conditions occurs: correlation length divergence, exponential growth of relaxation time, or intensified fluctuation of order parameters. Correlation length divergence is manifested as a continuous expansion of the correlation of order parameters in the spatial range; exponential growth of relaxation time is manifested as a significant increase in the time required for order parameters to recover to a stable state after being disturbed; and intensified fluctuation of order parameters is manifested as a significant increase in the fluctuation amplitude of order parameters within a continuous time window. When any of the above preset conditions is detected to be met, it is determined that the current environmental state is approaching the phase transition critical point, and risk control instructions are generated accordingly.

[0027] Furthermore, the method provided in the application embodiments also includes: Identify critical characteristics of behavior and determine whether they meet preset conditions. If they do, it is determined to be approaching the critical point of phase transition. Among them, at least one of the following preset conditions is met: correlation length divergence, exponential growth of relaxation time, and intensified fluctuation of order parameter.

[0028] In this embodiment, after the high-dimensional order parameter tensor space is continuously updated, the critical features of the identified behavior are transformed into three indicators that can be directly calculated: the spatial correlation range, the recovery time scale, and the fluctuation intensity. These indicators are repeatedly calculated within a continuous time window to form an indicator sequence. When the indicator sequence satisfies at least one of the preset conditions of correlation length divergence, exponential growth of relaxation time, and intensified fluctuation of order parameter, it is judged to be approaching the critical point of phase transition.

[0029] The spatial association range is calculated using the association length. Within the k-th time window, a reference spatial location is first determined, and the ordinal parameter value of this location within the time window is recorded as the reference value. Then, all spatial locations at a distance d from the reference location are selected sequentially. The ordinal parameter values ​​of these locations are subtracted from the reference value, and the absolute values ​​are taken. All absolute values ​​obtained at the same distance d are then summed and divided by the number of spatial locations at that distance d to obtain the average difference. The distance d is gradually increased from small to large, and the above average difference calculation is repeated to obtain a set of results showing the average difference changing with distance. A fixed threshold is set in this result as the criterion for maintaining a strong association. The maximum distance where the average difference is less than or equal to this threshold is found, and this maximum distance is defined as the association length within the time window. The larger the association length, the larger the spatial association range. The association length is repeatedly calculated within multiple consecutive time windows to form an association length sequence. When the association length sequence gradually increases within multiple consecutive time windows, it is determined that association length divergence has occurred.

[0030] The recovery timescale is calculated using the relaxation time. Within the k-th time window, the time series of the order parameter at the same spatial location are selected. The average value obtained by summing all sampled values ​​within this time window and dividing by the number of sample points is used as the stable benchmark. Then, the absolute value of the difference between each sampled value and the stable benchmark is calculated point by point within this time window. The sampled point with the largest absolute value of the difference is identified, and its corresponding time is taken as the disturbance start point. Starting from the disturbance start point, the absolute value of the difference is checked point by point along time. A fixed tolerance threshold is set. When the absolute value of the difference is less than or equal to the tolerance threshold for the first time, this time is taken as the recovery time. The relaxation time is obtained by subtracting the disturbance start point from the recovery time. The larger the relaxation time, the longer the recovery timescale. The relaxation time is obtained by repeating this process within multiple consecutive time windows to form a relaxation time series. When the relaxation time series increases exponentially within multiple consecutive time windows, it is determined that the relaxation time is exponentially growing.

[0031] The fluctuation intensity is calculated by the fluctuation amplitude. Within the k-th time window, the time series of the order parameter at the same spatial location is selected, and the maximum and minimum sample values ​​within the time window are found. The fluctuation amplitude is obtained by subtracting the minimum sample value from the maximum sample value, and this fluctuation amplitude is taken as the fluctuation intensity within the time window. The fluctuation intensity is repeatedly calculated within multiple consecutive time windows to form a fluctuation intensity sequence. When the fluctuation intensity sequence continues to increase within multiple consecutive time windows and exceeds a preset threshold, it is determined that the fluctuation of the order parameter has intensified.

[0032] After completing the continuous calculation of the three indicators of spatial correlation range, recovery time scale and fluctuation intensity, check whether at least one of the preset conditions of correlation length divergence, relaxation time exponential growth and order parameter fluctuation intensification is met within the same judgment period. If at least one condition is met, it is judged to be approaching the phase transition critical point.

[0033] Furthermore, in the method provided in the application embodiments, before activating the disaster management module, the construction of the disaster management module further includes: A first predictor is deployed with a vulnerability heatmap based on fire development and environmental conditions as the first generation target and the worst development path as the second generation target; a second predictor is deployed with intervention measures based on fire development and environmental conditions; and a disaster management module is generated by performing two-stage training on the first and second predictors.

[0034] In this embodiment of the application, when constructing the disaster management module, the fire development and environmental conditions are first structurally modeled. The fire development is represented by the change process of fire source intensity, burning area and spread speed in the form of time series. The environmental conditions include the ventilation status, spatial structure layout, floor partition structure and combustible material distribution density of the bank building. The fire development data and environmental condition data are aligned with a unified time scale and spatial grid, and the bank building is divided into regular spatial grid units. The corresponding ventilation parameters, structural connectivity and material property parameters are recorded in each spatial grid unit to form a multi-dimensional feature vector for training.

[0035] When deploying a first predictor with a vulnerability heatmap based on fire development and environmental conditions as the first generation target and the worst-case development path as the second generation target, a multi-layer feedforward neural network structure is used as the first predictor model. The multi-layer feedforward neural network structure includes an input layer, several hidden layers, and an output layer. The input layer receives a feature vector composed of fire development data and environmental condition data. The hidden layers extract spatial and temporal features through linear weighting and nonlinear activation operations. The output layer outputs the risk probability value and fire spread direction probability of each spatial grid cell, respectively. During the training phase, historical fire case data is used, with the real vulnerability heatmap as the supervision label and the historical fire spread trajectory as the worst-case development path label. The difference between the predicted result and the real label is calculated through the error backpropagation algorithm, and the network weights are updated through the gradient descent method, so that the first predictor gradually approaches the real fire development law, thereby realizing the joint generation of vulnerability heatmap and worst-case development path.

[0036] When deploying the second predictor, the goal is to generate intervention measures based on fire development and environmental conditions. A multi-layer feedforward neural network structure is used as the second predictor model. Its input includes the vulnerability heatmap, worst-case development path, and current environmental condition data output by the first predictor. The output is the intervention measure category code and execution time series for the corresponding spatial grid cell. During training, the actual intervention measures in historical fire response records are used as supervision labels. The difference between the predicted intervention measures and the actual intervention measures is calculated using the cross-entropy loss function. The network parameters are updated using the error backpropagation algorithm, so that the second predictor can generate intervention measures that match the actual response logic based on the risk distribution and environmental structure characteristics.

[0037] After deploying the first and second predictors, two-stage training is performed on them. First, a judge is introduced to construct an adversarial training framework. The first and second predictors are used as generators to execute the training process in parallel. The judge is used to determine the authenticity and rationality of the vulnerability heatmap and worst-case development path output by the first predictor, as well as the intervention measures output by the second predictor. The parameters of the first and second predictors are updated in reverse based on the judgment results, thus completing the first-stage adversarial training. Then, after completing the first-stage adversarial training, the trained first and second predictors are placed in an incomplete information dynamic game model, acting as different decision-making agents. Under the premise of uncertainty in fire development and environmental conditions, strategies are iteratively updated. Strategy convergence is achieved by optimizing the risk-reward function, thus completing the second-stage incomplete information dynamic game training and generating a disaster management module.

[0038] Furthermore, the method provided in the application embodiments, which involves two-stage training of the first predictor and the second predictor, further includes: By introducing a judge, the adversarial training of the first predictor and the second predictor in parallel is used as the first-order training method; the incomplete information dynamic game training of the first predictor and the second predictor after the first-order training is used as the second-order training method.

[0039] In this embodiment, during the first-order training phase, an adversarial training process is constructed by introducing a judge. First, a training dataset is prepared. Each sample includes fire development data and environmental condition data as input, along with corresponding real vulnerability heatmaps, real worst-case development paths, and real intervention measures as labels. A first predictor receives the fire development data and environmental condition data, and outputs a predicted vulnerability heatmap and a predicted worst-case development path. A second predictor receives the same input and the output of the first predictor, and generates predicted intervention measures. The judge's input consists of two types of data: real label data and data generated by the first and second predictors. The judge performs forward computation on the input data and outputs a value between 0 and 1, representing the probability that the input is a real sample. A target value of 1 is set for the real label data, and a target value of 0 is set for the generated data. When calculating the judge's error, the difference is obtained by subtracting the judge's output value from the target value. The difference is then squared and averaged over the same batch of samples to obtain the judge's loss value. The backpropagation algorithm is used to calculate the impact of the weights of each layer of the judge on the loss value, and the weights are updated according to a set learning rate, enabling the judge to improve its ability to distinguish between real and generated data. After updating the judge, the judge parameters are fixed, and the generated data is input into the judge again. This time, the target value of the generated data is set to 1, and the new difference is calculated and the squared average is taken to obtain the generated loss value. The parameters of the first and second predictors are updated separately using the backpropagation algorithm, so that the generated results are closer to the real data distribution under the judgment of the judge. The above steps are repeated in multiple training rounds until the change range of the judge loss value and the generated loss value is lower than the preset value in consecutive training rounds, thus completing the first-order adversarial training.

[0040] In the second-order training phase, the first and second predictors, trained using incomplete information dynamic game theory, are trained. First, the parameters of the second predictor are fixed, and fire development data and environmental condition data are input into the first predictor to obtain a predicted vulnerability heatmap and a predicted worst-case development path. The predicted vulnerability heatmap is a numerical matrix arranged by spatial grid indices, where the value corresponding to each spatial grid is its risk value. To calculate the risk error, the risk value of each spatial grid in the predicted vulnerability heatmap is subtracted from its corresponding actual risk value. The difference is squared and summed over all grids to obtain the heatmap error. The spatial coordinates of each path point in the predicted worst-case development path are squared, and the sum of the squared differences is taken to obtain the distance error for each path point. The path error is then summed over all path point errors. The heatmap error and the path error are added to obtain the total risk error. The parameters of the first predictor are updated using a backpropagation algorithm to reduce the total risk error. Then, the parameters of the first predictor are fixed, and its output is input into the second predictor to obtain predicted intervention measures. The risk value after intervention is recalculated based on the predicted intervention measures. The risk reduction is obtained by subtracting the risk value after intervention from the risk value before intervention. The total reduction value is obtained by summing the risk reduction values ​​for all spatial grids and is used as the optimization objective. The parameters of the second predictor are updated through the backpropagation algorithm to increase the total reduction value. The above steps are performed alternately, and each update of the first predictor and the second predictor is considered as one round of game iteration. When the change in the total risk error and the change in the total reduction value are both less than a preset threshold in several consecutive rounds, the training is considered to have reached a stable state, thus completing the second-order incomplete information dynamic game training and obtaining the final disaster management module.

[0041] Step S300: Activate the disaster management module according to the risk control instruction, determine the first vulnerability heat map, the first worst-case development path and the first intervention measure, and conduct proactive monitoring and collaborative risk control management of the bank area based on drone swarm; wherein, the disaster management module includes a first predictor and a second predictor based on first-order adversarial training and second-order game training.

[0042] In this embodiment, the disaster management module is first activated according to risk control instructions, and the dynamically updated high-dimensional ordered parametric tensor space is input into the disaster management module as a unified data input source. The disaster management module contains a first predictor and a second predictor optimized through first-order adversarial training and second-order game training. The first predictor generates risk evolution results, and the second predictor generates intervention strategy results. In the activated state, the forward computation processes of the first and second predictors are executed in parallel. The first predictor outputs a first vulnerability heatmap and a first worst-case development path within a preset time window, and the second predictor outputs a first intervention measure within the same preset time window. Subsequently, the first vulnerability heatmap and the first worst-case development path are used for targeted monitoring task allocation and flight trajectory adjustment of the drone swarm, and the first intervention measure is used for risk control early warning instruction generation and execution control, thereby implementing proactive monitoring and collaborative risk control management of the bank area based on the drone swarm.

[0043] Furthermore, the method provided in the application embodiments, which determines a first vulnerability heatmap, a first worst-case development path, and a first intervention measure, and conducts proactive monitoring and collaborative risk control management of the bank area based on drone swarms, also includes: According to the risk control command, the disaster management module is activated, and the dynamically updated high-dimensional order parameter tensor space is input into the disaster management module; the first predictor and the second predictor are driven in parallel to output the first prediction data and the second prediction data within a preset time window; based on the first prediction data and the second prediction data, the directional monitoring management and risk control early warning management of the drone swarm are executed.

[0044] In this embodiment, a signal is written into the control register unit of the disaster management module, causing the disaster management module to switch from standby to running state, and simultaneously locking the preset time window parameters corresponding to this prediction. After the state switch is completed, the data loading process is started, and the corresponding data is read from the dynamically updated high-dimensional order parameter tensor space according to the current time index and spatial grid index. The high-dimensional order parameter tensor space is a data structure organized according to the time dimension and spatial grid dimension, which internally stores the thermodynamic entropy gradient, information structure complexity, environmental resonance mode and material phase transition barrier values ​​corresponding to each time window and each spatial location. During the reading process, the four types of order parameters of each spatial grid in each time window are spliced ​​in a fixed order to form an order parameter vector, and arranged into an order parameter matrix according to the spatial grid index order. Then, the order parameter matrices of multiple time windows are superimposed in time order to form an input data block, and the input data block is loaded into the disaster management module.

[0045] After the disaster management module is running, the first and second predictors are driven in parallel to perform forward calculations. The first predictor performs layer-by-layer weighted calculations and nonlinear mapping calculations based on the input data blocks, and outputs a first vulnerability heatmap within a preset time window. Based on the connectivity between spatial grids and the order of the values ​​of each spatial grid in the first vulnerability heatmap, it selects adjacent grids in descending order of value to form a risk expansion trajectory, thereby obtaining the first worst-case development path. The two together constitute the first prediction data. The second predictor performs layer-by-layer weighted calculations under the same input conditions, outputs the intervention priority value and intervention action code corresponding to each spatial grid, and converts the intervention action code into specific intervention measures according to the preset action mapping rules to form the first intervention measures, which serve as the second prediction data.

[0046] Finally, based on the first and second prediction data, targeted monitoring and risk control early warning management of the drone swarm is implemented. Specifically, when implementing targeted monitoring management of the drone swarm, the first monitoring target is determined based on the first vulnerability heatmap in the first prediction data, and the second monitoring target is determined based on the first worst-case development path. Subsequently, the first and second monitoring targets are prioritized, and monitoring trajectories are planned by combining the spatial grid structure of the bank area with the current location and status of the drones to generate a pre-monitoring strategy. This pre-monitoring strategy is then distributed to each drone control unit in the drone swarm, and each drone control unit adjusts its flight path and patrols key areas according to the pre-monitoring strategy, thus achieving targeted monitoring management.

[0047] When implementing risk control and early warning management, for the first intervention measure in the second prediction data, the first intervention measure is decoupled using the smallest risk control unit as the basic execution unit and broken down into multiple intervention points corresponding to different spatial locations and time nodes. Based on the spatiotemporal relationship of risk control, the multiple intervention points are identified by spatiotemporal codes to form multi-threaded early warning instructions corresponding to spatial grid indexes and time window parameters. Subsequently, the multi-threaded early warning instructions are uniformly scheduled and distributed through the site central control system, and multi-threaded issuance management is carried out according to the execution subject corresponding to the intervention point, thereby realizing the parallelization and refined execution of risk control and early warning management.

[0048] Furthermore, the method provided in the application embodiments, for performing directional monitoring and management of the drone swarm, further includes: Based on the first vulnerability heatmap in the first prediction data, a first monitoring target is determined, and a second monitoring target is determined based on the first worst-case development path. The first and second monitoring targets are prioritized and their monitoring trajectories are planned to determine a pre-monitoring strategy. The pre-monitoring strategy is then distributed to the central control units of each UAV in the UAV cluster to perform targeted monitoring management.

[0049] In this embodiment of the application, when determining the first monitoring target based on the first vulnerability heat map in the first prediction data, the first vulnerability heat map is first read one by one according to the spatial grid index, the corresponding value of each spatial grid is obtained, and compared with the preset monitoring threshold, and spatial grids with values ​​greater than or equal to the preset monitoring threshold are selected; then the selected spatial grids are sorted from largest to smallest according to their corresponding values, and the spatial grids with the highest sorted values ​​are taken as the first monitoring target, and the first monitoring target is recorded in the form of spatial grid coordinates to indicate the key areas that need to be covered first.

[0050] When determining the second monitoring target based on the first worst-case development path, the path node sequence in the first worst-case development path is read. The path node sequence is a spatial grid coordinate arranged in chronological order. Then, a preset number of key path nodes are selected from the path node sequence as the second monitoring target. The selection rule is to prioritize the path nodes at the beginning of the sequence and record them as the second monitoring target in the form of spatial grid coordinates to indicate the key locations that need to be tracked along the risk expansion direction.

[0051] When prioritizing the first and second monitoring targets, a hierarchical sorting method is adopted. The monitoring targets are divided into overlapping targets and non-overlapping targets. Overlapping targets are spatial grids that belong to both the first and second monitoring targets, while non-overlapping targets are spatial grids that belong only to the first or only to the second monitoring target. During the sorting, overlapping targets are placed at the highest priority level and arranged in descending order of their corresponding values ​​in the first vulnerability heatmap within this level. After overlapping targets, spatial grids that belong only to the second monitoring target are arranged in ascending order of their appearance in the first worst-case development path. Finally, spatial grids that belong only to the first monitoring target are arranged in descending order of their corresponding values ​​in the first vulnerability heatmap, thus obtaining the monitoring priority list.

[0052] When planning the monitoring trajectory to determine the pre-monitoring strategy, the current position coordinates of each UAV in the UAV cluster are first read. The spatial grid coordinates in the monitoring priority list are used as target points in sequence. The straight-line distance from the current position to the first target point is calculated. The straight-line distance is obtained by taking the square root of the sum of the squares of the differences between the coordinates of the two points. After selecting the UAV with the shortest distance to perform the first target point task, the current position of the UAV is updated to the coordinates of the first target point. The distance calculation and UAV selection steps are repeated for the next target point until all target points are assigned. Then, for each UAV, the assigned target points are connected according to the monitoring priority to form a flight waypoint sequence, and the flight waypoint sequence is used as the content of the pre-monitoring strategy for that UAV.

[0053] Finally, when the pre-monitoring strategy is distributed to the central controllers of each UAV in the UAV swarm, the flight waypoint sequence corresponding to each UAV is encoded into a command data packet. The command data packet contains the waypoint coordinates and the waypoint execution order, and is sent to the corresponding UAV central controller through the communication link. After receiving the command data packet, each UAV central controller parses the waypoint coordinates and sequence information, converts the waypoint coordinates into the target heading, target speed and target altitude control quantities required for flight control, and updates the flight control command queue, so that the UAVs fly to the target point in sequence according to the pre-monitoring strategy and perform the patrol task, thereby completing the directional monitoring and management.

[0054] Furthermore, the method provided in the application embodiments, in performing risk control and early warning management, also includes: For the first intervention measure in the second predicted data, the first intervention measure is decoupled using the smallest risk control unit to determine multiple intervention points; the multiple intervention points are identified by spatiotemporal codes based on the spatiotemporal relationship of risk control to generate multi-threaded early warning instructions; and the multi-threaded early warning instructions are managed by site-centralized early warning and multi-threaded distribution.

[0055] In this embodiment, when performing risk control and early warning management on the first intervention measure in the second prediction data, the structure of the first intervention measure is first analyzed. The intervention actions arranged according to the spatial grid index in the first intervention measure are read one by one and classified and marked according to the execution type. In the classification process, smoke exhaust operations are marked as smoke exhaust control type, power outage operations are marked as power outage control type, and evacuation guidance operations are marked as evacuation guidance control type, and the corresponding spatial grid coordinates are used for positioning. Then, decoupling is performed using the smallest risk control unit as the basic execution unit. The smallest risk control unit is a physical control node that can independently execute a single control action, such as a single smoke exhaust valve control unit, a single floor power outage circuit control unit, or a single evacuation indicator control unit. In the decoupling process, operations of different execution types within the same spatial grid are split separately, and a corresponding intervention point record is established for each smallest risk control unit. The intervention point is composed of spatial grid coordinates and the smallest risk control unit identifier, thereby obtaining multiple intervention points.

[0056] After identifying multiple intervention points, these points are identified using spatiotemporal codes based on the spatiotemporal relationship of risk control. This spatiotemporal relationship is a rule that encodes and associates the spatial location of intervention points with their execution time sequence. Specifically, for each intervention point, its spatial grid index is extracted as a spatial identifier, and a time identifier is generated according to a preset execution time window and intervention priority order. The spatial identifier and time identifier are then concatenated into a spatiotemporal code identifier according to a fixed encoding order. Subsequently, each intervention point is combined with its corresponding spatiotemporal code identifier and execution type to generate an independent early warning instruction entry. Each early warning instruction entry corresponds to a minimum risk control unit and a specific execution time, thus forming a multi-threaded early warning instruction set, enabling different intervention points to be executed in parallel at different time nodes.

[0057] After generating multi-threaded early warning commands, the site central control system manages the early warning and multi-threaded distribution. In this process, the site central control system first receives all early warning command items and sorts them according to the time identifier in the spatiotemporal code, forming an execution order queue. Then, based on the spatial identifier, the early warning command items are distributed to the corresponding smallest risk control unit control terminal, and control commands are issued via wired or wireless communication links. Upon receiving the corresponding early warning command, each smallest risk control unit triggers an action according to the time parameter in the spatiotemporal code, achieving parallel execution of smoke exhaust control, power outage control, and evacuation guidance control, thus completing multi-threaded risk control and early warning management.

[0058] Furthermore, the method provided in the application embodiments, after conducting proactive monitoring and collaborative risk control management of the bank area based on drone swarms, also includes: Obtain risk control record data, mine strategy deviation features based on the risk control record data, wherein the strategy deviation features satisfy a preset generalization degree; and learn and update the first predictor and the second predictor based on the strategy deviation features.

[0059] In this embodiment, when acquiring risk control record data, historical records of completed handling processes are extracted from the site control system and the drone cluster operation logs. The risk control record data includes the historical first vulnerability heatmap, the historical first worst-case development path, the historical first intervention measure, and the actual monitoring data and execution feedback data within the corresponding time window. The actual monitoring data includes information on changes in ambient temperature, smoke concentration, combustible gas concentration, and air flow status. The execution feedback data includes the execution time, execution status, and execution result of each intervention measure. Subsequently, the above data are aligned according to the same time window and spatial grid index, so that the historical first vulnerability heatmap, the historical first worst-case development path, the historical first intervention measure, and the actual monitoring data establish a correspondence within the same spatial location and the same time period.

[0060] When mining strategy deviation characteristics based on risk control record data, a control result is first generated based on actual monitoring data. Within the same time window, statistical values ​​are calculated for environmental temperature change information, smoke concentration change information, combustible gas concentration change information, and air flow status information according to spatial grid index. The statistical values ​​are then combined according to preset rules to obtain the control vulnerability distribution results arranged by spatial grid index. Subsequently, the historical first vulnerability heatmap and the control vulnerability distribution results are compared one by one at the same spatial grid to calculate the difference, and the spatial grid with the absolute value of the difference greater than the preset deviation threshold is recorded as the deviation position. The above control generation and difference calculation are repeated in multiple historical records. The number of times and the proportion of occurrence of the deviation position and its corresponding difference pattern are counted. When a certain difference pattern occurs repeatedly in different historical records and the proportion of occurrence reaches the preset proportion threshold, the difference pattern is determined as a strategy deviation feature. The preset proportion threshold is used to characterize the preset generalization degree.

[0061] When learning and updating the first and second predictors based on strategy deviation characteristics, historical records containing strategy deviation characteristics are selected from the risk control record data as update samples. The corresponding fire development data and environmental condition data are used as inputs, and the control vulnerability distribution results, actual monitoring data, and execution feedback data are used as references. During the update process, the difference between the output of the first predictor and the control vulnerability distribution results, as well as the difference between the historical first intervention measures output by the second predictor and the execution feedback data, are calculated respectively. The parameters of the first and second predictors are updated with the goal of reducing the difference. After completing multiple rounds of updates, the same control generation and difference statistics method is used to check the newly generated risk control record data. When the occurrence ratio of strategy deviation characteristics is lower than the preset ratio threshold, the learning update is completed.

[0062] In summary, the embodiments of this application have at least the following technical effects: This application acquires multi-dimensional monitoring data based on pre-programmed monitoring by a drone swarm; through an ordinal parameter conversion component, the multi-dimensional monitoring data is mapped to a high-dimensional ordinal parameter tensor space, critical behavioral features are extracted, and a phase transition critical point is approximated to generate risk control instructions; based on the risk control instructions, a disaster management module is activated to determine a first vulnerability heatmap, a first worst-case development path, and a first intervention measure, enabling proactive monitoring and collaborative risk control management of the bank area based on the drone swarm; wherein, the disaster management module includes a first predictor and a second predictor based on first-order adversarial training and second-order game training. This invention solves the technical problem of difficulty in timely and accurate identification of early fire risks in bank areas in existing technologies. By performing ordinal parameter conversion on the multi-dimensional monitoring data acquired by the drone swarm and approximating the phase transition critical point, it achieves the technical effect of high-sensitivity early warning and early intervention before a fire occurs.

[0063] Example 2 is based on the same inventive concept as the drone-based early warning method for bank area fires in the previous examples, such as... Figure 2 As shown, this application provides an early warning system for bank area fires based on unmanned aerial vehicles (UAVs). The system and method embodiments in this application are based on the same inventive concept. The system includes: The monitoring module 11 is used to acquire multi-dimensional monitoring data based on pre-programmed monitoring of the drone swarm; the proximity judgment module 12 is used to map the multi-dimensional monitoring data to a high-dimensional order parameter tensor space through an order parameter conversion component, extract critical behavioral features, and perform proximity judgment of phase transition critical points to generate risk control instructions; the risk control management module 13 is used to activate the disaster management module according to the risk control instructions, determine the first vulnerability heatmap, the first worst-case development path, and the first intervention measures, and conduct proactive monitoring and collaborative risk control management of the bank area based on the drone swarm; wherein, the disaster management module includes a first predictor and a second predictor based on first-order adversarial training and second-order game training.

[0064] Furthermore, the system is also used to implement the following functions: The order parameters include thermodynamic entropy gradient, information structure complexity, environmental resonance mode, and material phase transition barrier. By performing order parameter transformation on the multivariate monitoring data, the high-dimensional order parameter tensor space is dynamically updated, and critical behavioral features are extracted. If the phase transition point is approached, the risk control instruction is generated.

[0065] Furthermore, the system is also used to implement the following functions: Identify critical characteristics of behavior and determine whether they meet preset conditions. If they do, it is determined to be approaching the critical point of phase transition. Among them, at least one of the following preset conditions is met: correlation length divergence, exponential growth of relaxation time, and intensified fluctuation of order parameter.

[0066] Furthermore, the system is also used to implement the following functions: A first predictor is deployed with a vulnerability heatmap based on fire development and environmental conditions as the first generation target and the worst development path as the second generation target; a second predictor is deployed with intervention measures based on fire development and environmental conditions; and a disaster management module is generated by performing two-stage training on the first and second predictors.

[0067] Furthermore, the system is also used to implement the following functions: By introducing a judge, the adversarial training of the first predictor and the second predictor in parallel is used as the first-order training method; the incomplete information dynamic game training of the first predictor and the second predictor after the first-order training is used as the second-order training method.

[0068] Furthermore, the system is also used to implement the following functions: According to the risk control command, the disaster management module is activated, and the dynamically updated high-dimensional order parameter tensor space is input into the disaster management module; the first predictor and the second predictor are driven in parallel to output the first prediction data and the second prediction data within a preset time window; based on the first prediction data and the second prediction data, the directional monitoring management and risk control early warning management of the drone swarm are executed.

[0069] Furthermore, the system is also used to implement the following functions: Based on the first vulnerability heatmap in the first prediction data, a first monitoring target is determined, and a second monitoring target is determined based on the first worst-case development path. The first and second monitoring targets are prioritized and their monitoring trajectories are planned to determine a pre-monitoring strategy. The pre-monitoring strategy is then distributed to the central control units of each UAV in the UAV cluster to perform targeted monitoring management.

[0070] Furthermore, the system is also used to implement the following functions: For the first intervention measure in the second predicted data, the first intervention measure is decoupled using the smallest risk control unit to determine multiple intervention points; the multiple intervention points are identified by spatiotemporal codes based on the spatiotemporal relationship of risk control to generate multi-threaded early warning instructions; and the multi-threaded early warning instructions are managed by site-centralized early warning and multi-threaded distribution.

[0071] Furthermore, the system is also used to implement the following functions: Obtain risk control record data, mine strategy deviation features based on the risk control record data, wherein the strategy deviation features satisfy a preset generalization degree; and learn and update the first predictor and the second predictor based on the strategy deviation features.

[0072] It should be noted that the order of the embodiments described above is merely for descriptive purposes and does not represent the superiority or inferiority of the embodiments. Furthermore, the above description focuses on specific embodiments of this specification. The processes depicted in the accompanying drawings do not necessarily require a specific or sequential order to achieve the desired results. In some implementations, multitasking and parallel processing are possible or may be advantageous.

[0073] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention in any way. Although the present invention has been disclosed above with reference to preferred embodiments, it is not intended to limit the present invention. Any person skilled in the art can make some modifications or alterations to the above-disclosed technical content to create equivalent embodiments without departing from the scope of the present invention. Any modifications, equivalent changes, and alterations made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the scope of the present invention.

Claims

1. A method for early warning of fires in bank areas based on unmanned aerial vehicles (UAVs), characterized in that, The method includes: Based on pre-programmed monitoring of drone swarms, obtain diverse monitoring data; The multidimensional monitoring data is mapped to a high-dimensional order parameter tensor space through the order parameter conversion component, critical behavior features are extracted and the approach to the phase transition critical point is determined to generate risk control instructions. The disaster management module is activated according to the risk control instructions to determine the first vulnerability heat map, the first worst-case development path and the first intervention measures, and to carry out proactive monitoring and collaborative risk control management of the bank area based on drone swarms. The disaster management module includes a first predictor and a second predictor based on first-order adversarial training and second-order game training.

2. The method for early warning of fires in bank areas based on unmanned aerial vehicles as described in claim 1, characterized in that, The order parameters include the thermodynamic entropy gradient, information structure complexity, environmental resonance modes, and material phase transition barriers; By performing ordinal parameter transformation on the multivariate monitoring data, the high-dimensional ordinal parameter tensor space is dynamically updated, and critical behavioral features are extracted. If the data approaches a phase transition point, the risk control instruction is generated.

3. The method for early warning of fires in bank areas based on unmanned aerial vehicles as described in claim 2, characterized in that, Identify critical characteristics of behavior and determine whether they meet preset conditions. If they do, it is determined that the behavior is approaching the critical point of phase transition. Among them, at least one of the following is a preset condition: divergence of correlation length, exponential growth of relaxation time, and intensified fluctuation of order parameter.

4. The method for early warning of fires in bank areas based on unmanned aerial vehicles as described in claim 1, characterized in that, Before activating the disaster management module, the construction of the disaster management module includes: The first predictor is deployed with a vulnerability heat map based on fire development and environmental conditions as the first generation target and the worst development path as the second generation target. A second predictor is deployed based on intervention measures generated according to fire development and environmental conditions; The first and second predictors are trained in two stages to generate a disaster management module.

5. The method for early warning of fires in bank areas based on unmanned aerial vehicles as described in claim 4, characterized in that, Two-stage training is performed on the first predictor and the second predictor, including: By introducing a judge, the adversarial training of the first predictor and the second predictor is performed in parallel as a first-order training method. The incomplete information dynamic game training between the first and second predictors after first-order training is used as the second-order training method.

6. The method for early warning of fires in bank areas based on unmanned aerial vehicles as described in claim 5, characterized in that, Identify the primary vulnerability heatmap, the first worst-case scenario, and the primary intervention measures; conduct proactive monitoring and collaborative risk control management of the banking area based on drone swarms, including: According to the risk control command, the disaster management module is activated, and the dynamically updated high-dimensional order parameter tensor space is input into the disaster management module. The first and second predictors are driven in parallel to output the first and second prediction data within a preset time window. Based on the first and second prediction data, targeted monitoring and risk control early warning management of the drone swarm is implemented.

7. The method for early warning of fires in bank areas based on unmanned aerial vehicles as described in claim 6, characterized in that, Perform targeted monitoring and management of drone swarms, including: Based on the first vulnerability heatmap in the first prediction data, the first monitoring target is determined, and based on the first worst-case development path, the second monitoring target is determined. Prioritize and plan monitoring trajectories for the first and second monitoring targets to determine pre-monitoring strategies; The pre-monitoring strategy is distributed to the central control units of each drone in the drone cluster for targeted monitoring and management.

8. The method for early warning of fires in bank areas based on unmanned aerial vehicles as described in claim 6, characterized in that, Implement risk control and early warning management, including: For the first intervention measure in the second predicted data, the first intervention measure is decoupled using the smallest risk control unit to determine multiple intervention points; Based on the spatiotemporal relationship of risk control, the multiple intervention points are identified by spatiotemporal codes to generate multi-threaded early warning instructions; The multi-threaded early warning commands are managed for site-wide central control early warning and multi-threaded distribution.

9. The method for early warning of fires in bank areas based on unmanned aerial vehicles as described in claim 1, characterized in that, After implementing proactive monitoring and collaborative risk control management of the banking area based on drone swarms, the following is included: Obtain risk control record data, and mine strategy deviation features based on the risk control record data, wherein the strategy deviation features satisfy a preset generalization degree; The first predictor and the second predictor are learned and updated based on the policy bias characteristics.

10. An early warning system for fires in bank areas based on unmanned aerial vehicles (UAVs), characterized in that: The system is used to execute the drone-based early warning method for bank area fires as described in any one of claims 1-9, and the system includes: The monitoring module is used to acquire diverse monitoring data based on pre-programmed monitoring of the drone swarm; The proximity determination module is used to map the multi-dimensional monitoring data to a high-dimensional order parameter tensor space through the order parameter conversion component, extract critical behavior features, determine the proximity of the phase transition critical point, and generate risk control instructions. The risk control management module is used to activate the disaster management module according to the risk control instructions, determine the first vulnerability heat map, the first worst development path and the first intervention measures, and conduct proactive monitoring and collaborative risk control management of the bank area based on drone swarms. The disaster management module includes a first predictor and a second predictor based on first-order adversarial training and second-order game training.

Images