Low altitude flight safety intelligent monitoring device based on big data analysis

Abstract

This invention discloses an intelligent monitoring device for low-altitude flight safety based on big data analysis, comprising: a multi-source heterogeneous data acquisition and preprocessing module that outputs structured feature vectors, raw data streams, and multi-source time series data; an attention mechanism adaptive fusion module that dynamically weights and fuses multi-source time series data according to the airspace context in the airspace data; a deep nonlinear risk assessment core module that constructs an airspace entity relationship graph based on the fused feature vectors and performs nonlinear feature cross-calculation, outputting the aircraft's dynamic risk index and feature contribution explanation; a two-layer adaptive early warning decision module that triggers graded early warnings based on the comparison results between the dynamic risk index and the real-time airspace risk baseline, and optimizes the early warning response actions based on a reinforcement learning strategy; and an abnormal behavior detection submodule that performs real-time pattern analysis on the raw data stream to detect abnormal aircraft operation behaviors.

Description

Technical Field

[0001] This invention relates to the field of low-altitude flight technology, and in particular to a low-altitude flight safety intelligent monitoring device based on big data analysis. Background Technology

[0002] In the context of the surge in flight frequency, increased airspace complexity, and prominent potential safety risks brought about by the widespread application of aircraft such as drones, light aircraft, and helicopters in the low-altitude flight field, traditional manual monitoring or single-sensor monitoring modes have limitations such as limited coverage, delayed real-time response, insufficient data fusion capabilities, and low accuracy in risk prediction. This device aims to integrate multi-source heterogeneous data acquisition modules, construct a low-altitude flight big data platform, and use machine learning algorithms to analyze and process key parameters such as flight trajectory, meteorological environment, electromagnetic interference, and equipment status in real time. It realizes multi-dimensional intelligent monitoring functions such as dynamic trajectory tracking of aircraft in low-altitude airspace, abnormal behavior identification, potential collision warning, illegal intrusion interception, and rapid response to emergency events. This improves the efficiency and accuracy of low-altitude flight safety monitoring, effectively reduces the probability of flight accidents, protects the safety of people and property and public airspace order, further promotes the healthy and orderly development of the low-altitude economy, and facilitates the improvement of relevant technical standards and regulatory systems. It has significant socio-economic benefits and technological innovation value.

[0003] In existing technologies, the fusion of data from different sources (such as status, environment, and airspace) often adopts fixed or simple rule-based weights, which cannot adapt to the ever-changing specific scenarios. Therefore, a low-altitude flight safety intelligent monitoring device based on big data analysis is proposed. Summary of the Invention

[0004] The purpose of this invention is to address the shortcomings of existing technologies by proposing a low-altitude flight safety intelligent monitoring device based on big data analysis.

[0005] To achieve the above objectives, the present invention adopts the following technical solution: A low-altitude flight safety intelligent monitoring device based on big data analysis includes a multi-source heterogeneous data acquisition and preprocessing module, an attention mechanism adaptive fusion module, a deep nonlinear risk assessment core module, and a two-layer adaptive early warning decision module connected in sequence; an abnormal behavior detection sub-module connected to the multi-source heterogeneous data acquisition and preprocessing module and the deep nonlinear risk assessment core module; and a feedback closed loop and online learning module connected to the two-layer adaptive early warning decision module and the deep nonlinear risk assessment core module. The multi-source heterogeneous data acquisition and preprocessing module is used to acquire aircraft status, environment and airspace data and perform robust filtering processing to output structured feature vectors, raw data streams and multi-source time series data; The attention mechanism adaptive fusion module is used to dynamically weight and fuse multi-source time series data according to the spatial context in the spatial data, and then send the weighted fused feature vector to the deep nonlinear risk assessment core module. The deep nonlinear risk assessment core module is used to construct an airspace entity relationship graph based on the fused feature vectors and perform nonlinear feature cross calculations to output the dynamic risk index and feature contribution explanation of the aircraft. The dual-layer adaptive early warning decision module is used to trigger graded early warnings based on the comparison results between the dynamic risk index and the real-time airspace risk baseline, and to optimize the early warning response actions based on reinforcement learning strategies. The abnormal behavior detection submodule is used to perform real-time pattern analysis on the raw data stream to detect abnormal operating behavior of the aircraft, and synchronously feeds the detection results back to the deep nonlinear risk assessment core module to assist in correcting the dynamic risk index. The feedback loop and online learning module is used to perform incremental learning and dynamic optimization of the risk assessment model and fusion weight strategy based on the verification information of the early warning results.

[0006] The above technical solution further includes: Furthermore, the multi-source heterogeneous data acquisition and preprocessing module integrates a GNSS positioning module, weather radar, millimeter-wave radar, visual sensor, and ADS-B receiver to collect real-time data on the aircraft's position and speed, meteorological elements, near-range obstacle point cloud data, visual images and video streams, and airspace traffic broadcast information, forming the original multi-source heterogeneous data stream. The multi-source heterogeneous data stream is aligned and synchronized according to a unified timestamp reference, and the unstructured data in the multi-source heterogeneous data stream is converted into standardized data packets containing timestamps, sensor IDs, data values, and confidence fields, and integrated into multi-source time series data. For the position and velocity data provided by the GNSS positioning module and millimeter-wave radar, Kalman filtering is used for optimal estimation and prediction, and median filtering is coupled to perform sliding window processing on the filtered output sequence to suppress sudden impulse noise interference and generate smoothed high-confidence state estimation data. The difference between the smoothed high-confidence state estimation data and the corresponding original sensor readings is calculated to generate filtered residual data. The filtered residual data is statistically analyzed to extract its mean, variance, and abrupt change point features, forming a fault feature vector for sensor fault prediction. By fusing the smoothed high-confidence state estimation data, meteorological element data from weather radar, and airspace traffic broadcast information data from ADS-B receiver, feature vectors of dimensions such as aircraft attitude angle, ambient wind speed, and airspace target density are extracted and assembled into structured feature vectors.

[0007] Furthermore, the core module of the deep nonlinear risk assessment uses a graph neural network to construct an airspace entity relationship graph with aircraft, obstacles and weather stations as nodes and spatial distance, relative speed and meteorological influence as edges; This module also includes a deep neural network feature cross layer, which is used to automatically generate the product of airspace density and wind speed, the product of aircraft type and visibility, and other nonlinear interactive features. The integrated SHAP value calculation engine generates a feature contribution heatmap describing the degree of contribution of each input feature to the current risk while outputting the dynamic risk index, thereby making the model decision-making process interpretable.

[0008] Furthermore, the attention mechanism adaptive fusion module is equipped with a multi-head spatiotemporal attention mechanism, which can dynamically calculate the feature weight coefficients of different data sources according to the real-time spatial domain scene. The module also includes an airspace type encoder and a time encoder. The airspace type encoder is used to identify scenarios such as airport terminal areas, urban canyons or rural airspace, and the time encoder is used to identify time patterns such as peak hours or night. The contextual instructions generated by the encoder are used to guide the attention mechanism to adjust weights, such as automatically reducing the weight of GNSS data and increasing the weight of visual and radar data in urban canyon scenes, or significantly increasing the weight of meteorological data under severe weather conditions.

[0009] Furthermore, the dual-layer adaptive early warning decision module includes a first-layer dynamic baseline calibration unit and a second-layer reinforcement learning strategy optimization unit; The dynamic baseline calibration unit uses a recurrent neural network to analyze the overall airspace situation in real time and calculate the dynamic risk baseline value. The original risk index output by the deep nonlinear risk assessment core module is compared with the baseline value to obtain an effective risk index for triggering early warning. The reinforcement learning strategy optimization unit constructs a reinforcement learning environment with risk index, airspace traffic and aircraft status as the state space and warning level and specific response action as the action space, and iteratively optimizes the warning strategy through a reward function that integrates safety and efficiency indicators.

[0010] Furthermore, the dual-layer adaptive early warning decision module is also equipped with a human-machine collaboration interface, which provides a strategy sandbox interface that allows airspace administrators to intervene. Administrators can temporarily adjust the warning threshold at specific times or for specific areas through the policy sandbox interface, or insert policy templates based on expert experience. These manual adjustment instructions will work in conjunction with the policies automatically generated by the reinforcement learning policy optimization unit to jointly constitute the final warning decision.

[0011] Furthermore, the abnormal behavior detection submodule is equipped with an anomaly detector based on the isolated forest algorithm, which continuously analyzes the raw sensor data stream and filtered residual data from the preprocessing module. The abnormal behavior detection submodule identifies aircraft behaviors that deviate from historical normal operating patterns, including sudden abnormal changes in acceleration, unexpected course deviations, or abnormal interruptions in communication signals. It synchronously transmits the identified abnormal behavior events and their confidence levels to the deep nonlinear risk assessment core module and the dual-layer adaptive early warning decision module to trigger local recalculation of the risk assessment or upgrade the early warning level.

[0012] Furthermore, the feedback loop and online learning module includes an incremental learning pipeline and an offline reinforcement learning simulation environment; The incremental learning pipeline is used to automatically generate labeled training samples after each early warning action is confirmed or disproven, and to fine-tune the model parameters in the deep nonlinear risk assessment core module using an online support vector machine classifier. The offline reinforcement learning simulation environment is used to run agent replicas of the reinforcement learning policy optimization unit, verify the effectiveness and stability of the new policy in a simulated spatial scenario, and when the simulation performance index is consistently better than the current policy, the verified new policy is migrated to the real decision-making module for gradual application.

[0013] Furthermore, the device internally forms a clear data flow logic, including a positive data flow from the sensor through preprocessing, fusion, evaluation to decision execution; And the data flow that optimizes model parameters and fusion weights from the early warning decision results through the feedback closed-loop module, and then back-acts on the risk assessment module and the fusion module; It also includes a horizontal auxiliary data flow that provides auxiliary judgment information directly from the anomaly detection submodule to the risk assessment module.

[0014] The present invention has the following beneficial effects: In this invention, an attention mechanism is used to autonomously and dynamically assign importance weights to data from different times and sources. In urban canyons, visual / radar data is automatically given higher weights to compensate for GNSS signal attenuation. In severe weather, the influence of meteorological data is significantly enhanced. This context-aware fusion method ensures that the most critical information for risk assessment is highlighted in any scenario, significantly reducing assessment bias caused by improper data fusion. Attached Figure Description

[0015] Figure 1 This is a system block diagram of a low-altitude flight safety intelligent monitoring device based on big data analysis proposed in this invention. Detailed Implementation

[0016] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0017] Please see Figure 1 As shown, the present invention is a low-altitude flight safety intelligent monitoring device based on big data analysis, comprising a multi-source heterogeneous data acquisition and preprocessing module, an attention mechanism adaptive fusion module, a deep nonlinear risk assessment core module, and a two-layer adaptive early warning decision module connected in sequence; an abnormal behavior detection sub-module connected to the multi-source heterogeneous data acquisition and preprocessing module and the deep nonlinear risk assessment core module; and a feedback closed loop and online learning module connected to the two-layer adaptive early warning decision module and the deep nonlinear risk assessment core module. The multi-source heterogeneous data acquisition and preprocessing module is used to acquire aircraft status, environment and airspace data and perform robust filtering processing to output structured feature vectors, raw data streams and multi-source time series data; The attention mechanism adaptive fusion module is used to dynamically weight and fuse multi-source time series data according to the spatial context in the spatial data, and then send the weighted fused feature vector to the deep nonlinear risk assessment core module. The deep nonlinear risk assessment core module is used to construct an airspace entity relationship graph based on the fused feature vectors and perform nonlinear feature cross calculations to output the dynamic risk index and feature contribution explanation of the aircraft. The dual-layer adaptive early warning decision module is used to trigger graded early warnings based on the comparison results between the dynamic risk index and the real-time airspace risk baseline, and to optimize the early warning response actions based on reinforcement learning strategies. The abnormal behavior detection submodule is used to perform real-time pattern analysis on the raw data stream to detect abnormal operating behavior of the aircraft, and synchronously feeds the detection results back to the deep nonlinear risk assessment core module to assist in correcting the dynamic risk index. The feedback loop and online learning module is used to perform incremental learning and dynamic optimization of the risk assessment model and fusion weight strategy based on the verification information of the early warning results.

[0018] In one embodiment, the multi-source heterogeneous data acquisition and preprocessing module integrates a GNSS positioning module, a weather radar, a millimeter-wave radar, a visual sensor, and an ADS-B receiver. It collects the aircraft's position and speed data, meteorological element data, near-range obstacle point cloud data, visual image and video stream data, and airspace traffic broadcast information data in real time, forming the original multi-source heterogeneous data stream. The multi-source heterogeneous data stream is aligned and synchronized according to a unified timestamp reference, and the unstructured data in the multi-source heterogeneous data stream is converted into standardized data packets containing timestamps, sensor IDs, data values, and confidence fields, and integrated into multi-source time series data. For the position and velocity data provided by the GNSS positioning module and millimeter-wave radar, Kalman filtering is used for optimal estimation and prediction, and median filtering is coupled to perform sliding window processing on the filtered output sequence to suppress sudden impulse noise interference and generate smoothed high-confidence state estimation data. The difference between the smoothed high-confidence state estimation data and the corresponding original sensor readings is calculated to generate filtered residual data. The filtered residual data is statistically analyzed to extract its mean, variance, and abrupt change point features, forming a fault feature vector for sensor fault prediction. By fusing the smoothed high-confidence state estimation data, meteorological element data from weather radar, and airspace traffic broadcast information data from ADS-B receiver, feature vectors of dimensions such as aircraft attitude angle, ambient wind speed, and airspace target density are extracted and assembled into structured feature vectors.

[0019] 3. The device according to claim 1, wherein the deep nonlinear risk assessment core module uses a graph neural network to construct an airspace entity relationship graph with aircraft, obstacles and weather stations as nodes and spatial distance, relative speed and meteorological influence as edges; This module also includes a deep neural network feature cross layer, which is used to automatically generate the product of airspace density and wind speed, the product of aircraft type and visibility, and other nonlinear interactive features. The integrated SHAP value calculation engine generates a feature contribution heatmap describing the degree of contribution of each input feature to the current risk while outputting the dynamic risk index, thereby making the model decision-making process interpretable.

[0020] In one embodiment, the attention mechanism adaptive fusion module is deployed with a multi-head spatiotemporal attention mechanism, which can dynamically calculate the feature weight coefficients of different data sources according to the real-time spatial domain scene. The module also includes an airspace type encoder and a time encoder. The airspace type encoder is used to identify scenarios such as airport terminal areas, urban canyons or rural airspace, and the time encoder is used to identify time patterns such as peak hours or night. The contextual instructions generated by the encoder are used to guide the attention mechanism to adjust weights, such as automatically reducing the weight of GNSS data and increasing the weight of visual and radar data in urban canyon scenes, or significantly increasing the weight of meteorological data under severe weather conditions.

[0021] In one embodiment, the two-layer adaptive early warning decision module includes a first-layer dynamic baseline calibration unit and a second-layer reinforcement learning policy optimization unit. The dynamic baseline calibration unit uses a recurrent neural network to analyze the overall airspace situation in real time and calculate the dynamic risk baseline value. The original risk index output by the deep nonlinear risk assessment core module is compared with the baseline value to obtain an effective risk index for triggering early warning. The reinforcement learning strategy optimization unit constructs a reinforcement learning environment with risk index, airspace traffic and aircraft status as the state space and warning level and specific response action as the action space, and iteratively optimizes the warning strategy through a reward function that integrates safety and efficiency indicators.

[0022] In one embodiment, the dual-layer adaptive early warning decision module is further provided with a human-machine collaboration interface, which provides a strategy sandbox interface that allows airspace administrators to intervene. Administrators can temporarily adjust the warning threshold at specific times or for specific areas through the policy sandbox interface, or insert policy templates based on expert experience. These manual adjustment instructions will work in conjunction with the policies automatically generated by the reinforcement learning policy optimization unit to jointly constitute the final warning decision.

[0023] In one embodiment, the abnormal behavior detection submodule is equipped with an anomaly detector based on the isolated forest algorithm, which continuously analyzes the raw sensor data stream and filtered residual data from the preprocessing module. The abnormal behavior detection submodule identifies aircraft behaviors that deviate from historical normal operating patterns, including sudden abnormal changes in acceleration, unexpected course deviations, or abnormal interruptions in communication signals. It synchronously transmits the identified abnormal behavior events and their confidence levels to the deep nonlinear risk assessment core module and the dual-layer adaptive early warning decision module to trigger local recalculation of the risk assessment or upgrade the early warning level.

[0024] In one embodiment, the feedback loop and online learning module includes an incremental learning pipeline and an offline reinforcement learning simulation environment; The incremental learning pipeline is used to automatically generate labeled training samples after each early warning action is confirmed or disproven, and to fine-tune the model parameters in the deep nonlinear risk assessment core module using an online support vector machine classifier. The offline reinforcement learning simulation environment is used to run agent replicas of the reinforcement learning policy optimization unit, verify the effectiveness and stability of the new policy in a simulated spatial scenario, and when the simulation performance index is consistently better than the current policy, the verified new policy is migrated to the real decision-making module for gradual application.

[0025] In one embodiment, the device internally forms a clear data flow logic, including a positive data flow from the sensor through preprocessing, fusion, evaluation to decision execution; And the data flow that optimizes model parameters and fusion weights from the early warning decision results through the feedback closed-loop module, and then back-acts on the risk assessment module and the fusion module; It also includes a horizontal auxiliary data flow that provides auxiliary judgment information directly from the anomaly detection submodule to the risk assessment module.

[0026] All data obtained in this invention has been authorized by the user.

[0027] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and variations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A low-altitude flight safety intelligent monitoring device based on big data analysis, characterized in that, It includes a multi-source heterogeneous data acquisition and preprocessing module, an attention mechanism adaptive fusion module, a deep nonlinear risk assessment core module, and a two-layer adaptive early warning decision module connected in sequence; an abnormal behavior detection sub-module connected to the multi-source heterogeneous data acquisition and preprocessing module and the deep nonlinear risk assessment core module; and a feedback loop and online learning module connected to the two-layer adaptive early warning decision module and the deep nonlinear risk assessment core module. The multi-source heterogeneous data acquisition and preprocessing module is used to acquire aircraft status, environment and airspace data and perform robust filtering processing to output structured feature vectors, raw data streams and multi-source time series data; The attention mechanism adaptive fusion module is used to dynamically weight and fuse multi-source time series data according to the spatial context in the spatial data, and then send the weighted fused feature vector to the deep nonlinear risk assessment core module. The deep nonlinear risk assessment core module is used to construct an airspace entity relationship graph based on the fused feature vectors and perform nonlinear feature cross calculations to output the dynamic risk index and feature contribution explanation of the aircraft. The dual-layer adaptive early warning decision module is used to trigger graded early warnings based on the comparison results between the dynamic risk index and the real-time airspace risk baseline, and to optimize the early warning response actions based on reinforcement learning strategies. The abnormal behavior detection submodule is used to perform real-time pattern analysis on the raw data stream to detect abnormal operating behavior of the aircraft, and synchronously feeds the detection results back to the deep nonlinear risk assessment core module to assist in correcting the dynamic risk index. The feedback loop and online learning module is used to perform incremental learning and dynamic optimization of the risk assessment model and fusion weight strategy based on the verification information of the early warning results.

2. The apparatus according to claim 1, characterized in that, The multi-source heterogeneous data acquisition and preprocessing module integrates a GNSS positioning module, weather radar, millimeter-wave radar, visual sensor, and ADS-B receiver. It collects the aircraft's position and velocity data, meteorological element data, near-range obstacle point cloud data, visual image and video stream data, and airspace traffic broadcast information data in real time, forming the original multi-source heterogeneous data stream. The multi-source heterogeneous data stream is aligned and synchronized according to a unified timestamp reference, and the unstructured data in the multi-source heterogeneous data stream is converted into standardized data packets containing timestamps, sensor IDs, data values, and confidence fields, and integrated into multi-source time series data. For the position and velocity data provided by the GNSS positioning module and millimeter-wave radar, Kalman filtering is used for optimal estimation and prediction, and median filtering is coupled to perform sliding window processing on the filtered output sequence to generate smoothed high-confidence state estimation data. The difference between the smoothed high-confidence state estimation data and the corresponding original sensor readings is calculated to generate filtered residual data. The filtered residual data is statistically analyzed to extract its mean, variance, and abrupt change point features, forming a fault feature vector for sensor fault prediction. The smoothed high-confidence state estimation data, meteorological element data from weather radar, and airspace traffic broadcast information data from ADS-B receiver are fused together to extract feature vectors and assemble them into structured feature vectors.

3. The apparatus according to claim 1, characterized in that, The core module of the deep nonlinear risk assessment uses a graph neural network to construct an airspace entity relationship graph with aircraft, obstacles and weather stations as nodes and spatial distance, relative speed and meteorological influence as edges; This module also includes a deep neural network feature cross layer, which is used to automatically generate the product of airspace density and wind speed, the product of aircraft type and visibility, and other nonlinear interactive features. The integrated SHAP value calculation engine generates a feature contribution heatmap describing the degree of contribution of each input feature to the current risk while outputting the dynamic risk index, thereby making the model decision-making process interpretable.

4. The apparatus according to claim 1, characterized in that, The adaptive fusion module of the attention mechanism is deployed with a multi-head spatiotemporal attention mechanism, which can dynamically calculate the feature weight coefficients of different data sources according to the real-time spatial domain scene. The module also includes an airspace type encoder and a time encoder. The airspace type encoder is used to identify scenarios such as airport terminal areas, urban canyons or rural airspace, and the time encoder is used to identify time patterns such as peak hours or night. The contextual instructions generated by the encoder are used to guide the attention mechanism to adjust weights, such as automatically reducing the weight of GNSS data and increasing the weight of visual and radar data in urban canyon scenes, or significantly increasing the weight of meteorological data under severe weather conditions.

5. The apparatus according to claim 1, characterized in that, The dual-layer adaptive early warning decision module includes a first-layer dynamic baseline calibration unit and a second-layer reinforcement learning strategy optimization unit. The dynamic baseline calibration unit uses a recurrent neural network to analyze the overall airspace situation in real time and calculate the dynamic risk baseline value. The original risk index output by the deep nonlinear risk assessment core module is compared with the baseline value to obtain an effective risk index for triggering early warning. The reinforcement learning strategy optimization unit constructs a reinforcement learning environment with risk index, airspace traffic and aircraft status as the state space and warning level and specific response action as the action space, and iteratively optimizes the warning strategy through a reward function that integrates safety and efficiency indicators.

6. The apparatus according to claim 5, characterized in that, The dual-layer adaptive early warning decision module is also equipped with a human-machine collaboration interface, which provides a strategy sandbox interface that allows airspace administrators to intervene. Administrators can temporarily adjust the warning threshold at specific times or for specific areas through the policy sandbox interface, or insert policy templates based on expert experience. These manual adjustment instructions will work in conjunction with the policies automatically generated by the reinforcement learning policy optimization unit to jointly constitute the final warning decision.

7. The apparatus according to claim 1, characterized in that, The abnormal behavior detection submodule is equipped with an anomaly detector based on the isolated forest algorithm. The anomaly detector continuously analyzes the raw sensor data stream and filtered residual data from the preprocessing module. The abnormal behavior detection submodule identifies aircraft behaviors that deviate from historical normal operating patterns, including sudden abnormal changes in acceleration, unexpected course deviations, or abnormal interruptions in communication signals. It synchronously transmits the identified abnormal behavior events and their confidence levels to the deep nonlinear risk assessment core module and the dual-layer adaptive early warning decision module to trigger local recalculation of the risk assessment or upgrade the early warning level.

8. The apparatus according to claim 1, characterized in that, The feedback loop and online learning module includes an incremental learning pipeline and an offline reinforcement learning simulation environment. The incremental learning pipeline is used to automatically generate labeled training samples after each early warning action is confirmed or disproven, and to fine-tune the model parameters in the deep nonlinear risk assessment core module using an online support vector machine classifier. The offline reinforcement learning simulation environment is used to run agent replicas of the reinforcement learning policy optimization unit, verify the effectiveness and stability of the new policy in a simulated spatial scenario, and when the simulation performance index is consistently better than the current policy, the verified new policy is migrated to the real decision-making module for gradual application.

9. The apparatus according to claim 1, characterized in that, The device internally forms a clear data flow logic, including a positive data flow from the sensor through preprocessing, fusion, evaluation to decision execution; And the data flow that optimizes model parameters and fusion weights from the early warning decision results through the feedback closed-loop module, and then back-acts on the risk assessment module and the fusion module; It also includes a horizontal auxiliary data flow that provides auxiliary judgment information directly from the anomaly detection submodule to the risk assessment module.

Images