Monitoring data management method and system based on integrated fusion of air, space, land and sea

Abstract

The embodiment of the present application provides a monitoring data management method and system based on the integrated fusion of air, space, land and sea. By dividing the candidate air, space, land and sea monitoring area into multiple monitoring unit areas and collecting data, multi-dimensional fusion feature data of each monitoring unit area is generated, and an attention coefficient is introduced, which can characterize the participation factor of each monitoring unit area in the target abnormal scene category data of the overall candidate air, space, land and sea monitoring area. By weighted fusion of the multi-dimensional fusion feature data of each monitoring unit area and the corresponding attention coefficient, the target multi-dimensional fusion feature data of each monitoring unit area is further optimized, which can more accurately reflect the characteristics of each monitoring unit area and the degree of its influence on the overall abnormal scene. Finally, based on the target multi-dimensional fusion feature data of all monitoring unit areas, a decision is made on the entire candidate air, space, land and sea monitoring area, and target abnormal scene category data with higher accuracy is generated.

Description

Technical Field

[0001] The present invention relates to the field of integrated air-space-ground-sea monitoring technology, and in particular to a monitoring data management method and system based on the integrated fusion of air-space-ground-sea. Background Art

[0002] With the development of science and technology, especially the advancement of remote sensing and drone technology, it has become possible to monitor multiple environments, including land, sea, and air. However, due to the complexity and wide range of monitoring areas, how to effectively and efficiently monitor large-scale, multi-environment candidate air, land, and sea monitoring areas is an urgent problem that needs to be solved.

[0003] Traditional monitoring methods typically collect and process data for each individual monitoring unit area separately, then simply fuse or compare this data to detect anomalies. However, this approach has several drawbacks. First, it ignores the potential interdependencies between different monitoring units, which can lead to biased judgments about the overall situation. Second, because it fails to consider the potential differences in the contribution of each monitoring unit area to the overall target anomaly scenario category, simple fusion or comparison often fails to produce accurate results.

[0004] Therefore, developing a method that can effectively monitor candidate air, land, and sea monitoring areas while fully considering the characteristics and importance of each monitoring unit area is a current technical challenge. The present invention is a solution to this challenge. Summary of the Invention

[0005] In order to at least overcome the above-mentioned deficiencies in the prior art, the embodiment of the present application aims to provide a monitoring data management method and system based on the integrated fusion of air, land, and sea.

[0006] Allocating monitoring units to the candidate air, space, and sea monitoring areas to generate a plurality of monitoring unit areas included in the candidate air, space, and sea monitoring areas;

[0007] Collect data for each of the monitoring unit areas to generate multi-dimensional fusion feature data for each of the monitoring unit areas, wherein the multi-dimensional fusion feature data is fusion feature data of feature data collected separately by satellites, drones, ground sensor networks, and ocean detectors;

[0008] Based on the multi-dimensional fusion feature data of each monitoring unit area, performing a first abnormal feature pattern decision on each monitoring unit area to generate abnormal feature pattern decision data for each monitoring unit area;

[0009] For each of the monitoring unit areas, determining an attention coefficient of the monitoring unit area based on the abnormal feature pattern decision data of the monitoring unit area, and performing weighted fusion on the multidimensional fusion feature data of the monitoring unit area and the attention coefficient of the monitoring unit area to generate target multidimensional fusion feature data of the monitoring unit area, wherein the attention coefficient represents a participation factor of the monitoring unit area in the target abnormal scene category data of the candidate air, land, and sea monitoring areas;

[0010] Based on the target multi-dimensional fusion feature data of the multiple monitoring unit areas, a second abnormal feature pattern decision is performed on the candidate air, space, land and sea monitoring area to generate target abnormal scene category data of the candidate air, space, land and sea monitoring area.

[0011] In a possible implementation of the first aspect, allocating monitoring units to the candidate air, space, and sea monitoring areas to generate a plurality of monitoring unit areas included in the candidate air, space, and sea monitoring areas includes:

[0012] Determine a moving observation frame having target scale measurement parameters corresponding to the candidate air, space, land and sea monitoring area, and a moving interval of the moving observation frame;

[0013] By moving the mobile observation frame according to the moving interval, monitoring units are allocated to the candidate air-space-ground-sea monitoring area, and a plurality of monitoring unit areas included in the candidate air-space-ground-sea monitoring area are generated.

[0014] In a possible implementation of the first aspect, collecting data from each monitoring unit area to generate multi-dimensional fusion feature data for each monitoring unit area includes:

[0015] Performing data collection at a first feature collection depth for each monitoring unit area to generate reference multi-dimensional fusion feature data at the first feature collection depth for each monitoring unit area;

[0016] Fusing the reference multi-dimensional fusion feature data of the first feature acquisition depths corresponding to the multiple monitoring unit areas to generate a reference multi-dimensional fusion feature data array for the candidate air, space, land and sea monitoring area;

[0017] Data collection at a second feature collection depth is performed on the reference multidimensional fusion feature data array to generate a multidimensional fusion feature data array for the candidate air, space, and sea monitoring areas, wherein the multidimensional fusion feature data array includes the multidimensional fusion feature data of the second feature collection depth for each of the monitoring unit areas, and the second feature collection depth is less than the first feature collection depth.

[0018] In a possible implementation of the first aspect, determining the attention coefficient of the monitoring unit area based on the abnormal feature pattern decision data of the monitoring unit area includes:

[0019] Obtaining the number of abnormal feature patterns determined by the first abnormal feature pattern, and determining an equal probability distribution of the number of abnormal feature patterns;

[0020] Determine the deviation between the abnormal feature pattern decision data of the monitoring unit area and the equal probability distribution, and use the deviation as the attention coefficient of the monitoring unit area.

[0021] In a possible implementation of the first aspect, performing a second abnormal feature pattern decision on the candidate air, space, and sea monitoring area based on the target multi-dimensional fusion feature data of the multiple monitoring unit areas to generate target abnormal scene category data for the candidate air, space, and sea monitoring area includes:

[0022] Integrating the target multi-dimensional fusion feature data of the plurality of monitoring unit areas to generate integrated multi-dimensional fusion feature data;

[0023] Based on the integrated multi-dimensional fusion feature data, a second abnormal feature pattern decision is performed on the candidate air-space-ground-sea monitoring area to generate target abnormal scene category data of the candidate air-space-ground-sea monitoring area.

[0024] In a possible implementation of the first aspect, the first abnormal feature pattern decision is implemented by a single prediction subnetwork of an abnormal feature pattern decision network, and the second abnormal feature pattern decision is implemented according to a group prediction subnetwork of the abnormal feature pattern decision network;

[0025] The method further comprises:

[0026] Acquire a sample multidimensional fusion feature data sequence for parameter learning of the abnormal feature pattern decision network, the sample multidimensional fusion feature data sequence comprising a plurality of sample multidimensional fusion feature data, the sample multidimensional fusion feature data carrying abnormal scene category annotation data, the sample multidimensional fusion feature data comprising a plurality of sample monitoring unit areas, the sample monitoring unit areas carrying abnormal feature pattern annotation data;

[0027] Obtaining the sample monitoring unit features of each of the sample monitoring unit areas in the sample multi-dimensional fusion feature data, performing a first abnormal feature pattern decision on each of the sample monitoring unit areas based on the sample monitoring unit features of each of the sample monitoring unit areas according to the individual prediction subnetwork, and generating individual prediction data for each of the sample monitoring unit areas;

[0028] For each of the sample multidimensional fusion feature data, based on the abnormal feature pattern annotation data of each of the sample monitoring unit areas in the sample multidimensional fusion feature data, a second abnormal feature pattern decision is performed on the sample multidimensional fusion feature data through the group prediction subnetwork to generate group prediction data of the sample multidimensional fusion feature data;

[0029] Determining a network cost parameter value of the abnormal feature pattern decision network based on the deviation between the individual prediction data and the abnormal feature pattern annotation data of each sample monitoring unit area, and the deviation between the group prediction data and the abnormal scene category annotation data of the sample multi-dimensional fusion feature data;

[0030] According to the network cost parameter value, the network parameters of the abnormal feature pattern decision network are optimized to perform knowledge learning on the abnormal feature pattern decision network.

[0031] In a possible implementation of the first aspect, determining a network cost parameter value of the abnormal feature pattern decision network based on the deviation between the individual prediction data and the abnormal feature pattern annotation data of each sample monitoring unit area, and the deviation between the group prediction data of the sample multi-dimensional fusion feature data and the abnormal scene category annotation data, includes:

[0032] For each of the sample monitoring unit areas, determining a candidate monomer training cost of the abnormal feature pattern decision network according to a degree of deviation between the monomer prediction data and the abnormal feature pattern labeled data of the sample monitoring unit area;

[0033] Aggregating the training costs of the candidate monomers of the abnormal feature pattern decision network to generate the monomer training cost of the abnormal feature pattern decision network;

[0034] Determining a group training cost of the abnormal feature pattern decision network based on the group prediction data of the sample multi-dimensional fusion feature data and the deviation of the abnormal scene category labeling data;

[0035] According to the first training error influence factor of the individual training cost and the second training error influence factor of the group training cost, the individual training cost and the group training cost are fused and calculated to generate a network cost parameter value of the abnormal feature pattern decision network.

[0036] In a possible implementation of the first aspect, based on the abnormal feature pattern labeling data of each of the sample monitoring unit areas in the sample multidimensional fusion feature data, performing a second abnormal feature pattern decision on the sample multidimensional fusion feature data by the group prediction subnetwork to generate group prediction data for the sample multidimensional fusion feature data includes:

[0037] For each of the sample monitoring unit areas in the sample multidimensional fusion feature data, obtaining the number of abnormal feature patterns determined by the first abnormal feature pattern, determining an equal probability distribution of the number of abnormal feature patterns, determining a degree of deviation between the abnormal feature pattern labeled data of the sample monitoring unit area and the equal probability distribution, using the degree of deviation as a unit attention coefficient of the sample monitoring unit area, and determining a fused sample monitoring unit feature of each of the sample monitoring unit areas based on the sample monitoring unit features and the unit attention coefficient of each of the sample monitoring unit areas;

[0038] Based on the fused sample monitoring unit features of multiple sample monitoring unit areas in the sample multidimensional fusion feature data, a second abnormal feature pattern decision is performed on the sample multidimensional fusion feature data according to the group prediction subnetwork to generate group prediction data of the sample multidimensional fusion feature data.

[0039] In a possible implementation of the first aspect, for each of the sample multidimensional fusion feature data, based on the abnormal feature pattern annotation data of each of the sample monitoring unit areas in the sample multidimensional fusion feature data, performing a second abnormal feature pattern decision on the sample multidimensional fusion feature data by the group prediction subnetwork to generate group prediction data for the sample multidimensional fusion feature data, the method further includes:

[0040] Determining, from the plurality of sample monitoring unit areas in the sample multi-dimensional fusion feature data sequence, key monitoring unit areas of each abnormal feature pattern generated by the first abnormal feature pattern decision, based on the monomer prediction data of each sample monitoring unit area in the sample multi-dimensional fusion feature data sequence;

[0041] For each of the abnormal feature patterns, obtaining a template knowledge representation vector of the abnormal feature pattern and a corresponding template abnormal feature pattern, and momentum optimizing the template knowledge representation vector based on a key monitoring unit area of ​​the abnormal feature pattern to generate an optimized template knowledge representation vector of the abnormal feature pattern;

[0042] For each of the sample monitoring unit areas, determining the degree of match between the multidimensional fusion feature data of the sample monitoring unit area and the optimized template knowledge representation vector of each of the abnormal feature patterns, and taking the template abnormal feature pattern corresponding to the target optimized template knowledge representation vector with the largest degree of match as the unit template abnormal feature pattern of the sample monitoring unit area;

[0043] For each of the sample monitoring unit areas, momentum optimizes the abnormal feature pattern labeling data of the sample monitoring unit area according to the unit template abnormal feature pattern of the sample monitoring unit area to generate optimized abnormal feature pattern labeling data of the sample monitoring unit area;

[0044] The abnormal feature pattern labeling data of each sample monitoring unit area in the sample multidimensional fusion feature data is based on the abnormal feature pattern labeling data, and the second abnormal feature pattern decision is performed on the sample multidimensional fusion feature data by the group prediction subnetwork to generate group prediction data of the sample multidimensional fusion feature data, including:

[0045] Based on the optimized abnormal feature pattern labeling data of each of the sample monitoring unit areas in the sample multidimensional fusion feature data, a second abnormal feature pattern decision is performed on the sample multidimensional fusion feature data according to the group prediction subnetwork to generate group prediction data of the sample multidimensional fusion feature data;

[0046] Determining the network cost parameter value of the abnormal feature pattern decision network based on the deviation between the individual prediction data and the abnormal feature pattern annotation data of each sample monitoring unit area, and the deviation between the group prediction data of the sample multi-dimensional fusion feature data and the abnormal scene category annotation data, includes:

[0047] The network cost parameter value of the abnormal feature pattern decision network is determined based on the deviation between the individual prediction data of each sample monitoring unit area and the optimized abnormal feature pattern annotation data, as well as the deviation between the group prediction data of the sample multi-dimensional fusion feature data and the abnormal scene category annotation data.

[0048] For example, in a possible implementation of the first aspect, the monomer prediction data includes a first confidence level of the sample monitoring unit area matching each of the abnormal feature patterns;

[0049] The determining, based on the monomer prediction data of each sample monitoring unit area in the sample multi-dimensional fusion feature data sequence, from multiple sample monitoring unit areas of the sample multi-dimensional fusion feature data sequence, of key monitoring unit areas of each abnormal feature pattern generated by the first abnormal feature pattern decision includes:

[0050] For each of the abnormal feature patterns, determining, from a plurality of sample monitoring unit areas of the sample multi-dimensional fusion feature data sequence, top N target sample monitoring unit areas that match the abnormal feature pattern and are arranged in descending order of first confidence, where N is an integer greater than 1;

[0051] The first N target sample monitoring unit areas are used as key monitoring unit areas of the abnormal feature pattern;

[0052] The step of momentum optimizing the template knowledge representation vector based on the key monitoring unit area of ​​the abnormal feature pattern to generate the optimized template knowledge representation vector of the abnormal feature pattern includes:

[0053] Momentum-optimize the template knowledge representation vector according to the first key monitoring unit area of ​​the abnormal feature pattern to generate a first intermediate template knowledge representation vector of the abnormal feature pattern;

[0054] According to the kth key monitoring unit area of ​​the abnormal feature pattern, momentum optimize the (k-1)th intermediate template knowledge representation vector of the abnormal feature pattern to generate the kth intermediate template knowledge representation vector of the abnormal feature pattern, where k is greater than 0 and not greater than N;

[0055] Polling the k, generating an Nth intermediate template knowledge representation vector of the abnormal feature pattern, and using the Nth intermediate template knowledge representation vector of the abnormal feature pattern as the optimized template knowledge representation vector of the abnormal feature pattern;

[0056] Alternatively, key multidimensional fusion feature data of each key monitoring unit area of ​​the abnormal feature pattern is obtained, and average multidimensional fusion feature data of a plurality of key multidimensional fusion feature data is determined;

[0057] Obtaining a first importance coefficient of the average multidimensional fusion feature data and a second importance coefficient of the template knowledge representation vector;

[0058] performing a fusion calculation on the average multi-dimensional fusion feature data and the template knowledge representation vector according to the first importance coefficient and the second importance coefficient to generate an optimized template knowledge representation vector of the abnormal feature pattern;

[0059] The step of momentum optimizing the abnormal feature pattern labeling data of the sample monitoring unit area based on the abnormal feature pattern of the unit template of the sample monitoring unit area to generate the optimized abnormal feature pattern labeling data of the sample monitoring unit area includes:

[0060] Obtaining a first mode attention coefficient of the abnormal feature pattern of the unit template and a second mode attention coefficient of the abnormal feature pattern annotation data;

[0061] According to the first pattern attention coefficient and the second pattern attention coefficient, the unit template abnormal feature pattern and the abnormal feature pattern annotation data are fused and calculated to generate optimized abnormal feature pattern annotation data of the sample monitoring unit area.

[0062] According to one aspect of an embodiment of the present application, a remote monitoring system is provided, which includes a processor and a machine-readable storage medium, wherein the machine-readable storage medium stores machine-executable instructions, and the machine-executable instructions are loaded and executed by the processor to implement the monitoring data management method based on the integrated fusion of air, land, and sea in any one of the aforementioned possible implementations.

[0063] According to one aspect of an embodiment of the present application, a computer program product or computer program is provided. The computer program product or computer program includes computer instructions stored in a computer-readable storage medium. A processor of a computer device reads the computer instructions from the computer-readable storage medium and executes the computer instructions, causing the computer device to perform the methods provided in various optional implementations of the above three aspects.

[0064] In the technical solutions provided in some embodiments of the present application, effective monitoring of candidate air, space, and sea monitoring areas is mainly achieved, and corresponding target abnormal scene category data is generated. By dividing the candidate air, space, and sea monitoring area into multiple monitoring unit areas and collecting data, multi-dimensional fusion feature data of each monitoring unit area is generated, thereby achieving detailed and comprehensive monitoring of each monitoring unit area. More importantly, an attention coefficient is also introduced, which can characterize the contribution factor of each monitoring unit area to the target abnormal scene category data of the overall candidate air, space, and sea monitoring area. By weighted fusion of the multi-dimensional fusion feature data of each monitoring unit area and the corresponding attention coefficient, the target multi-dimensional fusion feature data of each monitoring unit area is further optimized, so that it more accurately reflects the characteristics of each monitoring unit area and its influence on the overall abnormal scene. Finally, based on the target multi-dimensional fusion feature data of all monitoring unit areas, a second abnormal feature pattern decision is made for the entire candidate air, space, and sea monitoring area, generating target abnormal scene category data with higher accuracy, which not only provides a basis for further abnormal processing, but also helps in the management and decision-making of the monitoring area. Thus, accurate monitoring and abnormal identification of the candidate air, space, and sea monitoring area are achieved, effectively improving monitoring efficiency and accuracy. BRIEF DESCRIPTION OF THE DRAWINGS

[0065] In order to more clearly illustrate the technical solutions of the embodiments of the present application, the following is a brief introduction to the drawings required to be used in the embodiments. It should be understood that the following drawings only show certain embodiments of the present application and therefore should not be regarded as limiting the scope. For ordinary technicians in this field, other relevant drawings can be extracted in combination with these drawings without creative work.

[0066] Figure 1 A flow chart of a monitoring data management method based on the integrated fusion of air, land, and sea provided in an embodiment of the present application;

[0067] Figure 2 A schematic structural block diagram of a remote monitoring system for implementing the above-mentioned monitoring data management method based on the integrated fusion of air, land, and sea provided in an embodiment of the present application. DETAILED DESCRIPTION

[0068] The following description is intended to enable one of ordinary skill in the art to practice and incorporate the present application, and is provided in the context of a specific application scenario and its requirements. It will be apparent to one of ordinary skill in the art that various modifications may be made to the disclosed embodiments, and that the general principles defined herein may be applied to other embodiments and application scenarios without departing from the principles and scope of the present application. Therefore, the present application is not limited to the described embodiments, but should be accorded the broadest scope consistent with the claims.

[0069] Figure 1 This is a flow chart of a monitoring data management method based on the integrated fusion of air, land, and sea provided by an embodiment of the present application. The monitoring data management method based on the integrated fusion of air, land, and sea is introduced in detail below.

[0070] Step S110 : Allocating monitoring units to the candidate air-space-ground-sea monitoring area to generate a plurality of monitoring unit areas included in the candidate air-space-ground-sea monitoring area.

[0071] For example, if a remote monitoring system is used for forest fire monitoring, the entire monitoring area can be divided into multiple monitoring units, each of which includes a certain range of forest area. These monitoring units can be divided based on factors such as terrain, vegetation type, and climate.

[0072] The candidate air-space, land-space and sea monitoring areas may refer to areas that need to be monitored, and may be areas of any range such as land, sea, sky or space.

[0073] The monitoring unit area refers to dividing the candidate air, space, land and sea monitoring area into multiple sub-areas, and each sub-area is used as a monitoring unit for data collection and analysis.

[0074] Step S120, data is collected for each of the monitoring unit areas to generate multi-dimensional fusion feature data for each of the monitoring unit areas, wherein the multi-dimensional fusion feature data is fusion feature data of feature data collected separately by satellites, drones, ground sensor networks, and ocean detectors.

[0075] For example, for each monitoring unit area, satellite imagery can be used to obtain characteristics such as vegetation cover and temperature, drones can be used to obtain high-altitude image and video data, ground sensor networks can be used to collect environmental data such as temperature, humidity, and wind speed, and ocean probes can be used to obtain data such as seawater temperature and salinity. These data can then be fused together to form multidimensional fused feature data for each monitoring unit area.

[0076] This embodiment can collect various types of data from the monitoring unit area through different sensors and monitoring equipment, such as satellite imagery, drone video, ground sensor data, ocean detector data, etc. In other words, the multi-dimensional fused feature data refers to the fusion of feature data from different sources and data types to form a data representation containing multiple dimensions.

[0077] For example, for each monitoring unit area, satellite imagery can be used to capture characteristics such as vegetation cover and temperature. UAVs can be used to obtain high-altitude imagery and video data. Ground sensor networks can be used to collect environmental data such as temperature, humidity, and wind speed. Ocean detectors can also be used to obtain data such as seawater temperature, salinity, and marine biological activity. This data can then be fused together to form multidimensional fused feature data for each monitoring unit area. This multidimensional fused feature data can more comprehensively reflect the various characteristic information of the monitoring unit area, thereby improving the accuracy and reliability of monitoring.

[0078] Step S130 , performing a first abnormal feature pattern decision on each monitoring unit area based on the multi-dimensional fusion feature data of each monitoring unit area, and generating abnormal feature pattern decision data for each monitoring unit area.

[0079] For example, a machine learning algorithm can be used to train the multidimensional fused feature data of each monitoring unit area to learn characteristic patterns under various abnormal conditions, thereby generating abnormal feature pattern decision data. This abnormal feature pattern decision data may include the abnormal probability value of each abnormal feature pattern, indicating whether an abnormal feature pattern exists in the monitoring unit area, and the category of the abnormal feature pattern. For example, the category of abnormal feature patterns can be determined based on the specific application scenario and monitoring objectives of the remote monitoring system and may include multiple types. For example, in forest fire monitoring, abnormal feature patterns may include abnormal fire source temperature, abnormal smoke concentration, abnormal flame radiation, etc. In earthquake monitoring, abnormal feature patterns may include abnormal seismic waves, abnormal ground displacement, abnormal magnetic field, etc. In marine pollution monitoring, abnormal feature patterns may include abnormal pollutant concentration, abnormal water quality, abnormal marine life, etc. These abnormal feature patterns can be obtained by analyzing, processing, and identifying the multidimensional fused feature data. In the remote monitoring system, various machine learning algorithms can be used to identify these abnormal feature patterns. By analyzing, processing, and standardizing different types of abnormal feature patterns, it is possible to more accurately determine whether an abnormal situation exists in the monitoring area and provide more reliable data support for decision-making and analysis.

[0080] Step S140: For each of the monitoring unit areas, the attention coefficient of the monitoring unit area is determined based on the abnormal feature pattern decision data of the monitoring unit area, and the multidimensional fusion feature data of the monitoring unit area and the attention coefficient of the monitoring unit area are weightedly fused to generate the target multidimensional fusion feature data of the monitoring unit area. The attention coefficient represents the participation factor of the monitoring unit area in the target abnormal scene category data of the candidate air, space, land and sea monitoring areas.

[0081] For example, this step can utilize an attention mechanism to assign an attention coefficient to each monitoring unit area based on its abnormal feature pattern decision data. This attention coefficient represents the contribution of that monitoring unit area to the target abnormal scene category data for the entire monitoring area. The multidimensional fused feature data for each monitoring unit area can then be weightedly fused with the attention coefficient to generate the target multidimensional fused feature data.

[0082] In other words, the purpose of this process is to achieve weighted fusion of features from different regions by weighted fusion of the attention coefficients of different monitoring unit areas, thereby generating more accurate target multi-dimensional fusion feature data that contains global information. Specifically, in some cases, certain monitoring unit areas may have more abnormal features or be more important, so the attention coefficients of these areas need to be weighted so that they occupy a larger proportion in the target multi-dimensional fusion feature data. This weighted fusion method based on the attention mechanism can effectively improve the accuracy and reliability of the monitoring system, can more accurately identify abnormal scenes, and provide strong support for further decision-making and analysis.

[0083] Step S150 , performing a second abnormal feature pattern decision on the candidate air, space, land, and sea monitoring area based on the target multi-dimensional fusion feature data of the multiple monitoring unit areas, and generating target abnormal scene category data for the candidate air, space, land, and sea monitoring area.

[0084] For example, another machine learning algorithm can be used to train the target multi-dimensional fusion feature data of multiple monitoring unit areas in the entire candidate air, space, and sea monitoring area to learn the characteristic patterns under various abnormal scene categories. Then, target abnormal scene category data can be generated for the entire candidate air, space, and sea monitoring area.

[0085] That is, this process can be regarded as a global anomaly detection process, which is to fuse and analyze the local anomaly features of multiple monitoring unit areas, so as to obtain the abnormal scene category data of the entire candidate air, space, land and sea monitoring area.

[0086] Specifically, the multi-dimensional feature data of targets across multiple monitoring units is first fused to form a global feature representation. A machine learning algorithm is then used to learn and predict this global feature representation, identifying the types of abnormal scenarios present across the entire candidate air, land, and sea monitoring area.

[0087] The target abnormal scene category data refers to the types of abnormal conditions in the entire candidate air, space, and sea monitoring area monitored by the remote monitoring system. These abnormal conditions may include natural disasters, environmental pollution, man-made damage, and the like. For example, in forest fire monitoring, abnormal scene categories may include forest fires, fire source spread, smoke spread, and the like. In earthquake monitoring, abnormal scene categories may include earthquakes, crustal movement, volcanic eruptions, and the like. In marine pollution monitoring, abnormal scene categories may include marine pollution, water quality deterioration, marine life death, and the like. These abnormal scene categories can be obtained by analyzing, processing, and identifying multi-dimensional fusion feature data. In the remote monitoring system, various machine learning algorithms can be used to identify these abnormal scene categories, and by analyzing, processing, and standardizing different types of abnormal scene categories, it is possible to more accurately determine whether there are abnormal conditions in the monitoring area and provide more reliable data support for decision-making and analysis.

[0088] Therefore, through this global anomaly detection method, the accuracy and reliability of the monitoring system can be effectively improved, the false alarm and missed alarm rates can be reduced, and more intelligent and efficient monitoring and management can be achieved.

[0089] Based on the above steps, the embodiment of the present application mainly realizes the effective monitoring of the candidate air, space, land and sea monitoring area and generates the corresponding target abnormal scene category data. By dividing the candidate air, space, land and sea monitoring area into multiple monitoring unit areas and collecting data, the multi-dimensional fusion feature data of each monitoring unit area is generated, and detailed and comprehensive monitoring of each monitoring unit area is achieved. More importantly, the attention coefficient is also introduced, which can characterize the participation factor of each monitoring unit area in the target abnormal scene category data of the overall candidate air, space, land and sea monitoring area. By weighted fusion of the multi-dimensional fusion feature data of each monitoring unit area and the corresponding attention coefficient, the target multi-dimensional fusion feature data of each monitoring unit area is further optimized, so that it more accurately reflects the characteristics of each monitoring unit area and its influence on the overall abnormal scene. Finally, based on the target multi-dimensional fusion feature data of all monitoring unit areas, a second abnormal feature pattern decision is made for the entire candidate air, space, land and sea monitoring area, generating target abnormal scene category data with higher accuracy, which can not only provide a basis for further abnormal processing, but also help to manage and make decisions on the monitoring area. Thus, accurate monitoring and abnormal identification of the candidate air, space, land and sea monitoring area are achieved, effectively improving monitoring efficiency and accuracy.

[0090] In other application scenarios, to meet the need for comprehensive monitoring of air, land, and sea, candidate monitoring areas are first carefully allocated to monitoring units. This process ensures that each monitoring unit area is effectively covered, providing a solid foundation for subsequent data collection and anomaly detection.

[0091] Data collection is the core of the entire system's operation. Within each monitoring unit, multi-dimensional data collection is carried out using satellites, drones, ground-based sensor networks, and specialized oceanographic probes. This data not only includes geographical and environmental information, but also specifically captures key data on the marine environment, specifically relevant to submarine cable operations.

[0092] During submarine cable operations, ocean detectors monitor the cable's status and the surrounding marine environment in real time. Data collection on the seven key meteorological elements—temperature, humidity, air pressure, wind direction, wind speed, precipitation, and visibility—is crucial for assessing the operating environment and predicting weather changes. Furthermore, to ensure the safety of operators, special attention is paid to vital signs such as heart rate, blood oxygen saturation, and body temperature. This data provides real-time insights into the operator's physical condition and provides timely warnings of potential health risks.

[0093] Furthermore, wave height is a crucial safety indicator during marine operations. Wave height sensors monitor wave height and frequency in real time, providing robust data support for operational safety. Furthermore, considering the potential for toxic and flammable gas risks during onboard operations, gas detectors are deployed for real-time monitoring, ensuring that any leaks are detected and the system can respond immediately to protect personnel safety.

[0094] The collected multi-dimensional fusion feature data will serve as the basis for the first abnormal feature pattern decision. By analyzing this data, abnormal feature patterns can be promptly identified and abnormal feature pattern decision data generated for each monitoring unit area. This step is crucial for preventing and responding to emergencies.

[0095] Next, an attention coefficient is determined for each monitoring unit area based on the abnormal feature pattern decision data. This coefficient represents the importance of that area to the target abnormal scene category data for the entire candidate air, land, and sea monitoring area. By weightedly fusing the multidimensional fused feature data with the attention coefficient, more accurate target multidimensional fused feature data can be obtained.

[0096] Finally, based on this multi-dimensional fusion of target feature data, a second anomaly feature pattern decision is made, accurately generating target anomaly scene category data for candidate air, space, land, and sea monitoring areas. This process not only improves the accuracy of anomaly detection but also provides strong data support for subsequent emergency response and decision-making.

[0097] In a possible implementation, step S110 may include:

[0098] Step S110 , determining a moving observation frame with target scale measurement parameters corresponding to the candidate air-space, land-sea monitoring area and a moving interval of the moving observation frame.

[0099] Step S120 , allocating monitoring units to the candidate air-space-ground-sea monitoring area by moving the mobile observation frame according to the moving interval, and generating a plurality of monitoring unit areas included in the candidate air-space-ground-sea monitoring area.

[0100] In the remote monitoring system, the remote monitoring system may use a mobile observation frame-based technology to allocate monitoring units to candidate air, space, land, and sea monitoring areas, and generate multiple monitoring unit areas included in the candidate air, space, land, and sea monitoring areas.

[0101] First, the remote monitoring system needs to determine a moving observation frame with a target scale measurement parameter corresponding to the candidate air, space, and sea monitoring area, as well as the movement interval of the moving observation frame. This moving observation frame can be a virtual rectangular frame, and its size and shape can be determined based on the size and shape of the candidate air, space, and sea monitoring area. The target scale measurement parameter can be a scalar value used to measure the size of the moving observation frame.

[0102] The remote monitoring system can then allocate monitoring units to the candidate air, space, and sea monitoring area by moving the mobile observation frame according to the movement interval, thereby generating multiple monitoring unit areas included in the candidate air, space, and sea monitoring area. Specifically, the remote monitoring system can move the mobile observation frame over the candidate air, space, and sea monitoring area according to the movement interval, thereby dividing the candidate air, space, and sea monitoring area into multiple monitoring unit areas. In each monitoring unit area, the remote monitoring system can collect and analyze data, thereby achieving comprehensive monitoring of the candidate air, space, and sea monitoring area.

[0103] For example, if the candidate air, land, and sea monitoring areas include a large forested area, the remote monitoring system can use a large mobile observation frame to cover the entire forested area. The mobile observation frame can then be moved a certain distance at the desired movement interval to divide the forested area into multiple smaller monitoring units, each containing, for example, an independent forested area. Within each forested area, the remote monitoring system can use sensors and monitoring equipment to collect data on multiple dimensions, including temperature, humidity, and light, thereby achieving comprehensive monitoring of the large forested area.

[0104] In a possible implementation, step S120 may include:

[0105] Step S121 : collecting data at a first feature collection depth for each monitoring unit area, and generating reference multi-dimensional fusion feature data at the first feature collection depth for each monitoring unit area.

[0106] In the remote monitoring system, a multi-level feature collection technology may be used to collect data from each monitoring unit area and generate multi-dimensional fusion feature data for each monitoring unit area.

[0107] First, data is collected at a first feature acquisition depth for each monitoring unit area to generate reference multidimensional fusion feature data for each monitoring unit area. This first feature acquisition depth can be a relatively large depth value, for example, data from all sensors and monitoring equipment in each monitoring unit area can be collected to obtain comprehensive, high-dimensional reference multidimensional fusion feature data.

[0108] Step S122 , fusing the reference multi-dimensional fusion feature data of the first feature acquisition depths corresponding to the plurality of monitoring unit areas, to generate a reference multi-dimensional fusion feature data array of the candidate air-space, land-sea monitoring area.

[0109] For example, the reference multi-dimensional fusion feature data array may be a high-dimensional data matrix, in which each row corresponds to the reference multi-dimensional fusion feature data of a monitoring unit area.

[0110] Step S123, performing data collection at a second feature collection depth on the reference multidimensional fusion feature data array to generate a multidimensional fusion feature data array of the candidate air, space, and sea monitoring area, wherein the multidimensional fusion feature data array includes the multidimensional fusion feature data of the second feature collection depth of each monitoring unit area, and the second feature collection depth is less than the first feature collection depth.

[0111] The second feature acquisition depth can be a smaller depth value, for example, some feature values ​​can be extracted from the reference multidimensional fused feature data array to obtain a multidimensional fused feature data array of lower dimension. The multidimensional fused feature data array can include the multidimensional fused feature data of the second feature acquisition depth for each monitoring unit area.

[0112] For example, if the candidate air, space, land and sea monitoring areas include a large forest area, the data of all sensors and monitoring equipment in each monitoring unit area can be collected first to obtain a comprehensive, high-dimensional reference multi-dimensional fusion feature data. Then, these reference multi-dimensional fusion feature data are fused to generate a high-dimensional data matrix. Finally, some eigenvalues ​​are extracted from the data matrix to obtain a lower-dimensional multi-dimensional fusion feature data array. This multi-dimensional fusion feature data array may include the multi-dimensional fusion feature data of the second feature acquisition depth of each monitoring unit area, such as temperature, humidity, light and other feature values. In this way, this multi-dimensional fusion feature data array can be used to more accurately monitor and analyze the large forest area.

[0113] In a possible implementation, step S140 may include:

[0114] Step S141 : obtaining the number of abnormal feature patterns determined by the first abnormal feature pattern, and determining an equal probability distribution of the number of abnormal feature patterns.

[0115] Step S142: determining the deviation between the abnormal feature pattern decision data of the monitoring unit area and the equal probability distribution, and using the deviation as the attention coefficient of the monitoring unit area.

[0116] In the remote monitoring system, a method based on the attention mechanism can be used to determine the attention coefficient of each monitoring unit area, so that more attention can be paid to important monitoring unit areas when generating the target multi-dimensional fusion feature data of candidate air, space, land and sea monitoring areas.

[0117] Specifically, the number of abnormal characteristic patterns determined by the first abnormal characteristic pattern decision may be first obtained, and an equal probability distribution of the number of abnormal characteristic patterns may be determined. This equal probability distribution may be a probability density function representing the probability that each monitoring unit area has the same number of abnormal characteristic patterns.

[0118] Then, the degree of deviation between the abnormal feature pattern decision data of the monitoring unit area and the equal probability distribution can be determined, and the degree of deviation can be used as the attention coefficient of the monitoring unit area. This degree of deviation can be a numerical value used to represent the difference between the abnormal feature pattern decision data of the monitoring unit area and the equal probability distribution. The degree of deviation can be used as the attention coefficient of the monitoring unit area, so that when generating the target multi-dimensional fusion feature data of the candidate air, space, and sea monitoring areas, more attention can be paid to the monitoring unit areas with higher attention coefficients.

[0119] For example, if the candidate air, space, land, and sea monitoring area is a large forest area, each monitoring unit area may contain multiple sensors and monitoring equipment. By calculating the attention coefficient of each monitoring unit area, it is possible to determine which monitoring unit areas are more important, thereby generating more accurate target multi-dimensional fusion feature data. Specifically, the number of abnormal feature patterns of each monitoring unit area can be obtained first, and the equal probability distribution of the number of abnormal feature patterns can be determined. Then, the deviation between the abnormal feature pattern decision data of each monitoring unit area and the equal probability distribution can be determined, and the deviation can be used as the attention coefficient of the monitoring unit area. Finally, the attention coefficient can be used as a weight to perform weighted fusion on the multi-dimensional fusion feature data of multiple monitoring unit areas to generate the target multi-dimensional fusion feature data of the candidate air, space, and sea monitoring area. In this way, anomaly detection and identification can be performed based on the target multi-dimensional fusion feature data, thereby achieving comprehensive monitoring of large forest areas.

[0120] In a possible implementation, step S150 may include:

[0121] Step S151 : integrating the target multi-dimensional fusion feature data of the plurality of monitoring unit areas to generate integrated multi-dimensional fusion feature data.

[0122] Step S152: performing a second abnormal feature pattern decision on the candidate air, space, land and sea monitoring area based on the integrated multi-dimensional fusion feature data, and generating target abnormal scene category data for the candidate air, space, land and sea monitoring area.

[0123] In a remote monitoring system, the target multi-dimensional fusion feature data of multiple monitoring unit areas can be integrated to generate integrated multi-dimensional fusion feature data, and a second abnormal feature pattern decision can be made based on the integrated multi-dimensional fusion feature data to generate target abnormal scene category data for the candidate air, space, land and sea monitoring areas.

[0124] Specifically, the target multidimensional fused feature data from multiple monitoring unit areas can be integrated to generate integrated multidimensional fused feature data. This integration process can utilize various ensemble learning methods, such as bagging, boosting, and stacking. By integrating the target multidimensional fused feature data from multiple monitoring unit areas, a more stable and reliable feature representation can be obtained, thereby improving the accuracy and robustness of subsequent anomaly detection.

[0125] Then, based on the integrated multi-dimensional fusion feature data, a second abnormal feature pattern decision can be made for the candidate air, space, land, and sea monitoring area to generate target abnormal scene classification data for the candidate air, space, land, and sea monitoring area. This decision-making process can employ various anomaly detection methods, such as statistical, distance, density, and deep learning-based methods. By performing anomaly detection based on integrated multi-dimensional fusion feature data, more accurate and robust abnormal scene classification data can be obtained.

[0126] For example, if the candidate air, land, and sea monitoring area is a large forested area, the target multidimensional fusion feature data from multiple monitoring unit areas can be integrated to generate integrated multidimensional fusion feature data. Based on this integrated multidimensional fusion feature data, anomaly detection can then be performed on the large forested area to obtain target abnormal scene classification data, such as fire, pests and diseases, and water shortages. This allows for further analysis and decision-making based on this target abnormal scene classification data, thereby achieving comprehensive monitoring and management of large forested areas.

[0127] In a possible implementation, the first abnormal feature pattern decision is implemented by a single prediction subnetwork of an abnormal feature pattern decision network, and the second abnormal feature pattern decision is implemented based on a group prediction subnetwork of the abnormal feature pattern decision network.

[0128] The method further comprises:

[0129] Step S101: Obtain a sample multidimensional fusion feature data sequence for parameter learning of the abnormal feature pattern decision network, wherein the sample multidimensional fusion feature data sequence includes a plurality of sample multidimensional fusion feature data, and the sample multidimensional fusion feature data carries abnormal scene category annotation data. The sample multidimensional fusion feature data includes a plurality of sample monitoring unit areas, and the sample monitoring unit areas carry abnormal feature pattern annotation data.

[0130] In this step, some sample multidimensional fusion feature data sequences can be extracted from the historical monitoring data as training samples for parameter learning of the abnormal feature pattern decision network. Each sample multidimensional fusion feature data sequence can include multiple sample monitoring unit areas, each of which carries abnormal feature pattern annotation data indicating the abnormal feature pattern present in the sample monitoring unit area.

[0131] Step S102: Obtain the sample monitoring unit features of each of the sample monitoring unit areas in the multi-dimensional fusion feature data of each of the sample areas, and perform a first abnormal feature pattern decision on each of the sample monitoring unit areas based on the sample monitoring unit features of each of the sample monitoring unit areas according to the monomer prediction subnetwork to generate monomer prediction data for each of the sample monitoring unit areas.

[0132] In this step, the sample monitoring unit features of each sample monitoring unit area can be extracted and input into the abnormal feature pattern decision network as input data. Then, the individual prediction subnetwork can be used to perform a first abnormal feature pattern decision on each sample monitoring unit area, generating individual prediction data for the sample monitoring unit area. This individual prediction data may include the prediction result of whether an abnormal feature pattern exists in the area, as well as the probability distribution of various abnormal feature patterns.

[0133] Step S103: For each of the sample multidimensional fusion feature data, based on the abnormal feature pattern annotation data of each of the sample monitoring unit areas in the sample multidimensional fusion feature data, the group prediction subnetwork performs a second abnormal feature pattern decision on the sample multidimensional fusion feature data to generate group prediction data of the sample multidimensional fusion feature data.

[0134] In this step, the group prediction subnetwork can be used to perform a second abnormal feature pattern decision on each sample multidimensional fusion feature data to generate group prediction data for the sample multidimensional fusion feature data. This group prediction data may include the prediction result of the global abnormal scene category in the sample multidimensional fusion feature data, as well as the probability distribution of various abnormal scene categories.

[0135] Step S104, determining the network cost parameter value of the abnormal feature pattern decision network based on the deviation between the individual prediction data and the abnormal feature pattern annotation data of each sample monitoring unit area, and the deviation between the group prediction data and the abnormal scene category annotation data of the sample multi-dimensional fusion feature data.

[0136] Exemplarily, when determining the network cost parameter value of the abnormal feature pattern decision network, it can be achieved by calculating the following formula:

[0137] Cost(W)=Σ|Prediction(i)-Label(i)|+Σ|GroupPrediction(j)-SceneLabel(j)|

[0138] Among them, Cost(W) represents the network cost parameter value, Prediction(i) represents the individual prediction data of the i-th sample monitoring unit area, Label(i) represents the abnormal feature pattern annotation data of the i-th sample monitoring unit area, GroupPrediction(j) represents the group prediction data of the j-th sample multidimensional fusion feature data, and SceneLabel(j) represents the abnormal scene category annotation data of the j-th sample multidimensional fusion feature data.

[0139] Specifically, the network cost parameter value may include the following two parts:

[0140] The first part is the sum of the deviations between the individual prediction data of each sample monitoring unit area and the corresponding abnormal feature pattern annotation data.

[0141] The second part is the sum of the deviations between the group prediction data of the multi-dimensional fusion feature data of each sample and the corresponding abnormal scene category labeling data.

[0142] The deviation between the single prediction data and the abnormal feature pattern annotation data can be obtained by calculating the Euclidean distance. The specific formula is as follows:

[0143] Dis(Prediction(i), Label(i))=sqrt(Σ(Prediction(i, k)-Label(i, k))^2)

[0144] Among them, Dis(Prediction(i), Label(i)) represents the deviation between the individual prediction data of the i-th sample monitoring unit area and the corresponding abnormal feature pattern annotation data, Prediction(i, k) represents the k-th eigenvalue in the individual prediction data of the i-th sample monitoring unit area, Label(i, k) represents the k-th eigenvalue in the abnormal feature pattern annotation data of the i-th sample monitoring unit area, and k represents the feature index.

[0145] The deviation between the group prediction data and the abnormal scene category labeling data of the sample multi-dimensional fusion feature data can also be obtained by calculating the Euclidean distance. The specific formula is as follows:

[0146] Dis(GroupPrediction(j), SceneLabel(j))=sqrt(Σ(GroupPrediction(j, m)-SceneLabel(j, m))^2)

[0147] Among them, Dis(GroupPrediction(j), SceneLabel(j)) represents the deviation between the group prediction data of the j-th sample multidimensional fusion feature data and the corresponding abnormal scene category labeling data, GroupPrediction(j, m) represents the m-th eigenvalue in the group prediction data of the j-th sample multidimensional fusion feature data, SceneLabel(j, m) represents the m-th eigenvalue in the abnormal scene category labeling data of the j-th sample multidimensional fusion feature data, and m represents the feature index.

[0148] By calculating the deviation between the individual prediction data and the abnormal feature pattern annotation data and the deviation between the group prediction data of the sample multi-dimensional fusion feature data and the abnormal scene category annotation data, the accuracy of the abnormal feature pattern decision network in predicting abnormal feature patterns and abnormal scene categories can be evaluated, which helps to optimize network parameters and improve the prediction ability of the model.

[0149] By calculating the network cost parameter value, the accuracy of the abnormal feature pattern decision network in predicting abnormal feature patterns and abnormal scene categories can be evaluated, which helps optimize network parameters and improve the prediction ability of the model.

[0150] Step S105 : optimizing the network parameters of the abnormal feature pattern decision network according to the network cost parameter value, so as to perform knowledge learning on the abnormal feature pattern decision network.

[0151] In this step, the network parameters of the abnormal feature pattern decision network can be updated through an optimization algorithm based on the network cost parameter value. These network parameters may include various weights and biases in the abnormal feature pattern decision network. By updating the network parameters, the abnormal feature pattern decision network can be trained, thereby improving the accuracy of the abnormal feature pattern decision network.

[0152] In a possible implementation, step S104 may include:

[0153] Step S1041 : For each of the sample monitoring unit regions, determining a candidate monomer training cost of the abnormal feature pattern decision network according to a degree of deviation between the monomer prediction data and the abnormal feature pattern annotation data of the sample monitoring unit region.

[0154] Specifically, for each sample monitoring unit area, the deviation between the monomer prediction data and the abnormal feature pattern annotation data of the area can be calculated as described above, and used as the candidate monomer training cost of the area.

[0155] Step S1042 , summarizing the candidate monomer training costs of the abnormal feature pattern decision network to generate the monomer training cost of the abnormal feature pattern decision network.

[0156] Next, the candidate monomer training costs of all sample monitoring unit areas can be aggregated to generate the monomer training cost of the abnormal feature pattern decision network.

[0157] Step S1043 , determining the group training cost of the abnormal feature pattern decision network according to the deviation between the group prediction data of the sample multi-dimensional fusion feature data and the abnormal scene category labeling data.

[0158] Similarly, the deviation of the group prediction data and the abnormal scene category labeling data of the sample multi-dimensional fusion feature data and the abnormal scene category labeling data can be calculated by referring to the above description part, and used as the group training cost of the abnormal feature pattern decision network.

[0159] Step S1044: Based on the first training error influence factor of the individual training cost and the second training error influence factor of the group training cost, the individual training cost and the group training cost are fused and calculated to generate a network cost parameter value of the abnormal feature pattern decision network.

[0160] Specifically, the individual training cost and the group training cost can be weighted and fused to generate the network cost parameter value. The weighted fusion operation can be calculated using the following formula:

[0161] NetworkCost=α*MonolithicCost+(1-α)*GroupCost

[0162] Wherein, NetworkCost represents the network cost parameter value, MonolithicCost represents the individual training cost, GroupCost represents the group training cost, and α represents the first training error influencing factor, which ranges from 0 to 1. By adjusting the value of α, a trade-off can be made between considering the individual prediction error and the group prediction error, thereby generating a more accurate network cost parameter value.

[0163] In a possible implementation, step S103 may include:

[0164] Step S1031: For each of the sample monitoring unit areas in the sample multidimensional fusion feature data, obtain the number of abnormal feature patterns decided by the first abnormal feature pattern, determine the equal probability distribution of the number of abnormal feature patterns, determine the deviation between the abnormal feature pattern annotation data of the sample monitoring unit area and the equal probability distribution, use the deviation as the unit attention coefficient of the sample monitoring unit area, and determine the fused sample monitoring unit characteristics of each of the sample monitoring unit areas based on the sample monitoring unit characteristics and unit attention coefficient of each of the sample monitoring unit areas.

[0165] Specifically, for each sample monitoring unit region, the number of abnormal feature patterns in that region after the first abnormal feature pattern decision is obtained and converted into an equiprobable distribution. Next, the deviation between the annotated data of the abnormal feature pattern in that region and the equiprobable distribution is calculated, and this deviation is used as the unit attention coefficient for that region. Finally, the sample monitoring unit feature and the unit attention coefficient are weightedly fused to generate a fused sample monitoring unit feature for that region.

[0166] Step S1032: Based on the fused sample monitoring unit features of multiple sample monitoring unit areas in the sample multidimensional fusion feature data, a second abnormal feature pattern decision is performed on the sample multidimensional fusion feature data according to the group prediction subnetwork to generate group prediction data for the sample multidimensional fusion feature data.

[0167] Next, the fused sample monitoring unit features of multiple sample monitoring unit areas in the sample multi-dimensional fused feature data are integrated and input as input features into the group prediction subnetwork to perform a second abnormal feature pattern decision, thereby generating group prediction data for the sample multi-dimensional fused feature data. Specifically, the group prediction subnetwork can be implemented using a deep learning model such as a convolutional neural network, which can automatically learn the feature mapping relationship and contextual relationship in the input features and generate high-precision group prediction data.

[0168] In a possible implementation, before step S103, the method further includes:

[0169] Step A110, based on the monomer prediction data of each sample monitoring unit area in the sample multidimensional fusion feature data sequence, determine the key monitoring unit area of ​​each abnormal feature pattern generated by the first abnormal feature pattern decision from multiple sample monitoring unit areas in the sample multidimensional fusion feature data sequence.

[0170] Specifically, first, based on the individual prediction data of each sample monitoring unit area in each sample multi-dimensional fusion feature data sequence, the occurrence probability of the abnormal feature pattern of each sample monitoring unit area is determined. Then, from the multiple sample monitoring unit areas in the sample multi-dimensional fusion feature data sequence, the sample monitoring unit area with the highest occurrence probability is selected as the key monitoring unit area of ​​each abnormal feature pattern generated by the first abnormal feature pattern decision.

[0171] Step A120: For each of the abnormal feature patterns, obtain the template knowledge representation vector of the abnormal feature pattern and the corresponding template abnormal feature pattern, and momentum optimize the template knowledge representation vector based on the key monitoring unit area of ​​the abnormal feature pattern to generate an optimized template knowledge representation vector of the abnormal feature pattern.

[0172] Next, for each abnormal feature pattern, a template knowledge representation vector for the abnormal feature pattern and the corresponding template abnormal feature pattern are obtained. Then, based on the key monitoring unit area of ​​the abnormal feature pattern, momentum optimization is performed on the template knowledge representation vector of the abnormal feature pattern to generate an optimized template knowledge representation vector for the abnormal feature pattern. Specifically, an optimization algorithm such as gradient descent can be used to perform momentum optimization on the template knowledge representation vector so that the optimized template knowledge representation vector of the abnormal feature pattern can better represent the key features of the abnormal feature pattern.

[0173] Step A130: For each of the sample monitoring unit areas, determine the degree of matching between the multi-dimensional fusion feature data of the sample monitoring unit area and the optimized template knowledge representation vector of each of the abnormal feature patterns, and use the template abnormal feature pattern corresponding to the target optimized template knowledge representation vector with the largest matching degree as the unit template abnormal feature pattern of the sample monitoring unit area.

[0174] Then, for each sample monitoring unit area, the degree of match between the multidimensional fused feature data of the sample monitoring unit area and the optimized template knowledge representation vector of each abnormal feature pattern is calculated, and the template abnormal feature pattern corresponding to the target optimized template knowledge representation vector with the greatest matching degree is used as the unit template abnormal feature pattern of the sample monitoring unit area. Specifically, the degree of match can be determined by calculating the cosine similarity between the multidimensional fused feature data and the optimized template knowledge representation vector of each abnormal feature pattern.

[0175] Step A140 : For each of the sample monitoring unit areas, momentum optimize the abnormal feature pattern labeling data of the sample monitoring unit area according to the abnormal feature pattern of the unit template of the sample monitoring unit area to generate optimized abnormal feature pattern labeling data of the sample monitoring unit area.

[0176] Finally, for each sample monitoring unit area, momentum optimization is performed on the abnormal feature pattern labeling data of the sample monitoring unit area based on the abnormal feature pattern of the unit template of the sample monitoring unit area to generate optimized abnormal feature pattern labeling data for the sample monitoring unit area. Specifically, an optimization algorithm such as gradient descent can be used to perform momentum optimization on the abnormal feature pattern of the unit template so that the optimized abnormal feature pattern labeling data can better represent the true abnormal feature pattern of the sample monitoring unit area.

[0177] Based on the above description, step S103 includes: according to the optimized abnormal feature pattern labeling data of each sample monitoring unit area in the sample multidimensional fusion feature data, according to the group prediction subnetwork, a second abnormal feature pattern decision is made on the sample multidimensional fusion feature data, and group prediction data of the sample multidimensional fusion feature data is generated.

[0178] Specifically, the optimized abnormal feature pattern annotation data for each sample monitoring unit area in the sample multi-dimensional fusion feature data is input into the group prediction subnetwork, and a second abnormal feature pattern decision is performed to generate group prediction data for the sample multi-dimensional fusion feature data. The group prediction subnetwork can be implemented using a deep learning model such as a convolutional neural network, which can automatically learn the feature mapping relationship and contextual relationship in the input features and generate high-precision group prediction data.

[0179] The step S104 includes: determining the network cost parameter value of the abnormal feature pattern decision network based on the deviation between the individual prediction data of each sample monitoring unit area and the optimized abnormal feature pattern annotation data, and the deviation between the group prediction data of the sample multi-dimensional fusion feature data and the abnormal scene category annotation data.

[0180] During training, the network cost parameter value of the abnormal feature pattern decision network is determined based on the deviation between the individual prediction data and the optimized abnormal feature pattern annotation data for each sample monitoring unit area, as well as the deviation between the group prediction data and the abnormal scene category annotation data for the sample multi-dimensional fusion feature data. Specifically, the deviation can be calculated using a loss function such as mean squared error and used as the network cost parameter value to optimize the network parameters of the abnormal feature pattern decision network to improve the model's prediction accuracy and generalization ability.

[0181] For example, in one possible implementation, the monomer prediction data includes a first confidence level of the sample monitoring unit area matching each of the abnormal feature patterns.

[0182] The step A110 may include:

[0183] Step A111: For each of the abnormal feature patterns, determine the top N target sample monitoring unit areas that match the abnormal feature pattern and are arranged in descending order with a first confidence level from multiple sample monitoring unit areas of the sample multi-dimensional fusion feature data sequence, where N is an integer greater than 1.

[0184] Step A112: Using the first N target sample monitoring unit areas as key monitoring unit areas of the abnormal feature pattern.

[0185] Specifically, for each abnormal feature pattern, the top N target sample monitoring unit areas, ranked from largest to smallest with a first confidence level, that match the abnormal feature pattern are determined from the multiple sample monitoring unit areas in the sample multi-dimensional fusion feature data sequence. The first confidence level is the confidence level that the sample monitoring unit areas match each of the abnormal feature patterns, which can be predicted by the model. N is an integer greater than 1 and is used to determine the number of key monitoring unit areas.

[0186] Then, the first N target sample monitoring unit areas are used as key monitoring unit areas of the abnormal feature pattern. That is, the key monitoring unit areas are sample monitoring unit areas that match the abnormal feature pattern and have a high first confidence level, and can serve as an important basis for decision-making on the abnormal feature pattern.

[0187] The step A120 may include:

[0188] Step A121 : Momentum optimize the template knowledge representation vector based on the first key monitoring unit area of ​​the abnormal feature pattern to generate a first intermediate template knowledge representation vector of the abnormal feature pattern.

[0189] Specifically, the momentum optimization is a momentum-based optimization algorithm that can adjust the template knowledge representation vector according to the feature information of the key monitoring unit area.

[0190] For example, first, based on the first key monitoring unit area of ​​the abnormal feature pattern, feature information of the key monitoring unit area is extracted. Then, based on the template knowledge representation vector and the feature information of the key monitoring unit area, an optimization gradient is calculated. Next, based on the optimization gradient and the template knowledge representation vector, the template knowledge representation vector is updated to generate the first intermediate template knowledge representation vector of the abnormal feature pattern.

[0191] It is important to note that the momentum optimization algorithm can adjust the template knowledge representation vector based on the characteristic information of the key monitoring unit area to better adapt it to the characteristic information of the abnormal characteristic pattern. Furthermore, the momentum optimization algorithm can also prevent oscillation and fluctuation of the template knowledge representation vector during the optimization process, thereby improving the stability and accuracy of the template knowledge representation vector.

[0192] Step A122: Based on the kth key monitoring unit area of ​​the abnormal feature pattern, momentum optimize the (k-1)th intermediate template knowledge representation vector of the abnormal feature pattern to generate the kth intermediate template knowledge representation vector of the abnormal feature pattern, where k is greater than 0 and not greater than N.

[0193] Based on the kth key monitoring unit area of ​​the abnormal feature pattern, momentum optimization is performed on the (k-1)th intermediate template knowledge representation vector of the abnormal feature pattern to generate the kth intermediate template knowledge representation vector of the abnormal feature pattern. Wherein, k is greater than 0 and not greater than N, indicating that momentum optimization needs to be performed on the first N key monitoring unit areas of the abnormal feature pattern.

[0194] Exemplarily, first, based on the kth key monitoring unit area of ​​the abnormal feature pattern, feature information of the key monitoring unit area is extracted. Then, based on the (k-1)th intermediate template knowledge representation vector of the abnormal feature pattern and the feature information of the key monitoring unit area, an optimization gradient is calculated. Next, based on the optimization gradient and the (k-1)th intermediate template knowledge representation vector of the abnormal feature pattern, the (k-1)th intermediate template knowledge representation vector of the abnormal feature pattern is updated to generate the kth intermediate template knowledge representation vector of the abnormal feature pattern.

[0195] Assuming that the kth key monitoring unit area of ​​the abnormal feature pattern is A and the (k-1)th intermediate template knowledge representation vector is v(k-1), the feature information f(A) is first extracted based on area A. Then, the server calculates an optimization gradient g(k) based on the (k-1)th intermediate template knowledge representation vector v(k-1) of the abnormal feature pattern and the feature information f(A) of the key monitoring unit area A. Next, the server updates the (k-1)th intermediate template knowledge representation vector v(k-1) of the abnormal feature pattern based on the optimization gradient g(k) and the (k-1)th intermediate template knowledge representation vector v(k-1) of the abnormal feature pattern, and generates the kth intermediate template knowledge representation vector v(k) of the abnormal feature pattern. It should be noted that the momentum optimization algorithm can adjust the template knowledge representation vector based on the feature information of the key monitoring unit area, so as to better adapt to the feature information of the abnormal feature pattern. At the same time, the momentum optimization algorithm can also avoid oscillations and fluctuations in the template knowledge representation vector during the optimization process, thereby improving the stability and accuracy of the template knowledge representation vector.

[0196] Step A123 , polling the k, generating the Nth intermediate template knowledge representation vector of the abnormal feature pattern, and using the Nth intermediate template knowledge representation vector of the abnormal feature pattern as the optimized template knowledge representation vector of the abnormal feature pattern.

[0197] For example, the k is polled to generate the Nth intermediate template knowledge representation vector of the abnormal feature pattern. Finally, the Nth intermediate template knowledge representation vector of the abnormal feature pattern is used as the optimized template knowledge representation vector of the abnormal feature pattern for subsequent abnormality detection tasks.

[0198] Alternatively, in step A124, key multidimensional fusion feature data of each key monitoring unit area of ​​the abnormal feature pattern is obtained, and average multidimensional fusion feature data of multiple key multidimensional fusion feature data is determined.

[0199] For example, suppose there are five key monitoring units, and their multi-dimensional fusion feature data are F1, F2, F3, F4, and F5 respectively. Then, the average multi-dimensional fusion feature data F_avg can be calculated by the following formula:

[0200] F_avg=(F1+F2+F3+F4+F5) / 5

[0201] Step A125: Obtain a first importance coefficient of the average multi-dimensional fusion feature data and a second importance coefficient of the template knowledge representation vector.

[0202] In this step, two importance coefficients are evaluated: one for the average multidimensional fusion feature data (denoted as w1) and the other for the template knowledge representation vector (denoted as w2). These two coefficients can be assigned according to their respective importance, for example, through some optimization algorithm or based on prior knowledge.

[0203] In other words, the first importance coefficient (denoted as w1) reflects the weight of the average multidimensional fusion feature data in the optimized template knowledge representation vector. This parameter may be determined based on multiple factors, such as the consistency of key monitoring unit areas and the credibility of the feature data. The calculation formula may vary depending on the actual situation. One possible approach is to iteratively optimize using an optimization algorithm (such as gradient descent or genetic algorithm) until a certain set criterion is met or a preset number of iterations is reached.

[0204] The second importance coefficient (denoted as w2) reflects the weight of the template knowledge representation vector in the optimization of the template knowledge representation vector. This coefficient may be determined based on prior knowledge or model training results. Similarly, its calculation formula may vary depending on the specific application scenario. One possible approach is to set an initial value and then optimize it using a method similar to the first importance coefficient.

[0205] It should be noted that in order to ensure the balance between w1 and w2, it is often necessary to normalize the two, that is, to make the sum of w1 and w2 equal to 1. This can be done using the following formula:

[0206] w1=w1 / (w1+w2)

[0207] w2=1-w1

[0208] In this way, no matter what the original values ​​of w1 and w2 are, after normalization, the sum of the two is always equal to 1, thereby ensuring that the sum of the weights of the average multidimensional fusion feature data and the template knowledge representation vector is constant when generating the optimized template knowledge representation vector.

[0209] Step A126: performing a fusion calculation on the average multi-dimensional fusion feature data and the template knowledge representation vector based on the first importance coefficient and the second importance coefficient to generate an optimized template knowledge representation vector of the abnormal feature pattern.

[0210] In this step, the first importance coefficient and the second importance coefficient are used to perform weighted fusion on the average multidimensional fusion feature data and the template knowledge representation vector. Let the original template knowledge representation vector be V, then the optimized template knowledge representation vector V_opt can be calculated by the following formula:

[0211] V_opt=w1*F_avg+w2*V

[0212] In this way, an optimized template knowledge representation vector is obtained, which reflects the characteristic information of the key monitoring unit area to a greater extent, thereby improving the recognition accuracy of abnormal feature patterns.

[0213] In this embodiment, key multidimensional fusion feature data is first extracted from the key monitoring unit areas of the abnormal feature pattern. The average value of these key multidimensional fusion feature data is then calculated to obtain average multidimensional fusion feature data. A first importance coefficient of the average multidimensional fusion feature data and a second importance coefficient of the template knowledge representation vector are then calculated. These coefficients can be obtained by calculating the weight of each feature data.

[0214] Finally, based on the first and second importance coefficients, a weighted summation of the average multidimensional fusion feature data and the template knowledge representation vector is performed to obtain the optimized template knowledge representation vector, which can be used for subsequent anomaly detection and analysis.

[0215] Step A140 may include:

[0216] Step A141: Obtain a first mode attention coefficient of the abnormal feature pattern of the unit template and a second mode attention coefficient of the abnormal feature pattern annotation data.

[0217] Step A142: Based on the first pattern attention coefficient and the second pattern attention coefficient, a fusion calculation is performed on the unit template abnormal feature pattern and the abnormal feature pattern annotation data to generate optimized abnormal feature pattern annotation data of the sample monitoring unit area.

[0218] In this process, we first obtain the first-mode attention coefficient of the unit template abnormal feature pattern (denoted as a1) and the second-mode attention coefficient of the abnormal feature pattern annotation data (denoted as a2). These two attention coefficients represent the weights of the unit template abnormal feature pattern and the abnormal feature pattern annotation data when generating the optimized abnormal feature pattern annotation data.

[0219] First Pattern Attention Coefficient: This parameter reflects the weight of the abnormal feature pattern of a unit template in optimizing the anomaly feature pattern annotation data. The calculation formula may vary depending on the actual situation. One possible approach is to iterate the optimization algorithm until a certain set criteria is met or a preset number of iterations is reached.

[0220] Second-mode attention coefficient: This parameter reflects the weight of anomaly pattern annotation data in optimizing anomaly pattern annotation data. This coefficient may be determined based on prior knowledge or model training results. Similarly, its calculation formula may vary depending on the specific application scenario.

[0221] After obtaining these two attention coefficients, the unit template abnormal feature pattern (denoted as P1) and the abnormal feature pattern annotation data (denoted as P2) can be weighted fused using the following formula to generate the optimized abnormal feature pattern annotation data (denoted as P_opt):

[0222] P_opt=a1*P1+a2*P2

[0223] In this process, the values ​​of a1 and a2 may be dynamically adjusted through some optimization algorithm to better adapt to the actual situation of the sample monitoring unit area, thereby generating more accurate optimized abnormal feature pattern annotation data.

[0224] Figure 2 The hardware structure of the remote monitoring system 100 provided in the embodiment of the present application for implementing the above-mentioned monitoring data management method based on the integrated fusion of air, land, and sea is shown. Figure 2 As shown, remote monitoring system 100 may include a processor 110 , a machine-readable storage medium 120 , a bus 130 , and a communication unit 140 .

[0225] In an alternative embodiment, the remote monitoring system 100 may be a single server or a server group. The server group may be centralized or distributed (for example, the remote monitoring system 100 may be a distributed system). In an alternative embodiment, the remote monitoring system 100 may be local or remote. For example, the remote monitoring system 100 may access information and / or data stored in the machine-readable storage medium 120 via a network. For another example, the remote monitoring system 100 may be directly connected to the machine-readable storage medium 120 to access the stored information and / or data. In an alternative embodiment, the remote monitoring system 100 may be implemented on a cloud platform. By way of example only, the cloud platform may include a private cloud, a public cloud, a hybrid cloud, a community cloud, a distributed cloud, an internal cloud, a multi-layer cloud, or the like, or any combination thereof.

[0226] The machine-readable storage medium 120 can store data and / or instructions. In an alternative embodiment, the machine-readable storage medium 120 can store data obtained from an external terminal. In an alternative embodiment, the machine-readable storage medium 120 can store data and / or instructions used by the remote monitoring system 100 to execute or use to complete the exemplary methods described in this application. In an alternative embodiment, the machine-readable storage medium 120 may include a mass storage, a removable memory, a volatile read-write memory, a read-only memory, or the like, or any combination thereof. Exemplary mass storage may include a magnetic disk, an optical disk, a solid-state disk, or the like. Exemplary removable storage may include a flash drive, a floppy disk, an optical disk, a memory card, a compressed disk, a magnetic tape, or the like.

[0227] During the specific implementation process, multiple processors 110 execute computer executable instructions stored in the machine-readable storage medium 120, so that the processor 110 can execute the monitoring data management method based on the integrated integration of air, land, and sea as described in the above method embodiment. The processor 110, the machine-readable storage medium 120 and the communication unit 140 are connected through the bus 130, and the processor 110 can be used to control the sending and receiving actions of the communication unit 140.

[0228] The specific implementation process of the processor 110 can refer to the various method embodiments executed by the above-mentioned remote monitoring system 100. The implementation principles and technical effects are similar and will not be repeated here in this embodiment.

[0229] In addition, an embodiment of the present application also provides a readable storage medium, which has computer-executable instructions preset in the readable storage medium. When the processor executes the computer-executable instructions, the above-mentioned monitoring data management method based on the integrated fusion of air, land, and sea is implemented.

[0230] Similarly, it should be noted that, in order to simplify the description of the present disclosure and thus facilitate the understanding of one or more embodiments of the invention, in the foregoing description of the embodiments of the present disclosure, multiple features may sometimes be combined into one embodiment, figure, or description thereof. Similarly, it should be noted that, in order to simplify the description of the present disclosure and thus facilitate the understanding of one or more embodiments of the invention, in the foregoing description of the embodiments of the present disclosure, multiple features may sometimes be combined into one embodiment, figure, or description thereof.

Claims

1. A monitoring data management method based on the integration of air, land, and sea, characterized in that: The method comprises: Allocating monitoring units to the candidate air, space, and sea monitoring areas to generate a plurality of monitoring unit areas included in the candidate air, space, and sea monitoring areas; Collect data for each of the monitoring unit areas to generate multi-dimensional fusion feature data for each of the monitoring unit areas, wherein the multi-dimensional fusion feature data is fusion feature data of feature data collected separately by satellites, drones, ground sensor networks, and ocean detectors; Based on the multi-dimensional fusion feature data of each monitoring unit area, performing a first abnormal feature pattern decision on each monitoring unit area to generate abnormal feature pattern decision data for each monitoring unit area; For each of the monitoring unit areas, determining an attention coefficient of the monitoring unit area based on the abnormal feature pattern decision data of the monitoring unit area, and performing weighted fusion on the multidimensional fusion feature data of the monitoring unit area and the attention coefficient of the monitoring unit area to generate target multidimensional fusion feature data of the monitoring unit area, wherein the attention coefficient represents a participation factor of the monitoring unit area in the target abnormal scene category data of the candidate air, land, and sea monitoring areas; Performing a second abnormal feature pattern decision on the candidate air, space, and sea monitoring area based on the target multi-dimensional fusion feature data of the multiple monitoring unit areas, and generating target abnormal scene category data for the candidate air, space, and sea monitoring area; The first abnormal feature pattern decision is implemented by a single prediction sub-network of the abnormal feature pattern decision network, and the second abnormal feature pattern decision is implemented according to a group prediction sub-network of the abnormal feature pattern decision network; The method further comprises: Acquire a sample multidimensional fusion feature data sequence for parameter learning of the abnormal feature pattern decision network, the sample multidimensional fusion feature data sequence comprising a plurality of sample multidimensional fusion feature data, the sample multidimensional fusion feature data carrying abnormal scene category annotation data, the sample multidimensional fusion feature data comprising a plurality of sample monitoring unit areas, the sample monitoring unit areas carrying abnormal feature pattern annotation data; Obtaining the sample monitoring unit features of each of the sample monitoring unit areas in the sample multi-dimensional fusion feature data, performing a first abnormal feature pattern decision on each of the sample monitoring unit areas based on the sample monitoring unit features of each of the sample monitoring unit areas according to the individual prediction subnetwork, and generating individual prediction data for each of the sample monitoring unit areas; For each of the sample multidimensional fusion feature data, based on the abnormal feature pattern annotation data of each of the sample monitoring unit areas in the sample multidimensional fusion feature data, a second abnormal feature pattern decision is performed on the sample multidimensional fusion feature data through the group prediction subnetwork to generate group prediction data of the sample multidimensional fusion feature data; Determining a network cost parameter value of the abnormal feature pattern decision network based on the deviation between the individual prediction data and the abnormal feature pattern annotation data of each sample monitoring unit area, and the deviation between the group prediction data and the abnormal scene category annotation data of the sample multi-dimensional fusion feature data; Optimizing network parameters of the abnormal feature pattern decision network according to the network cost parameter value to perform knowledge learning on the abnormal feature pattern decision network; Before the step of performing a second abnormal feature pattern decision on the sample multidimensional fusion feature data by the group prediction subnetwork based on the abnormal feature pattern annotation data of each sample monitoring unit area in the sample multidimensional fusion feature data and generating group prediction data for the sample multidimensional fusion feature data, the method further comprises: Determining, from the plurality of sample monitoring unit areas in the sample multi-dimensional fusion feature data sequence, key monitoring unit areas of each abnormal feature pattern generated by the first abnormal feature pattern decision, based on the monomer prediction data of each sample monitoring unit area in the sample multi-dimensional fusion feature data sequence; For each of the abnormal feature patterns, obtaining a template knowledge representation vector of the abnormal feature pattern and a corresponding template abnormal feature pattern, and momentum optimizing the template knowledge representation vector based on a key monitoring unit area of ​​the abnormal feature pattern to generate an optimized template knowledge representation vector of the abnormal feature pattern; For each of the sample monitoring unit areas, determining the degree of match between the multidimensional fusion feature data of the sample monitoring unit area and the optimized template knowledge representation vector of each of the abnormal feature patterns, and taking the template abnormal feature pattern corresponding to the target optimized template knowledge representation vector with the largest degree of match as the unit template abnormal feature pattern of the sample monitoring unit area; For each of the sample monitoring unit areas, momentum optimizes the abnormal feature pattern labeling data of the sample monitoring unit area according to the unit template abnormal feature pattern of the sample monitoring unit area to generate optimized abnormal feature pattern labeling data of the sample monitoring unit area; The single prediction data includes a first confidence level of the sample monitoring unit area matching each of the abnormal feature patterns; The determining, based on the monomer prediction data of each sample monitoring unit area in the sample multi-dimensional fusion feature data sequence, from multiple sample monitoring unit areas of the sample multi-dimensional fusion feature data sequence, of key monitoring unit areas of each abnormal feature pattern generated by the first abnormal feature pattern decision includes: For each of the abnormal feature patterns, determining, from a plurality of sample monitoring unit areas of the sample multi-dimensional fusion feature data sequence, top N target sample monitoring unit areas that match the abnormal feature pattern and are arranged in descending order of first confidence, where N is an integer greater than 1; The first N target sample monitoring unit areas are used as key monitoring unit areas of the abnormal feature pattern; The step of momentum optimizing the template knowledge representation vector based on the key monitoring unit area of ​​the abnormal feature pattern to generate the optimized template knowledge representation vector of the abnormal feature pattern includes: Momentum-optimize the template knowledge representation vector according to the first key monitoring unit area of ​​the abnormal feature pattern to generate a first intermediate template knowledge representation vector of the abnormal feature pattern; According to the kth key monitoring unit area of ​​the abnormal feature pattern, momentum optimize the (k-1)th intermediate template knowledge representation vector of the abnormal feature pattern to generate the kth intermediate template knowledge representation vector of the abnormal feature pattern, where k is greater than 0 and not greater than N; Polling the k, generating an Nth intermediate template knowledge representation vector of the abnormal feature pattern, and using the Nth intermediate template knowledge representation vector of the abnormal feature pattern as the optimized template knowledge representation vector of the abnormal feature pattern; Alternatively, key multidimensional fusion feature data of each key monitoring unit area of ​​the abnormal feature pattern is obtained, and average multidimensional fusion feature data of a plurality of key multidimensional fusion feature data is determined; Obtaining a first importance coefficient of the average multidimensional fusion feature data and a second importance coefficient of the template knowledge representation vector; performing a fusion calculation on the average multi-dimensional fusion feature data and the template knowledge representation vector according to the first importance coefficient and the second importance coefficient to generate an optimized template knowledge representation vector of the abnormal feature pattern; The step of momentum optimizing the abnormal feature pattern labeling data of the sample monitoring unit area based on the abnormal feature pattern of the unit template of the sample monitoring unit area to generate the optimized abnormal feature pattern labeling data of the sample monitoring unit area includes: Obtaining a first mode attention coefficient of the abnormal feature pattern of the unit template and a second mode attention coefficient of the abnormal feature pattern annotation data; According to the first pattern attention coefficient and the second pattern attention coefficient, the unit template abnormal feature pattern and the abnormal feature pattern annotation data are fused and calculated to generate optimized abnormal feature pattern annotation data of the sample monitoring unit area.

2. The monitoring data management method based on the integrated fusion of air, land, and sea according to claim 1 is characterized in that: The step of allocating monitoring units to the candidate air, space, and sea monitoring areas to generate a plurality of monitoring unit areas included in the candidate air, space, and sea monitoring areas includes: Determine a moving observation frame having target scale measurement parameters corresponding to the candidate air, space, land and sea monitoring area, and a moving interval of the moving observation frame; By moving the mobile observation frame according to the moving interval, monitoring units are allocated to the candidate air-space-ground-sea monitoring area, and a plurality of monitoring unit areas included in the candidate air-space-ground-sea monitoring area are generated.

3. The monitoring data management method based on the integrated fusion of air, land, and sea according to claim 1 is characterized in that: The data collection for each monitoring unit area to generate multi-dimensional fusion feature data for each monitoring unit area includes: Performing data collection at a first feature collection depth for each monitoring unit area to generate reference multi-dimensional fusion feature data at the first feature collection depth for each monitoring unit area; Fusing the reference multi-dimensional fusion feature data of the first feature acquisition depths corresponding to the multiple monitoring unit areas to generate a reference multi-dimensional fusion feature data array for the candidate air, space, land and sea monitoring area; Data collection at a second feature collection depth is performed on the reference multidimensional fusion feature data array to generate a multidimensional fusion feature data array for the candidate air, space, and sea monitoring areas, wherein the multidimensional fusion feature data array includes the multidimensional fusion feature data of the second feature collection depth for each of the monitoring unit areas, and the second feature collection depth is less than the first feature collection depth.

4. The monitoring data management method based on the integrated fusion of air, land, and sea according to claim 1 is characterized in that: The determining of the attention coefficient of the monitoring unit area based on the abnormal feature pattern decision data of the monitoring unit area includes: Obtaining the number of abnormal feature patterns determined by the first abnormal feature pattern, and determining an equal probability distribution of the number of abnormal feature patterns; Determine the deviation between the abnormal feature pattern decision data of the monitoring unit area and the equal probability distribution, and use the deviation as the attention coefficient of the monitoring unit area.

5. The monitoring data management method based on the integrated fusion of air, land, and sea according to claim 1 is characterized in that: The step of performing a second abnormal feature pattern decision on the candidate air, space, and sea monitoring area based on the target multi-dimensional fusion feature data of the multiple monitoring unit areas to generate target abnormal scene category data for the candidate air, space, and sea monitoring area includes: Integrating the target multi-dimensional fusion feature data of the plurality of monitoring unit areas to generate integrated multi-dimensional fusion feature data; Based on the integrated multi-dimensional fusion feature data, a second abnormal feature pattern decision is performed on the candidate air-space-ground-sea monitoring area to generate target abnormal scene category data of the candidate air-space-ground-sea monitoring area.

6. The monitoring data management method based on the integrated fusion of air, land, and sea according to claim 1 is characterized in that: Determining the network cost parameter value of the abnormal feature pattern decision network based on the deviation between the individual prediction data and the abnormal feature pattern annotation data of each sample monitoring unit area, and the deviation between the group prediction data of the sample multi-dimensional fusion feature data and the abnormal scene category annotation data, includes: For each of the sample monitoring unit areas, determining a candidate monomer training cost of the abnormal feature pattern decision network according to a degree of deviation between the monomer prediction data and the abnormal feature pattern labeled data of the sample monitoring unit area; Aggregating the training costs of the candidate monomers of the abnormal feature pattern decision network to generate the monomer training cost of the abnormal feature pattern decision network; Determining a group training cost of the abnormal feature pattern decision network based on the group prediction data of the sample multi-dimensional fusion feature data and the deviation of the abnormal scene category labeling data; According to the first training error influence factor of the individual training cost and the second training error influence factor of the group training cost, the individual training cost and the group training cost are fused and calculated to generate a network cost parameter value of the abnormal feature pattern decision network.

7. The monitoring data management method based on the integrated fusion of air, land, and sea according to claim 1 is characterized in that: The abnormal feature pattern labeling data of each sample monitoring unit area in the sample multidimensional fusion feature data is based on the abnormal feature pattern labeling data, and the second abnormal feature pattern decision is performed on the sample multidimensional fusion feature data by the group prediction subnetwork to generate group prediction data of the sample multidimensional fusion feature data, including: For each of the sample monitoring unit areas in the sample multidimensional fusion feature data, obtaining the number of abnormal feature patterns determined by the first abnormal feature pattern, determining an equal probability distribution of the number of abnormal feature patterns, determining a degree of deviation between the abnormal feature pattern labeled data of the sample monitoring unit area and the equal probability distribution, using the degree of deviation as a unit attention coefficient of the sample monitoring unit area, and determining a fused sample monitoring unit feature of each of the sample monitoring unit areas based on the sample monitoring unit features and the unit attention coefficient of each of the sample monitoring unit areas; Based on the fused sample monitoring unit features of multiple sample monitoring unit areas in the sample multidimensional fusion feature data, a second abnormal feature pattern decision is performed on the sample multidimensional fusion feature data according to the group prediction subnetwork to generate group prediction data of the sample multidimensional fusion feature data.

8. The monitoring data management method based on the integrated fusion of air, land, and sea according to claim 1 is characterized in that: The abnormal feature pattern labeling data of each sample monitoring unit area in the sample multidimensional fusion feature data is based on the abnormal feature pattern labeling data, and the second abnormal feature pattern decision is performed on the sample multidimensional fusion feature data by the group prediction subnetwork to generate group prediction data of the sample multidimensional fusion feature data, including: Based on the optimized abnormal feature pattern labeling data of each of the sample monitoring unit areas in the sample multidimensional fusion feature data, a second abnormal feature pattern decision is performed on the sample multidimensional fusion feature data according to the group prediction subnetwork to generate group prediction data of the sample multidimensional fusion feature data; Determining the network cost parameter value of the abnormal feature pattern decision network based on the deviation between the individual prediction data and the abnormal feature pattern annotation data of each sample monitoring unit area, and the deviation between the group prediction data of the sample multi-dimensional fusion feature data and the abnormal scene category annotation data, includes: The network cost parameter value of the abnormal feature pattern decision network is determined based on the deviation between the individual prediction data of each sample monitoring unit area and the optimized abnormal feature pattern annotation data, as well as the deviation between the group prediction data of the sample multi-dimensional fusion feature data and the abnormal scene category annotation data.

9. A remote monitoring system, characterized in that: The remote monitoring system includes a processor and a machine-readable storage medium, wherein the machine-readable storage medium stores machine-executable instructions, and the machine-executable instructions are loaded and executed by the processor to implement the monitoring data management method based on the integrated fusion of air, land, and sea according to any one of claims 1 to 8.

Images