A real-time regional risk monitoring method in weak communication environment

Abstract

The present invention discloses a real-time regional risk monitoring method in a weak communication environment, including classifying time-sensitive target devices in an environmental risk area picture to obtain image sample data, dividing the image sample data into a data training set and a data test set according to a preset ratio; constructing a knowledge distillation model based on the relationship between soft labels and modules; training the knowledge distillation model through the data training set to obtain an optimal knowledge distillation model, and running the optimal knowledge distillation model to an edge device to realize target identification of time-sensitive target devices in a preset environmental risk area; constructing a risk dynamic update mechanism based on Bayesian reasoning, which solves the problem that the existing method has a reliable target identification model that often has a huge number of parameters and cannot run in real time on an edge device with relatively insufficient computing power, so it is necessary to seek a model that is suitable for the computing power of the edge device and has reliable results, and the technical problem that the existing situation cannot be updated in combination with new detection information.

Description

Technical Field

[0001] The present invention relates to the field of unmanned reconnaissance technology, and in particular to a real-time regional risk monitoring method in a weak communication environment. Background Art

[0002] Unmanned reconnaissance equipment has gradually played an important role in modern regional risk monitoring due to its low cost and adaptability to harsh environments. At present, land area risk monitoring is characterized by an increase in dynamic and uncertain tasks. Tactical teams equipped with unmanned platforms will become the core force for carrying out local area tasks. It is particularly important to make full use of the advantages of unmanned platforms, do a good job in defending key areas, respond to threats from time-sensitive targets (time-sensitive targets TCT), and realize regional reconnaissance and situation generation. At present, the existing regional risk monitoring methods usually involve people operating unmanned equipment or unmanned equipment transmitting the collected raw data back to the ground or other platforms through communication links for processing. This mode requires a stable communication link and a large bandwidth to support data transmission, but communications in this area are usually unstable, and low bandwidth problems or even communication denial may occur due to electronic interference, equipment damage, and other reasons.

[0003] However, there are still two problems that need to be solved in the existing regional risk monitoring methods: on the one hand, reliable target recognition models often have a huge number of parameters and cannot run in real time on edge devices with relatively insufficient computing power, so it is necessary to find models that are suitable for the computing power of edge devices and have reliable results. On the other hand, there is a core technical problem that the existing situation cannot be updated in combination with new detection information. Summary of the invention

[0004] The present invention provides a real-time regional risk monitoring method in a weak communication environment to overcome the above technical problems.

[0005] In order to achieve the above object, the technical solution of the present invention is:

[0006] A real-time regional risk monitoring method in a weak communication environment specifically comprises the following steps:

[0007] S1: Based on each edge device set in a weak communication environment, autonomously scout the preset environmental risk area to obtain reconnaissance data, and the reconnaissance data at least includes a picture of the environmental risk area;

[0008] And mark the target category labels of time-sensitive target devices in the environmental risk area pictures to obtain image sample data;

[0009] Divide the image sample data into a data training set and a data test set according to a preset ratio;

[0010] S2: Construct a knowledge distillation model based on the relationship between soft labels and modules;

[0011] The knowledge distillation model includes an inter-module association knowledge network for real-time regional risk target feature extraction based on a symmetric teacher-student target recognition network and a label dark knowledge module for target category label recognition, and the inter-module association knowledge network includes a pre-trained teacher target recognition network and a student target recognition network;

[0012] S3: Train the knowledge distillation model through the data training set to obtain the optimal knowledge distillation model, and run the optimal knowledge distillation model to the edge device to achieve target identification of time-sensitive target devices in the preset environmental risk area;

[0013] The model training method specifically includes:

[0014] The teacher target recognition network obtains knowledge distillation intermediate feature maps and knowledge distillation labels of different scales used as knowledge distillation information based on the input data training set; the student target recognition network obtains target recognition intermediate feature maps and target category recognition labels of different scales based on the input data training set, and obtains the Euclidean distance between the knowledge distillation intermediate feature map and the target recognition intermediate feature map;

[0015] The label dark knowledge module obtains the target category recognition probability based on the knowledge distillation label and the target category recognition label, and outputs the final target recognition result;

[0016] According to the target category recognition probability and Euclidean distance, the loss function of the model distillation training of the knowledge distillation model is constructed, and the network weights of the student target recognition network when the model output converges are confirmed based on the loss function, and the corresponding trained knowledge distillation model is used as the optimal knowledge distillation model;

[0017] S4: Construct a risk dynamic update mechanism based on Bayesian reasoning, which includes a risk update strategy based on Bayesian reasoning and a time-shifted iteration mechanism for risk confidence;

[0018] The risk update strategy based on Bayesian reasoning is used to obtain the real-time regional risk value of the preset environmental risk area according to the target identification results;

[0019] The time-shifted iteration mechanism of risk confidence is used to dynamically update the regional risk value at the current reconnaissance moment according to the real-time regional risk value obtained at the previous reconnaissance moment, so as to realize dynamic risk assessment of the preset environmental risk area.

[0020] Furthermore, the loss function of the model distillation training of the knowledge distillation model in S3 is obtained as follows:

[0021] S31: Define each network module in the pre-trained teacher target recognition network as ;

[0022] The network modules in the student target recognition network are: ;

[0023] The knowledge distillation intermediate feature map obtained through the teacher target recognition network is expressed as:

[0024] , where: Indicates the network module The output knowledge distillation intermediate feature map; Indicates the network module The output knowledge distillation intermediate feature map; Represents the input image sample data; Indicates the network module The output knowledge distillation intermediate feature map;

[0025] The target recognition intermediate feature map obtained by the student target recognition network is expressed as:

[0026] , where: Indicates the network module Intermediate feature map for target recognition; Indicates the network module The output knowledge distillation intermediate feature map; Indicates the network module The output knowledge distillation intermediate feature map;

[0027] Obtain the Euclidean distance between the knowledge distillation intermediate feature map and the target recognition intermediate feature map, and use it as the loss function of the inter-module association knowledge network to perform knowledge distillation on the student target recognition network through the knowledge distillation intermediate feature map;

[0028] The expression of the loss function of the inter-module association knowledge network is:

[0029] , where: Represents the Euclidean distance between the intermediate feature map of knowledge distillation and the intermediate feature map of target recognition; represents the total number of network modules in the teacher target recognition network;

[0030] S32: Obtaining the target category recognition probability according to the knowledge distillation label and the target category recognition label through the label dark knowledge module;

[0031] That is, the first target probability of target recognition in the knowledge distillation intermediate feature map output by the back-end network module in the teacher target recognition network is obtained, and the second target probability of target recognition in the target recognition intermediate feature map output by the back-end network module in the student target recognition network is obtained;

[0032] And according to the first target probability The probability of the second target Construct label dark knowledge loss function;

[0033] And the expression of the label dark knowledge loss function is:

[0034] ,

[0035] ,

[0036] Where: distilled soft labels representing object recognition; Indicates the category in the corresponding intermediate feature map The value of and They represent the back-end network modules in the teacher / student target recognition network, the output knowledge distillation intermediate feature map / target category in the target recognition intermediate feature map degree of awareness; represents the temperature coefficient used to reduce the difference between the target class and the non-target class in the feature map; represents the label dark knowledge loss function;

[0037] S33: According to the loss function of the inter-module association knowledge network and the label dark knowledge loss function, the loss function of the model distillation training is constructed as follows: , where: Represents the loss function of model distillation training; Represents the direct loss of the training data after passing through the network.

[0038] Furthermore, the risk dynamic update mechanism based on Bayesian reasoning in S4 specifically includes

[0039] S41: For pre-set environmental risk areas Perform grid division to obtain risk area grid ;in Indicates mission area subgrids and , Indicates the grid width, Indicates the grid length;

[0040] S42: confirm whether each task area sub-grid has been scouted;

[0041] If it is confirmed that the task area sub-grid has been detected, step S44 is executed;

[0042] Otherwise, execute step S43;

[0043] S43: Initializing the risk value of the task area sub-grid with a preset risk value;

[0044] S44: Based on the time offset iteration mechanism of risk confidence, the risk value formed by the historical reconnaissance results is iterated with respect to the current time offset, so as to update the current regional risk value of the sub-grid of the mission area;

[0045] And the regional risk value is the probability of inferring that the target exists in the sub-grid of the mission area.

[0046] Further, in S44, the current regional risk value of the task area subgrid is updated, specifically including

[0047] Set event Indicates that the target exists in the mission area sub-grid Events Indicates that the target is not identified, then the edge detection device The probability of successfully identifying the target when it exists for: , where: Indicates that the target exists in the mission area sub-grid The probability of not being detected during the attack;

[0048] According to probability Get the probability that the target exists in the task area sub-grid, and its expression is:

[0049] , where: Represents the task area subgrid Risk value before reconnaissance; Indicates edge detection equipment The success rate of reconnaissance targets; Indicates edge detection equipment The probability of successfully identifying a target if it exists;

[0050] Update and obtain the task area sub-grid according to the probability that the target exists in the task area sub-grid The current regional risk value is expressed as: , where: Representation Grid The risk value obtained after reconnaissance.

[0051] Furthermore, the expression of the time offset iteration mechanism based on risk confidence in S44 is:

[0052] ,

[0053] Where: Indicates the sensitivity of risk value to time shift; Represents the task area subgrid Value at Risk The time parameter since the last reconnaissance.

[0054] The present invention provides a real-time regional risk monitoring method in a weak communication environment, and the beneficial effects are:

[0055] 1) The present invention constructs a knowledge distillation model based on the relationship between soft labels and modules. Under the premise of ensuring the accuracy of the model, the knowledge distillation model can run in real time on edge devices, greatly improving the efficiency and accuracy of real-time regional risk target monitoring and identification in weak communication environments;

[0056] 2) The present invention designs and constructs a dynamic risk update mechanism based on Bayesian reasoning. This update mechanism can realize the serial update of historical information and acquired reconnaissance information through Bayesian reasoning. The mechanism also takes into account the impact of time offset on risk values, so that areas without reconnaissance coverage can scientifically update their risk values ​​in real time. BRIEF DESCRIPTION OF THE DRAWINGS

[0057] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative labor.

[0058] Figure 1 It is a flow chart of the real-time regional risk monitoring method in a weak communication environment of the present invention;

[0059] Figure 2 This is a schematic diagram of the architecture of the real-time regional risk monitoring method in this embodiment;

[0060] Figure 3 is an architecture diagram of the knowledge distillation model based on the relationship between soft labels and modules in this embodiment;

[0061] Figure 4 This is a simulation diagram of the change of risk value when the risk value time offset sensitivity is 0.6 in this embodiment;

[0062] Figure 5 This is a simulation diagram of the change of risk value when the risk value time offset sensitivity is 1 in this embodiment;

[0063] Figure 6 This is a simulation diagram of the change of risk value when the risk value time offset sensitivity is 3 in this embodiment;

[0064] Figure 7It is a simulation curve diagram of the Top-1 accuracy of each model in this embodiment;

[0065] Figure 8 It is a simulation curve diagram of the Top-3 accuracy of each model in this embodiment;

[0066] Fig. 9 It is a simulation curve chart of the Top-5 accuracy of each model in this embodiment. DETAILED DESCRIPTION

[0067] In order to make the purpose, technical solution and advantages of the embodiments of the present invention clearer, the technical solution in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0068] This embodiment provides a real-time regional risk monitoring method in a weak communication environment. Figure 1 As shown, the specific steps include:

[0069] S1: Based on each edge device set in a weak communication environment, autonomously scout the preset environmental risk area to obtain reconnaissance data, and the reconnaissance data at least includes a picture of the environmental risk area;

[0070] And classify the time-sensitive target devices in the environmental risk area pictures to obtain image sample data;

[0071] Divide the image sample data into a data training set and a data test set according to a preset ratio;

[0072] S2: Construct a knowledge distillation model based on the relationship between soft labels and modules;

[0073] The knowledge distillation model includes an inter-module association knowledge network for real-time regional risk target feature extraction based on a symmetric teacher-student target recognition network and a label dark knowledge module for target category label recognition, and the inter-module association knowledge network includes a pre-trained teacher target recognition network and a student target recognition network;

[0074] S3: Train the knowledge distillation model through the data training set to obtain the optimal knowledge distillation model, and run the optimal knowledge distillation model to the edge device to achieve target identification of time-sensitive target devices in the preset environmental risk area;

[0075] Specifically, in order to enable the trained optimal knowledge distillation model to run in real time on edge devices, this embodiment provides a knowledge distillation method SIKD (knowledge distillation based on soft label and Inter-module relationship), that is, the trained large model is used as the teacher model, and the large model to be run on the edge device is used as the student model. For example, the large model includes network models such as deep convolutional networks or residual networks; on the one hand, the dark knowledge contained in the labels is learned from the data perspective, and on the other hand, from the model perspective, the relationship between the modules in the teacher network is learned to reduce the differences between the modules of the teacher network and the student network. Under the premise of ensuring the performance of the model, the model parameters are significantly reduced, so that the edge device has the ability of intelligence; the overall framework of a knowledge distillation method provided is as follows Figure 2 As shown in the figure, the large model above corresponds to the teacher network, and the small model below corresponds to the student network. On the basis of training the student network with the training data, the knowledge of the teacher network is transmitted through the two types of supervision signals distilled from the teacher model, so that the performance of the student model is close to that of the teacher model.

[0076] like Figure 3 As shown, the model training method specifically includes:

[0077] The teacher target recognition network obtains knowledge distillation intermediate feature maps and knowledge distillation labels of different scales used as knowledge distillation information based on the input data training set; the student target recognition network obtains target recognition intermediate feature maps and target category recognition labels of different scales based on the input data training set, and obtains the Euclidean distance between the knowledge distillation intermediate feature map and the target recognition intermediate feature map;

[0078] The label dark knowledge module obtains the target category recognition probability based on the knowledge distillation label and the target category recognition label, and outputs the final target recognition result;

[0079] According to the target category recognition probability and Euclidean distance, the loss function of the model distillation training of the knowledge distillation model is constructed, and the network weights of the student target recognition network when the model output converges are confirmed based on the loss function, and the corresponding trained knowledge distillation model is used as the optimal knowledge distillation model;

[0080] Specifically, the method for obtaining the loss function of the model distillation training of the knowledge distillation model is:

[0081] S31: Define each network module in the pre-trained teacher target recognition network as ;

[0082] The network modules in the student target recognition network are: ;

[0083] The knowledge distillation intermediate feature map obtained through the teacher target recognition network is expressed as:

[0084] , where: Indicates the network module The output knowledge distillation intermediate feature map; Indicates the network module The output knowledge distillation intermediate feature map; Represents the input image sample data; Indicates the network module The output knowledge distillation intermediate feature map;

[0085] The target recognition intermediate feature map obtained by the student target recognition network is expressed as:

[0086] , where: Indicates the network module Intermediate feature map for target recognition; Indicates the network module The output knowledge distillation intermediate feature map; Indicates the network module The output knowledge distillation intermediate feature map;

[0087] Obtain the Euclidean distance between the knowledge distillation intermediate feature map and the target recognition intermediate feature map, and use it as the loss function of the inter-module association knowledge network to perform knowledge distillation on the student target recognition network through the knowledge distillation intermediate feature map;

[0088] The expression of the loss function of the inter-module association knowledge network is:

[0089] , where: Represents the Euclidean distance between the intermediate feature map of knowledge distillation and the intermediate feature map of target recognition; represents the total number of network modules in the teacher target recognition network;

[0090] S32: Obtaining the target category recognition probability according to the knowledge distillation label and the target category recognition label through the label dark knowledge module;

[0091] That is, the first target probability of target recognition in the knowledge distillation intermediate feature map output by the back-end network module in the teacher target recognition network is obtained, and the second target probability of target recognition in the target recognition intermediate feature map output by the back-end network module in the student target recognition network is obtained;

[0092] And according to the first target probability The probability of the second target Construct label dark knowledge loss function;

[0093] And the expression of the label dark knowledge loss function is:

[0094] , ,

[0095] Where: distilled soft labels representing object recognition; Indicates the category in the corresponding intermediate feature map The value of and They represent the back-end network modules in the teacher / student target recognition network, the output knowledge distillation intermediate feature map / target category in the target recognition intermediate feature map degree of awareness; represents the temperature coefficient used to reduce the difference between the target class and the non-target class in the feature map; represents the label dark knowledge loss function;

[0096] S33: According to the loss function of the inter-module association knowledge network and the label dark knowledge loss function, the loss function of the model distillation training is constructed as follows:

[0097] , where: Represents the loss function of model distillation training; Represents the direct loss of training data after passing through the network;

[0098] S4: Construct a risk dynamic update mechanism based on Bayesian reasoning, which includes a risk update strategy based on Bayesian reasoning and a time-shifted iteration mechanism for risk confidence;

[0099] The risk update strategy based on Bayesian reasoning is used to obtain the real-time regional risk value of the preset environmental risk area according to the target identification results; the time offset iteration mechanism of risk confidence is used to dynamically update the regional risk value at the current reconnaissance moment according to the real-time regional risk value obtained at the previous reconnaissance moment, so as to realize dynamic risk assessment of the preset environmental risk area.

[0100] In a specific embodiment, the edge device can independently complete the target recognition task under the support of the aforementioned SIKD method. In this embodiment, in order to solve how to dynamically evaluate and update the risks in various places in the region according to the result information generated by the edge device, a risk dynamic update mechanism based on Bayesian reasoning is designed, and the regional risk information before reconnaissance is used as prior knowledge, and the reconnaissance results of the edge device are used as a posteriori knowledge. An update strategy for regional risks is established, and a real-time update mechanism is designed considering the impact of risk confidence on time offset;

[0101] Specifically, the risk dynamic update mechanism based on Bayesian reasoning in S4 includes

[0102] S41: For pre-set environmental risk areas Perform grid division to obtain risk area grid ;in Indicates task area subgrids to represent the fine-grained area of ​​interest and , Indicates the grid width, Indicates the grid length;

[0103] S42: confirm whether each task area sub-grid has been scouted;

[0104] If it is confirmed that the task area sub-grid has been detected, step S44 is executed;

[0105] Otherwise, execute step S43;

[0106] S43: Initializing the risk value of the task area sub-grid to a preset risk value, specifically setting the risk value of each task area sub-grid to 0.5;

[0107] S44: Based on the time offset iteration mechanism of risk confidence, the risk value formed by the historical reconnaissance results is iterated with respect to the current time offset, so as to update the current regional risk value of the sub-grid of the mission area;

[0108] And the regional risk value is the probability of inferring that the target exists in the sub-grid of the mission area.

[0109] In this embodiment, the area is divided into a grid , the risk value of each grid in the area is initialized, which is divided into two cases. One is that the mission area has never been reconnaissanced, then the risk value of each grid is set to 0.5; the other is that there is historical reconnaissance information in the area, then the risk value formed by the historical reconnaissance result is iterated relative to the current time offset, and the iterative result is used as the initialized risk; let Representation Grid The risk value before reconnaissance, Representation Grid The risk value obtained after reconnaissance is Indicates the success rate of the edge reconnaissance equipment in detecting the target, It indicates the probability that the edge reconnaissance equipment can successfully identify the target when the target exists. If the target is found in the grid, then However, in the case where the target is not detected, the probability of the target's existence needs to be inferred to calculate the grid risk value;

[0110] In a specific embodiment, updating the current regional risk value of the task area subgrid in S44 specifically includes: setting the event Indicates that the target exists in the mission area sub-grid Events Indicates that the target is not identified, then the edge detection device The probability of successfully identifying the target when it exists for:

[0111] , where: Indicates that the target exists in the mission area sub-grid The probability of not being detected during the attack;

[0112] According to probability Get the probability that the target exists in the task area sub-grid, and its expression is:

[0113] , where: Represents the task area subgrid Risk value before reconnaissance; Indicates edge detection equipment The success rate of reconnaissance targets; Indicates edge detection equipment The probability of successfully identifying a target if it exists;

[0114] Update and obtain the task area sub-grid according to the probability that the target exists in the task area sub-grid The current regional risk value is expressed as:

[0115] , where: Representation Grid The risk value obtained after reconnaissance.

[0116] In a specific embodiment, in this embodiment, regional reconnaissance and guarding are usually carried out by edge devices along a specified route, and real-time monitoring of each grid cannot be achieved. Therefore, the confidence of the grid risk value will shift with the time from the last reconnaissance. Therefore, this embodiment designs a time-shifted iteration mechanism for risk confidence, so that the risk value of each grid can be updated in real time. For grids where targets existed before, the longer the time from the last reconnaissance, the more drastic the reduction of its risk value; and for grids where no targets have been detected, its risk value will increase with the increase of time from the last reconnaissance, and the increase will become more and more significant; in both cases, the risk value eventually stabilizes at around 0.5; the task area sub-grid Value at Risk As the time from the last reconnaissance The iteration method is a time-shift iteration mechanism based on risk confidence, and its expression is:

[0117] ,

[0118] Where: Indicates the sensitivity of risk value to time shift; Represents the task area subgrid Value at Risk The time parameter since the last reconnaissance.

[0119] This embodiment also includes a diagram of the effect of the time offset iteration mechanism designed based on risk confidence, such as Figures 4 to 6 shown. Figures 4 to 6 The changes in risk value when the risk value is greater than (0.7), less than (0.3) and equal to 0.5 under different time offset sensitivities are shown respectively.

[0120] In order to verify the feasibility of the proposed real-time regional risk monitoring method and test the performance of the designed SIKD algorithm, this embodiment also designs a regional reconnaissance and risk monitoring scenario, verifies the algorithm by drawing a risk heat map of the selected area, and proves the advancement and efficiency of the proposed algorithm by testing the compression rate, accuracy and operation rate of the distillation model.

[0121] Scenario description and experimental configuration: This embodiment intercepts part of the map of Changsha area as the monitoring and guard area, divides the grid into 0.04 latitude and 0.05 longitude, and sets three drones and three unmanned vehicles as edge devices for patrol tasks. The patrol route adopts a random route and the target adopts random coordinates. The parameters of the drone are set as shown in Table 1 and the parameters of the unmanned vehicle are set as shown in Table 2. Both are simulated using a virtual machine. The teacher model uses the ResNet-50 model, which is trained using the NVIDIA GeForce RTX 3080Ti Laptop graphics card using the cifar10 dataset. The student model uses the ResNet-18 model. Considering that the edge device needs to run other algorithms such as path planning, the model uses single-threaded mode operation.

[0122]

[0123]

[0124] Functional verification and performance analysis: Based on the scenario description and experimental configuration, the proposed regional risk monitoring method is functionally verified to prove the effectiveness of the proposed method and test the distillation model indicators. The SIKD algorithm is compared with the benchmark algorithm to prove the feasibility of the designed algorithm from three perspectives: compression rate, accuracy, and computing speed. It is more advanced than the existing algorithm.

[0125] The designed SIKD algorithm was tested separately, and the model sizes of the distillation model and the original model were compared, as well as the running speeds of the two models when deployed on edge devices, and the computational complexity of the two models was calculated. The computational complexity was calculated as follows:

[0126] For the convolutional layer For example, its computational complexity is:

[0127] , where: Represents the convolutional layer The computational complexity of Indicates the number of input channels; Indicates the number of output channels; Indicates the width of the convolution kernel; Indicates that its convolution kernel is high; and They represent the length and width of the feature map respectively.

[0128] For the fully connected layer Its computational complexity is:

[0129] , where: Represents a fully connected layer The computational complexity of Represents the number of its input neurons; Represents the number of output neurons in the fully connected layer, so according to the computational complexity of the model, it can be concluded that: As shown in Table 3, the SIKD algorithm of this embodiment can reduce the number of parameters of the knowledge distillation model by 52.34%, and achieve a compression rate of 43.71%. The computational complexity is greatly reduced to only 44.21% of the original model. After testing, the CPU running speed is three times higher than before, meeting the real-time computing requirements, further verifying the feasibility of the method of this embodiment.

[0130]

[0131] The above examples verify the feasibility of the SIKD algorithm. Furthermore, this example tests the model's Top-1 accuracy, Top-3 accuracy, and Top-5 accuracy, and compares and analyzes the results of the current advanced benchmark algorithms DKD and IRG. Figures 7 to 9 It can be concluded that the knowledge distillation algorithm designed in this embodiment has better accuracy indicators than the DKD algorithm and the IRG algorithm, and the iteration rounds at the same time are basically better than the other two algorithms, and the performance is relatively excellent.

[0132] This embodiment studies a real-time regional risk monitoring method based on the situation of communication restriction / denial in a regional risk monitoring environment. This method enables edge nodes such as drones / unmanned vehicles to perform real-time target recognition based on their own computing power without transmitting original data to the back-end command post for processing, and only transmits result information with a very small amount of data back to the back-end. The returned result information is used for risk update in the region based on Bayesian reasoning. The method architecture of the real-time regional risk monitoring method in a weak communication environment is as follows: Figure 2 As shown in the figure, our unmanned reconnaissance equipment has drones corresponding to the , the corresponding figure of the unmanned vehicle The trained target recognition model is transformed into a knowledge distillation model that can be deployed to the edge device through knowledge distillation. The edge device independently runs the knowledge distillation model to perform real-time target recognition on the collected data. For the discovered targets, the edge device transmits the target recognition information back to the back-end command post and uses it as the input of the Bayesian reasoning network to dynamically update the risks of the entire area, form regional situational awareness, and provide support for monitoring route planning or target fire strikes.

[0133] This embodiment proposes a real-time regional risk monitoring method in a weak communication environment by focusing on the difficulty of transmitting original video and image data in regional reconnaissance and guarding in a weak communication environment, and the problem of regional risk assessment shifting over time. The designed knowledge distillation model realizes the deployment and real-time operation of the model on the edge device, reducing the communication load. A series of experimental results show that the knowledge distillation model can maintain a high accuracy rate with a high compression rate, and can meet the requirements of real-time operation in a CPU environment. It constructs a risk dynamic update mechanism based on Bayesian reasoning, realizes the online dynamic update of risk values, and improves the reliability of regional risk assessment.

[0134] In summary, the main contributions of this embodiment can be summarized as follows:

[0135] 1) This embodiment is aimed at the risk monitoring environment of weak communication areas, and proposes a real-time regional risk monitoring method architecture. It studies how to effectively implement regional reconnaissance and guard duty and update the regional situation in real time when bandwidth is insufficient or communication is likely to be interrupted.

[0136] 2) This embodiment constructs a knowledge distillation model based on the relationship between soft labels and modules. Under the premise of ensuring the accuracy of the model, the knowledge distillation model can run in real time on the edge device.

[0137] 3) This embodiment designs a risk dynamic update mechanism based on Bayesian reasoning. The mechanism realizes the serial update of historical information and acquired reconnaissance information through Bayesian reasoning. The mechanism also takes into account the impact of time offset on risk value, so that the risk value of areas without reconnaissance coverage can be updated in real time and scientifically.

[0138] 4) This embodiment designs a scenario experiment to verify the feasibility of the method, and tests the performance of the proposed knowledge distillation algorithm and compares it with the current advanced algorithms. The results show that the algorithm accuracy has been significantly improved.

[0139] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or replace some or all of the technical features therein with equivalents. However, these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A real-time regional risk monitoring method in a weak communication environment, characterized in that: The specific steps include: S1: Based on each edge device set in a weak communication environment, autonomously scout the preset environmental risk area to obtain reconnaissance data, and the reconnaissance data at least includes a picture of the environmental risk area; And mark the target category labels of time-sensitive target devices in the environmental risk area pictures to obtain image sample data; Divide the image sample data into a data training set and a data test set according to a preset ratio; S2: Construct a knowledge distillation model based on the relationship between soft labels and modules; The knowledge distillation model includes an inter-module association knowledge network for real-time regional risk target feature extraction based on a symmetric teacher-student target recognition network and a label dark knowledge module for target category label recognition, and the inter-module association knowledge network includes a pre-trained teacher target recognition network and a student target recognition network; S3: Train the knowledge distillation model through the data training set to obtain the optimal knowledge distillation model, and run the optimal knowledge distillation model to the edge device to achieve target identification of time-sensitive target devices in the preset environmental risk area; The model training method specifically includes: The teacher target recognition network obtains knowledge distillation intermediate feature maps and knowledge distillation labels of different scales used as knowledge distillation information based on the input data training set; the student target recognition network obtains target recognition intermediate feature maps and target category recognition labels of different scales based on the input data training set, and obtains the Euclidean distance between the knowledge distillation intermediate feature map and the target recognition intermediate feature map; The label dark knowledge module obtains the target category recognition probability based on the knowledge distillation label and the target category recognition label, and outputs the final target recognition result; According to the target category recognition probability and Euclidean distance, the loss function of the model distillation training of the knowledge distillation model is constructed, and the network weights of the student target recognition network when the model output converges are confirmed based on the loss function, and the corresponding trained knowledge distillation model is used as the optimal knowledge distillation model; S4: Construct a risk dynamic update mechanism based on Bayesian reasoning, which includes a risk update strategy based on Bayesian reasoning and a time-shifted iteration mechanism for risk confidence; The risk update strategy based on Bayesian reasoning is used to obtain the real-time regional risk value of the preset environmental risk area according to the target identification results; The risk dynamic update strategy based on Bayesian reasoning includes: S41: Grid the preset environmental risk area Ω to obtain the risk area grid G ​​= {g1, g2, ..., g i , ...}; where g i represents the i-th task area subgrid and g i =l*d, d represents the grid width, l represents the grid length; S42: confirm whether each task area sub-grid has been scouted; If it is confirmed that the task area sub-grid has been detected, step S44 is executed; Otherwise, execute step S43; S43: Initializing the risk value of the task area sub-grid with a preset risk value; S44: Based on the time offset iteration mechanism of risk confidence, the risk value formed by the historical reconnaissance results is iterated with respect to the current time offset, so as to update the current regional risk value of the sub-grid of the mission area; And the regional risk value is the probability of inferring that the target exists in the sub-grid of the mission area; The time-shifted iteration mechanism of risk confidence is used to dynamically update the regional risk value at the current reconnaissance moment according to the real-time regional risk value obtained at the previous reconnaissance moment, so as to realize dynamic risk assessment of the preset environmental risk area.

2. According to the method of claim 1, the method is characterized in that: The method for obtaining the loss function of the model distillation training of the knowledge distillation model in S3 is: S31: Define each network module in the pre-trained teacher target recognition network as The network modules in the student target recognition network are: The knowledge distillation intermediate feature map obtained through the teacher target recognition network is expressed as Where: Indicates the network module The output knowledge distillation intermediate feature map; Indicates the network module The output knowledge distillation intermediate feature map; X represents the input image sample data; Indicates the network module The output knowledge distillation intermediate feature map; The target recognition intermediate feature map obtained by the student target recognition network is expressed as Where: Indicates the network module The output knowledge distillation intermediate feature map; Indicates the network module The output knowledge distillation intermediate feature map; Indicates the network module The output knowledge distillation intermediate feature map; Obtain the Euclidean distance between the knowledge distillation intermediate feature map and the target recognition intermediate feature map, and use it as the loss function of the inter-module association knowledge network to perform knowledge distillation on the student target recognition network through the knowledge distillation intermediate feature map; The expression of the loss function of the inter-module association knowledge network is Where: loss imr represents the Euclidean distance between the intermediate feature map of knowledge distillation and the intermediate feature map of target recognition; N represents the total number of network modules in the teacher target recognition network; S32: Obtaining the target category recognition probability according to the knowledge distillation label and the target category recognition label through the label dark knowledge module; That is, the first target probability of target recognition in the knowledge distillation intermediate feature map output by the back-end network module in the teacher target recognition network is obtained, and the second target probability of target recognition in the target recognition intermediate feature map output by the back-end network module in the student target recognition network is obtained; And according to the first target probability The probability of the second target Construct label dark knowledge loss function; And the expression of the label dark knowledge loss function is Where: p(z r , T) represents the distilled soft label of target recognition; z r Represents the value of category r in the corresponding intermediate feature map; and Respectively represent the most back-end network module in the teacher / student target recognition network, the output knowledge distillation intermediate feature map / target recognition intermediate feature map, the degree of recognition of the target category b; T represents the temperature coefficient used to reduce the difference between the target class and the non-target class in the feature map; loss soft represents the label dark knowledge loss function; S33: According to the loss function of the inter-module association knowledge network and the label dark knowledge loss function, the loss function of the model distillation training is constructed as follows: loss=loss imr +loss soft +loss label Where: loss represents the loss function of model distillation training; loss label Represents the direct loss of the training data after passing through the network.

3. The real-time regional risk monitoring method in a weak communication environment according to claim 1 is characterized in that: In S44, the current regional risk value of the task area subgrid is updated, specifically including Suppose event A i Indicates that the target exists in the task area subgrid g i In the case where event B indicates that the target is not recognized, the probability that the edge reconnaissance device j can successfully recognize the target when the target exists is for: Where: P(B|A i ) indicates that the target exists in the task area subgrid g i The probability of not being detected during the attack; According to probability Get the probability that the target exists in the task area sub-grid, and its expression is Where: Represents the task area subgrid g i Risk value before reconnaissance; represents the success rate of edge reconnaissance device j in reconnaissance of the target; represents the probability that the edge reconnaissance device j can successfully identify the target when the target exists; According to the probability that the target exists in the task area sub-grid, update and obtain the task area sub-grid g i The current regional risk value is expressed as Where: Represents the grid g i The risk value obtained after reconnaissance.

4. The real-time regional risk monitoring method in a weak communication environment according to claim 1 is characterized in that: The expression of the time shift iteration mechanism based on risk confidence described in S44 is: Where: α represents the sensitivity of the risk value to time offset; Δt represents the task area subgrid g i Value at Risk The time parameter since the last reconnaissance.

Images