Multi-sensor feature selection and collaborative optimization method and device for unmanned system

Abstract

The application discloses a kind of multi-sensor feature selection and collaborative optimization method and equipment for unmanned system, and its steps include: using the data acquisition of multiple sensors carried by unmanned system;Importance evaluation is carried out on features in initial data set using ReliefF algorithm, the weight of each feature is calculated, and important features are selected;Multiple sub-tasks are constructed using these important features and non-dominated sorting algorithm;Through environmental selection iteration, adaptive collaborative optimization is carried out on feature selection scheme, feature subset is evaluated, and finally the optimal feature selection scheme is obtained, which is used for decision support of unmanned system.The application can solve the problems of feature redundancy, large consumption of computing resources and low optimization efficiency when unmanned system performs multi-sensor data fusion in complex environment.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of unmanned systems, and in particular, proposes a multi-sensor feature selection and collaborative optimization method and device for unmanned systems. BACKGROUND

[0002] In recent years, unmanned systems have been widely applied in various fields, especially in autonomous navigation, environmental perception and intelligent decision-making. With the increasing demand for unmanned system applications, how to improve its performance and efficiency has become a hot research issue. Unmanned systems are usually equipped with multiple sensors, such as laser radar, camera, inertial measurement unit (IMU), temperature and humidity sensor, etc., to obtain environmental information and support decision-making. However, in the process of multi-sensor fusion, how to effectively select and optimize sensor features to improve the system's perception ability and decision-making accuracy still faces great challenges.

[0003] With the dramatic increase in sensor data dimensions and the complexity of multi-task, traditional feature selection methods face problems of high-dimensional data, redundant features and interdependence between data in unmanned systems. These problems not only increase the computational complexity, but also may lead to longer system response time, increased storage requirements, and even affect the real-time and accuracy of the system. Therefore, how to remove redundant information, select effective features, and use collaborative optimization methods to improve feature selection efficiency and system performance in the context of multi-sensor has become one of the key technologies for the development of unmanned systems.

[0004] Currently, feature selection methods for multi-sensor data mainly focus on feature extraction and selection based on single sensor data, or feature screening based on artificial rules. However, these methods often fail to adapt to different task requirements and environmental changes in complex and dynamic unmanned system tasks, and the computational efficiency and optimization effect are not ideal on high-dimensional data sets. Therefore, how to design adaptive and multi-task collaborative feature selection methods to improve the global optimization ability of the system has become a technical problem to be solved in the field of unmanned systems. SUMMARY

[0005] The present application proposes a multi-sensor feature selection and collaborative optimization method and device for unmanned systems to address the shortcomings of existing multi-sensor feature selection methods. Through this method, the time and computational resources required in the multi-sensor feature selection process can be effectively reduced. At the same time, through an adaptive mechanism, the most suitable auxiliary task feature selection scheme is dynamically selected according to the evolution state of the target task, so that the optimal feature subset is obtained in the optimization process.

[0006] The present application adopts the following technical solutions to solve the above technical problems:

[0007] Step 1: Collect raw data using various sensors (such as vision sensors, radar, lidar, etc.) on the unmanned system, taking the actual environmental state collected by the sensors as a sample value; preprocess the collected raw data, including noise reduction, normalization, and other operations.

[0008] Step 2: Preprocess the collected raw data and calculate the weight of each feature to select important features; these important features will serve as the initial feature selection scheme for subsequent subtask construction.

[0009] Step 3: Construct multiple subtasks using the selected important features and non-dominated ranking algorithms; the construction of these subtasks is based on the importance of the features and the optimization objectives of the algorithm, providing a specific task framework for subsequent collaborative optimization.

[0010] Step 4: Adaptively and collaboratively optimize the initial feature selection scheme through environmental selection iteration. During the iteration process, the optimization direction is continuously adjusted based on environmental feedback to finally obtain the optimal feature selection scheme. This scheme will provide accurate and efficient data support for the decision support of unmanned systems.

[0011] Furthermore, step two includes the following steps:

[0012] Step 2.1: Use the ReliefF method to evaluate the importance of the original features, assessing the correlation between each feature and the class label to identify key features highly correlated with classification performance. For each feature... The weights are calculated as follows:

[0013] (1)

[0014] In equation (1), Features The weight, For the number of iterations, Indicates the index of the sample. Indicates from 1 to Quantity index; Indicates the category of the sample; This represents the number of nearest neighbor samples; It is a sample Class label; and Indicate category and categories proportion, Indicates sample and In features Differences on and They are samples of the same class; representative sample with in feature , with are different class samples;

[0015] Step 2.2, according to the calculated feature weight values, all features are arranged in descending order, and the sorted feature set and the corresponding weight set are obtained;

[0016] Step 2.3, connecting the points of the minimum feature weight and the maximum feature weight forms a connecting line . Calculate the vertical distance of each point in the feature to , select the point with the largest distance as the knee point, and take its feature weight as the threshold ;

[0017] Step 2.4, keep the features with weight greater than or equal to the threshold , form an important feature set , which is used for subsequent low-dimensional task construction, and the features less than the threshold are regarded as redundant features and deleted;

[0018] Further, step three includes,

[0019] Step 3.1, after obtaining the important feature set , a population of appropriate size is randomly initialized for constructing subtasks, and each individual in the population represents a feature subset, where each feature is either selected or not selected;

[0020] Step 3.2, multi-objective evolutionary optimization of the population in step 3.1;

[0021] Step 3.3, the final population evolved is divided into layers by non-dominated sorting, and all feature subsets in the optimal Pareto layer are extracted as , and the feature subsets extracted in the remaining layers are ;

[0022] Step 3.4, according to the results of non-dominated sorting, construct subtasks { , ,…, }. Among them, the first subtask directly uses as the basic feature subset. For the remaining subtasks ( =2,3,…, ),Will and Combine to form sub-tasks The specific formula is as follows:

[0023] ( =2,3,…, (2)

[0024] Furthermore, step four includes,

[0025] Step 4.1: Initialize parameters and population;

[0026] Step 4.1.1 Initialize parameters;

[0027] Current iteration number t=1, maximum iteration number ;initialization A table, in which all elements are initialized to 0;

[0028] Step 4.1.2: Initialize the population, setting the population size to N; evenly distribute the N individuals in the population to each task. Corresponding to a subpopulation The subpopulation size is N / M; and ={ ,..., ,..., }, express The g-th individual in the dataset. Each individual is encoded using a real number. The position is ={ , ,...., ,..., },in Indicates the dimension of the dataset. Represents the g-th individual. The position of each feature, 0≤ ≤1, when When >0.6, it indicates that the first... One feature is selected and marked as '1', otherwise it is not selected and marked as '0'; Represents the g-th individual. The position of the last feature;

[0029] Step 4.1.2: Calculate the fitness function value;

[0030] The Middle The error rate for each individual is:

[0031] (3)

[0032] Equation (3) represents the error rate obtained when using the feature selection scheme represented by , where c is the number of classes in the dataset, represents the proportion of instances in class k that are correctly identified;

[0033] The fitness function of the th individual in the population is:

[0034] = (4)

[0035] Equation (4) represents the number of features obtained when using the feature selection scheme represented by ;

[0036] The fitness function of the th individual in the population is:

[0037] (5)

[0038] Equation (5) represents the fitness function value obtained when using the feature selection scheme represented by , where = 0.999999, used to adjust the weight value between classification accuracy and the number of selected features;

[0039] Using equation (3), equation (4), and equation (5), select the N / M feature selection schemes with the smallest fitness function values in each sub-population as the tth generation sub-population for each task, and combine the N feature selection schemes as the tth generation population to continue participating in evolution;

[0040] Step 4.2, calculate the similarity between tasks, and select a candidate auxiliary task set for the target task;

[0041] Step 4.2.1, calculate the similarity between each pair of sub-tasks using , and construct a symmetric similarity matrix , where the th element of the symmetric similarity matrix represents the structural similarity between and , and the median of all non-diagonal elements is taken as the global similarity threshold ;

[0042] Step 4.2.2, according to the threshold , construct a ​Candidate auxiliary relation matrix This is used to identify auxiliary relationships between tasks. ≥ ,remember =1 The first in the matrix The first line The value of each element in the column is marked as 1, indicating a task. Can be used as Candidate auxiliary tasks; conversely, if < , =0 means The first in the matrix The first line The value of each element in the column is marked as 0;

[0043] Step 4.2.3: For each target task From the matrix The Extract all values ​​of 1 from the row, and column index This constitutes its candidate auxiliary task set. ;

[0044] Finally, each target task Each corresponds to a static candidate auxiliary task set. This provides an initial search space for subsequent adaptive auxiliary task selection mechanisms;

[0045] Step 4.3: Based on the candidate auxiliary task set selected in Step 4.2, utilize... The reinforcement learning mechanism adaptively selects appropriate auxiliary tasks during the evolution of the target task;

[0046] Step 4.3.1: Design for each target task The table contains three key components: state space construction, state partitioning mechanism, and action space definition.

[0047] State space construction and introduction of migration survival rate And Elite Improvement Rate Two metrics, used to describe the global and local impacts of migration on the target population, are defined as follows:

[0048] (6)

[0049] In equation (6), where Indicates the previous generation of auxiliary population To the target population Offspring resulting from migration = ∩ denotes the number of offspring individuals that successfully survive in the t-th generation; and denote the target population of the (t-1)-th and t-th generations, respectively;

[0050] (7)

[0051] In formula (7), wherein and denote the fitness values of the optimal individuals in the target population of the (t-1)-th and t-th generations, respectively;

[0052] State division mechanism, the current state of the target task is divided into the following three categories by using formula (6) and formula (7):

[0053] (Significant migration gain): when ≥ 0.5; (Moderate migration gain): when <0.5 and >0; (Poor migration effect): when MSR<0.5 and = 0;

[0054] Action space definition, according to step 4.2, the candidate auxiliary task set screened out for it ;

[0055] Step 4.3.2, according to 4.3.1, for each feature selection task Table is constructed for each target task, and the task corresponding to the maximum value is selected as the optimal auxiliary task, and then the sub-population corresponding to the target task is evolved according to step 4.4;

[0056] Step 4.3.3, update the Table according to the reinforcement learning standard, and update the value iteratively by using the following mechanism:

[0057] (8)

[0058] In formula (8), wherein the current auxiliary task the immediate reward under the state , the state of the target population of the t-th generation, the maximum value in all possible auxiliary tasks under the state ; ​∈ [0, 1]: learning rate, controlling the influence of current reward on value update; ∈ [0, 1]: discount factor, measuring the value of future reward;

[0059] where, is defined as:

[0060] = (9);

[0061] Step 4.4, performing knowledge transfer. In each generation, for any target task , the current optimal auxiliary task is selected from the candidate set according to step 4.3.2. Then, the global optimal particle in the auxiliary task population participates in the update of the failed particles in the target population , which is realized by the improved CSO algorithm. Compare the current iteration number with 0.5* , if < 0.5* , evolve the population using formulas (10) and (12); if ≥ 0.5* , evolve the population using formulas (11) and (12):

[0062] (10)

[0063] (11)

[0064] (12)

[0065] where r1, r2, r3 ∈ [0, 1] are random numbers; and represent the current position information of the winning particles and the failed particles in the tth generation target population, respectively; represents the global optimal particle position in the auxiliary population ; represents the winning particle position in the auxiliary population , and represent the speed of the failed particles in the tth generation and the t+1th generation, respectively;

[0066] Step 4.5, after assigning t+1 to t, judge > If yes, output the feature selection scheme with the minimum fitness function value, and use it for unmanned system identification, so as to determine whether there is an obstacle in the environment of multi-sensor identification of the unmanned system, otherwise return to step four for execution.

[0067] In yet another aspect, the present application also discloses a computer readable storage medium storing a computer program, which, when executed by a processor, causes the processor to perform the steps of the above method.

[0068] In yet another aspect, the present application also discloses a computer device comprising a memory and a processor, wherein the memory stores a computer program, which, when executed by the processor, causes the processor to perform the steps of the above method.

[0069] Compared with the prior art, the present application has the following beneficial effects:

[0070] 1. The present application proposes a dimension reduction strategy for multi-sensor feature selection of unmanned systems, which effectively filters the high-dimensional feature space through the RelieF method and extracts the feature subset with the maximum discrimination ability. This strategy eliminates redundant or irrelevant features by evaluating the correlation and importance between features, thereby significantly reducing the feature dimension, providing a more streamlined and effective feature space for subsequent task generation, improving system processing efficiency, and laying a good foundation for multi-sensor collaborative optimization.

[0071] 2. The present application designs an auxiliary task adaptive selection strategy, which automatically selects the most suitable auxiliary task for the target task by combining static task similarity analysis and dynamic selection mechanism. This strategy can effectively reduce the risk of negative transfer and ensure efficient knowledge transfer between different tasks, thereby improving the accuracy and stability of overall task execution and facilitating subsequent collaborative optimization.

[0072] 3. The present application adopts a collaborative optimization approach, which performs cross-task knowledge transfer through an improved collaborative particle swarm optimization (CSO) algorithm. During collaborative optimization, the search direction and speed of particles are adjusted to enhance the collaboration between tasks and further improve the overall optimization effect. Combined with multi-task learning and evolutionary mechanisms, the collaborative optimization approach of the present application can efficiently share and optimize features between tasks, ultimately realizing adaptive decision-making and task execution of unmanned systems in complex environments. BRIEF DESCRIPTION OF DRAWINGS

[0073] Figure 1 is a flowchart of the method of the present application;

[0074] Figure 2 is a schematic diagram of the present application for selecting a candidate auxiliary task set for a target task;

[0075] Figure 3 To the present application is An illustration of selecting an optimal auxiliary task as a target task. DETAILED DESCRIPTION

[0076] In the embodiment of the present application, a multi-sensor feature selection and collaborative optimization method and device for unmanned systems are provided. The method includes the following steps:

[0077] Step one, using the sensors carried by the unmanned system to collect raw data, and taking the actual environment state collected by the sensors as a sample value;

[0078] Step two, preprocessing the collected raw data, calculating the weight of each feature, and screening out important features; these important features will be used as the initial feature selection scheme for subsequent subtask construction;

[0079] Step three, using the screened important features to construct multiple subtasks in combination with the non-dominated sorting algorithm; the construction of these subtasks is based on the importance of features and the optimization target of the algorithm, and provides a specific task framework for subsequent collaborative optimization;

[0080] Step four, through the environment selection iteration method, the initial feature selection scheme is adaptively and collaboratively optimized; in the iteration process, the optimization direction is continuously adjusted according to the environmental feedback, and finally the optimal feature selection scheme is obtained, which will provide accurate and efficient data basis for the decision support of the unmanned system.

[0081] Referring to Figure 1 the specific steps are as follows:

[0082] Step one, using the sensors carried by the unmanned system to collect raw data, and taking the actual environment state collected by the sensors as a sample value; Table 1 shows a feature set for unmanned system identification containing 9 statistical features and 3 samples. Features 1-3 are visual sensor features, features 4-6 are radar features, and features 7-9 are lidar features. Feature 1 is the mean brightness, 2 is the edge intensity, 3 is the texture complexity, 4 is the target distance, 5 is the target speed, 6 is the reflection intensity, 7 is the point cloud density, 8 is the height distribution, and 9 is the shape feature. Column 10 indicates the target category corresponding to the sample, where 0 represents non-obstacles (i.e., no obstacles in the environment) and 1 represents obstacles (i.e., obstacles exist in the environment).

[0083] Table 1 Environmental State Feature Set

[0084]

[0085] Step 2: Calculate the weight of each feature and select the important features;

[0086] Step 2.1: Use the ReliefF method to evaluate the importance of the original features, assessing the correlation between each feature and the class label to identify key features highly correlated with classification performance. For each feature... The weights are calculated as follows:

[0087] (1)

[0088] In equation (1), Features The weight, For the number of iterations, Indicates the index of the sample. Indicates from 1 to Quantity index; Indicates the category of the sample; This represents the number of nearest neighbor samples; It is a sample Class label; and Indicates category and categories proportion, Indicates sample and In features Differences on and They are samples of the same class; Indicates sample and In features Differences on and These are samples from different classes; Table 2 shows the importance of each feature:

[0089] Table 2 Feature importance

[0090]

[0091] Step 2.2, feature weight ordering. According to the calculated feature weight values, all features are arranged in descending order to obtain the sorted feature set and the corresponding weight set ;

[0092] Table 3 Feature set arranged in descending order of feature weight

[0093]

[0094] Step 2.3, connecting the points with the minimum feature weight and the maximum feature weight to form a connecting line . Calculate the vertical distance from each point in the remaining features to , and select the point with the largest distance as the knee point ;

[0095] Step 2.4, keep the features with feature weight greater than or equal to the threshold to form the important feature set , ={ , , , , , } for subsequent low-dimensional task construction, and features less than the threshold are considered redundant features and are deleted;

[0096] Table 4 Important feature set

[0097]

[0098] Step three, use these important features and non-dominated sorting algorithm to construct multiple subtasks;

[0099] Step 3.1, after obtaining the important feature set ={ , , , , , }, randomly initialize a population of appropriate size for constructing subtasks (this population is only used to construct tasks and does not participate in subsequent evolution, and the size of this population in this experiment is 10), each individual in the population represents a feature subset, where each feature is either selected or not selected;

[0100] Step 3.2, multi-objective evolutionary optimization of the population in step 3.1;

[0101] Step 3.3, the final population evolved is divided into layers in the experimental test =4, the optimal Pareto layer is extracted all feature subsets in the feature subsets extracted in the remaining layers ( =2,3,4) are ;

[0102] The individuals contained in layer 1 are {1, 0, 0, 0, 0, 0}, {0, 1, 0, 0, 0, 0} and {1, 1, 0, 0, 0, 0}, so the union of all selected features is taken, the index of which is {1, 1, 0, 0, 0, 0, 0}; The individuals contained in layer 2 are {1, 0, 0, 0, 0, 0} and {0, 0, 0, 0, 0, 1}, and the union of all features is taken, the index of which is {1, 0, 0, 0, 0, 0, 1}; The individuals contained in layer 3 are {1, 0, 0, 0, 0, 1} and {0, 0, 0, 0, 1, 1}, the index of which is {1, 0, 0, 0, 1, 1}; The individuals contained in layer 4 are {0, 0, 0, 0, 1, 0}, {0, 0, 1, 0, 0, 0, 0} and {0, 1, 0, 0, 0, 0}, the index of which is {0, 1, 1, 0, 1, 0};

[0103] Step 3.4, according to the results of non-dominated sorting, construct subtasks { , , , }. Among them, the first subtask directly uses as the basic feature subset. For the remaining subtasks ( =2,3,4), combine with to form subtask . The specific formula is as follows:

[0104] ( =2,3,4) (2)

[0105] According to steps 3.3 and 3.4, four subtasks are constructed ={1, 1, 0, 0, 0, 0}, ={1, 1, 0, 0, 0, 1}, ={1, 1, 0, 0, 1, 1} and ={1, 1, 1, 0, 1, 0};

[0106] Step 4: Through environment selection iteration, the feature selection scheme is collaboratively optimized to obtain the optimal feature selection scheme for decision support of unmanned systems.

[0107] Step 4.1: Initialize parameters and population;

[0108] Step 4.1.1 Initialize parameters;

[0109] Current iteration number t=1, maximum iteration number ;initialization A table, in which all elements are initialized to 0;

[0110] Step 4.1.2: Initialize the population, setting the population size to N; evenly distribute the N individuals in the population to each task. Corresponding to a subpopulation The subpopulation size is N / 4 (i.e., N / M, M=4); and ={ ,..., ,..., }, express The g-th individual in the dataset. Each individual is encoded using a real number. The position is ={ , ,...., ,..., },in Indicates the dimension of the dataset. Represents the g-th individual. The position of each feature, 0≤ ≤1, when When >0.6, it indicates that the first... One feature is selected and marked as '1', otherwise it is not selected and marked as '0'; Represents the g-th individual. The position of the last feature;

[0111] Step 4.1.2: Calculate the fitness function value;

[0112] The Middle The error rate for each individual is:

[0113] (3)

[0114] Equation (3) indicates the use of The error rate obtained when the feature selection scheme is used for classification is represented by c, where c is the number of classes in the dataset. This represents the proportion of correctly identified instances in class k;

[0115] The Middle The number of features for each individual is:

[0116] = (4)

[0117] Equation (4) indicates the use of The number of features obtained when classifying using the feature selection scheme represented;

[0118] The Middle The fitness function for each individual is:

[0119] (5)

[0120] Equation (5) indicates the use of The fitness function value obtained when classifying using the feature selection scheme represented by the example, where =0.999999, used to adjust the weight value between classification accuracy and the selected number of features;

[0121] Using formulas (3), (4), and (5), select the N / 4 feature selection schemes with the smallest fitness function values ​​in each subpopulation as the t-th generation subpopulation for each task, and merge these N feature selection schemes as the t-th generation population to continue participating in evolution;

[0122] Step 4.2: Calculate the similarity between tasks and select a set of candidate auxiliary tasks for the target task (the four sub-tasks are used in turn for evolutionary optimization of the target task).

[0123] Step 4.2.1, adopt Similarity calculations are performed between each pair of subtasks to construct a symmetric similarity matrix. The ( )th of the symmetric similarity matrix , ) elements represent and Structural similarity between them; such as Figure 2 As shown in (a), the similarity between tasks is calculated, and the task can be obtained from the third row of the table. With the task , and The similarity scores were 0.5, 0.25, and 0.4, respectively; the median of all off-diagonal elements was taken as the global similarity threshold. ( =0.29);

[0124] Step 4.2.2: Based on the threshold , build a Candidate auxiliary relation matrix This is used to identify auxiliary relationships between tasks. ≥ ,remember =1 The first in the matrix The first line The value of each element in the column is marked as 1, indicating a task. Can be used as Candidate auxiliary tasks; conversely, if < , =0 means The first in the matrix The first line The value of each element in the column is marked as 0; for example Figure 2 As shown in (b), according to the third row of the table, 0.5 > 0.29, therefore... =1, and the other two candidate elements are calculated in the same way. =0, =0. ( Figure 2 The element marked as NULL in (a) Figure 2 The corresponding position of (b) is directly marked as 0.

[0125] Step 4.2.3: For each target task From the matrix The Extract all values ​​of 1 from the row, and column index This constitutes its candidate auxiliary task set. ;

[0126] Finally, each target task Each corresponds to a static candidate auxiliary task set. This provides an initial search space for the subsequent adaptive auxiliary task selection mechanism, resulting in a candidate auxiliary task set for each task. ={ }, ={ }, ={ ,}and ={ , };

[0127] Step 4.3, setting the target task of current evolution as , the candidate auxiliary task set selected according to step 4.2 ={ , }, using reinforcement learning mechanism, in the evolution process of , self-adaptively select appropriate auxiliary tasks;

[0128] Step 4.3.1, for the current target task design table, including three key contents: state space construction, state division mechanism and action space definition;

[0129] State space construction, introduce migration survival rate and elite improvement rate Two measurement indexes, respectively used to describe the global and local influence of migration on target population, defined as follows:

[0130] (6)

[0131] In formula (6), wherein represents the offspring produced by the migration of the last generation auxiliary population to the target population , = ∩ represents the number of offspring individuals that successfully survive in the tth generation; and respectively represent the t-1th generation and the tth generation target population;

[0132] (7)

[0133] In formula (7), wherein and respectively represent the fitness values of the optimal individuals in the t-1th generation and the tth generation target population;

[0134] State division mechanism, using formula (6) and formula (7) to divide the current state of the target task into the following three categories:

[0135] (Significant migration gain): when ≥ 0.5; (Moderate migration gain): when <0.5 and > 0; (migration effect is poor): when MSR < 0.5 and = 0;

[0136] Action space definition, according to step 4.2, limit to the set of candidate auxiliary tasks screened out = { , },

[0137] Therefore, mark and as the action of the current table;

[0138] Step 4.3.2, according to 4.3.1, select the task constructed for each feature Table, select the task corresponding to the maximum value for each target task as the optimal auxiliary task, and then evolve the sub-population corresponding to the target task according to step 4.4; the population of the current evolution task , according to the table obtained in step 4.3.1 as shown in (a) of , calculate each Figure 3 value as shown in (b) of , set the evolution state of Figure 3 to , according to the second row of the table, 0.5>0.1, so select as the auxiliary task of ;

[0139] Step 4.3.3, update the table according to the reinforcement learning standard, and use the following mechanism to iteratively update the value:

[0140] (8)

[0141] In equation (8), where the immediate reward of the current auxiliary task in state , the state of the target population of the generation, the maximum value in all possible auxiliary tasks in state ; ∈[0,1]: learning rate, controls the influence of the current reward on the update of value; ∈[0,1]: discount factor, used to measure the value of future rewards; ​

[0142] in, Defined as:

[0143] = (9)

[0144] choose As After completing the auxiliary task, evolve according to step 4.4. According to formulas (6) and (7), let After successful evolution, it enters the next state. Using equations (8) and (9), The value was updated from 0.5 to 0.587, and the update process is as follows: Figure 3 As shown in (c);

[0145] Step 4.4: Perform knowledge transfer. In each generation, for the target task... According to step 4.3.2, the current optimal auxiliary task is selected from the candidate set. Subsequently, the auxiliary task population was... The globally optimal particle participates in the target population. The update of failed particles is implemented using an improved CSO algorithm. The current iteration number is then set. With 0.5* In comparison, if <0.5* The population is evolved using formulas (10) and (12); if ≥0.5* The population is evolved using formulas (11) and (12) (where i=3 and j=4):

[0146] (10)

[0147] (11)

[0148] (12)

[0149] Where r1, r2, r3 ∈ [0, 1] are random numbers; and These represent the current position information of the winning and losing particles in the target population of generation t, respectively. Indicates auxiliary population The global optimal particle position; Indicates auxiliary population The winning particle position in the middle, and These represent the velocities of the failed particles in generation t and generation (t+1), respectively.

[0150] Step 4.5, after assigning t+1 to t, judging whether it is true, if true, outputting a feature selection scheme with the minimum fitness function value for unmanned system identification, so as to judge whether there is an obstacle in the environment of multi-sensor identification of the unmanned system, otherwise returning to step four for execution.

[0151] In order to verify the effectiveness of the scheme, simulation experiments were conducted on 12 high-dimensional data sets. These data sets cover different feature dimensions and sample sizes; the purpose of the experiment is to optimize the size of the feature subset through the feature selection method of the scheme, while improving the classification accuracy;

[0152] Data set description

[0153] The following is the basic information of the 12 high-dimensional data sets:

[0154] Table 5 High-dimensional data sets

[0155]

[0156] Experimental results

[0157] On each data set, the feature selection method of the scheme is applied, and the size of the feature subset and the classification accuracy are recorded. The following is the detailed data of the experimental results:

[0158] Table 6 Experimental results

[0159]

[0160] In summary, the scheme optimizes feature selection, not only significantly reduces the size of the feature subset, but also greatly improves the classification accuracy, has high practical value and wide application prospect.

[0161] In another aspect, the application also discloses a computer readable storage medium, which stores a computer program, and the computer program is executed by a processor to make the processor execute steps of the above method.

[0162] In another aspect, the application also discloses a computer device, which comprises a memory and a processor, and the memory stores a computer program, and the computer program is executed by the processor to make the processor execute steps of the above method.

[0163] It can be understood that the system, device and storage medium provided by the embodiments of the application correspond to the method provided by the embodiments of the application, and the explanation, examples and beneficial effects of related contents can be referred to the corresponding part in the above method.

[0164] ​In the embodiments described above, all or some of the steps can be implemented by hardware, software, firmware or any combination thereof. When implemented by software, all or some of the steps can be implemented in the form of one or more computer programs or program elements. The computer programs reside (at least temporarily) in a memory of a computer during execution. The memory can be a RAM memory, a flash memory, a ROM memory, an EPROM memory, or any other suitable memory. The memory can be integral to or separate from the computer. The computer programs can be written in any suitable programming language, such as C, C++, Java, Visual Basic, etc. The computer programs can be written in assembly or machine language, if desired. The computer programs can be distributed over network coupled file servers, or can be distributed by any other suitable means.

[0165] It is to be understood that the terminology "first", "second", etc. used herein merely for the purpose of distinguishing one entity or action from another, and does not necessarily imply these entities or actions are mutually exclusive or are either temporal or spatial priorities of one another. Moreover, the terms "comprising", "including", or any other variation thereof, are intended to cover a non-exclusive inclusion, such that processes, methods, articles, or apparatuses that comprise a list of elements are not necessarily limited to those elements, but can include other elements not expressly listed or inherent to such processes, methods, articles, or apparatuses. Furthermore, the term "comprising" does not exclude the presence of elements or additional steps other than those listed in the process, method, article, or apparatus.

[0166] Each of the embodiments described in the specification adopt a related manner, and the same or similar parts between the embodiments can be referred to each other. Each embodiment focuses on the difference from other embodiments. In particular, for the system embodiments, since they are basically similar to the method embodiments, the description is relatively simple, and the relevant parts can be referred to the part of the method embodiments.

[0167] The above examples are only used to illustrate the technical solutions of the present application, and are not intended to limit the present application; although the present application has been described in detail with reference to the foregoing examples, those skilled in the art should understand that the technical solutions recorded in the foregoing examples can be modified, or some technical features can be replaced by equivalent features; and these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present application.

Claims

1. A method for multi-sensor feature selection and collaborative optimization for unmanned systems, the method comprising: The steps are as follows: Step one, using the sensor carried by the unmanned system to collect raw data, taking the actual environmental state collected by the sensor as a sample value; Step two, preprocessing the collected raw data, calculating the weight of each feature, and screening out important features; These important features will be used as the initial feature selection scheme for subsequent sub-task construction; Step three, using the important features selected to construct multiple sub-tasks by combining the non-dominated sorting algorithm; The construction of these sub-tasks is based on the importance of features and the optimization target of the algorithm, and provides a specific task framework for subsequent collaborative optimization; Step four, through the iterative method of environment selection, the initial feature selection scheme is adaptively and collaboratively optimized; in the iteration process, the optimization direction is adjusted according to the environmental feedback, and finally the optimal feature selection scheme is obtained, which will provide accurate and efficient data basis for the decision support of the unmanned system; Step four specifically includes: Step 4.1, initialization of parameters and population; Step 4.2, calculate the similarity between tasks, select the candidate auxiliary task set of the target task; finally each target task corresponds to a static candidate auxiliary task set provides an initial search space for the subsequent adaptive auxiliary task selection mechanism; Step 4.

3. According to the candidate auxiliary task set selected in step 4.2, using a reinforcement learning mechanism, to adaptively select appropriate auxiliary tasks in the evolution process of the target task. Step 4.3.1, Design for each target task Table, including three key content: state space construction, state division mechanism and action space definition; The state space construction includes introducing a migration survival rate and an elite improvement rate Two metrics, respectively describing the global and local impact of migration on the target population, are defined as follows: (6) In formula (6), where denotes the offspring produced by the previous generation of helper population to the target population migrating to the target population, = ∩ denotes the number of offspring individuals that successfully survive in the tth generation, and denote the (t-1)th and tth generations of target population, respectively; (7) In formula (7), wherein and respectively represent the fitness value of the optimal individual in the t 1 generation and the t generation target population; The state division mechanism includes dividing the current state of the target task into the following three categories by using formula (6) and formula (7): , significant migration gain: when ≥ 0.5; Medium migration gain: when < 0.5 and > 0; , poor migration effect: when MSR < 0.5 and = 0; The action space definition comprises, according to step 4.2, limiting the set of candidate secondary tasks filtered out for it ; Step 4.3.

2. Select the task corresponding to the maximum value of the table for each feature constructed according to 4.3.1 as the optimal auxiliary task, and then evolve the sub-population corresponding to the target task according to step 4.

4. Table, select the task corresponding to the maximum value of the table for each target task Step 4.3.

2. Select the task corresponding to the maximum value of the table for each feature constructed according to 4.3.1 as the optimal auxiliary task, and then evolve the sub-population corresponding to the target task according to step 4.

4. Step 4.3.3, update according to reinforcement learning criteria Table, the values are iteratively updated using the following mechanism: values are iteratively updated using the following mechanism: (8) In equation (8), where Indicates the current auxiliary task In state Instant rewards below, Indicates the first The state of the target population, In state The maximum of all possible auxiliary tasks value; ∈[0,1]: Learning rate, controlling the current reward pair The degree of impact of value updates; ∈[0,1]: Discount factor, used to measure the value of future rewards; wherein is defined as: = (9) Step 4.4, knowledge transfer; In each generation, for any target task , the current optimal auxiliary task is selected from the candidate set according to step 4.3.2 ; then, the global optimal particle in the auxiliary task population is allowed to participate in the update of the failed particles in the target population , and this process is implemented using the improved CSO algorithm; the current iteration number is compared with 0.5 ; if < 0.5 , the populations are evolved using formulas (10) and (12); if ≥ 0.5 , the populations are evolved using formulas (11) and (12). (10) (11) (12) where r1, r2, r3 ∈ [0, 1] are random numbers; and respectively represent the current position information of the winning particles and the losing particles in the tth generation target population; represent the global optimal particle position in the auxiliary population ; represent the winning particle position in the auxiliary population ; and respectively represent the speed of the losing particles in the tth generation and the t+1th generation; Step 4.5: After assigning t+1 to t, determine... > If the condition is met, the feature selection scheme with the minimum fitness function value is output for unmanned system recognition, thereby determining whether there are obstacles in the environment of the multi-sensor recognition of the unmanned system; otherwise, return to step four for execution.

2. The multi-sensor feature selection and co-optimization method for unmanned systems according to claim 1, wherein: Step two includes, Step 2.1, using the ReliefF method to evaluate the importance of the original features, evaluate the correlation between each feature and the class label, so as to identify the key features highly related to the classification performance; for each feature The weight is calculated as follows: (1) in formula (1), is a feature of weight, is a number of iterations, denotes an index of a sample, denotes a number index from 1 to ; denotes a class of a sample; is a number of nearest neighbor samples; is a class label of a sample ; and denote a proportion of classes and classes , denotes a difference between a sample and on a feature , and are samples of the same class; denotes a difference between a sample and on a feature , and are samples of different classes; Step 2.2, according to the calculated feature weight values, all features are arranged in descending order to obtain the sorted feature set and the corresponding weight set ; Step 2.

3. The points connecting the minimum feature weight and the maximum feature weight form a connecting line ; Calculate the connection line from each point in the feature. The vertical distance is used to select the point with the largest distance as the knee point, and the feature weight of the knee point is used as the threshold. ; Step 2.4, retaining features with significance greater than or equal to a threshold of features, forming a set of significant features for subsequent low-dimensional task construction, features with significance less than the threshold are considered redundant and deleted.

3. The multi-sensor feature selection and co-optimization method for unmanned systems according to claim 2, wherein: Step three specifically includes, Step 3.1, obtaining a set of important features After that, a population of a given size is randomly initialized for constructing the subtasks, each individual in the population represents a subset of features, where each feature is either selected or not selected; Step 3.2, multi-objective evolutionary optimization of the population in step 3.1; Step 3.3, the final population evolved is divided into layers, the optimal Pareto layer is extracted with all feature subsets in it and the feature subsets extracted in the remaining layers ; Step 3.

4. Constructing subtasks according to the result of non-dominated sorting , ,…, }; where the first subtask directly uses as the base feature subset; for the rest of the subtasks , =2,3,…, , combine and to form the subtask ; the specific formula is as follows: ， =2,3,…, (2)。 4. The multi-sensor feature selection and co-optimization method for unmanned systems according to claim 3, wherein: Step 4.2 specifically further includes: Step 4.2.1, adopt Similarity calculates the similarity between each pair of subtasks, and constructs a symmetric similarity matrix , where the (i, j)th element of the symmetric similarity matrix represents the structural similarity between , and , and the median of all non-diagonal elements is taken as the global similarity threshold ;​ Step 4.2.2, according to the threshold value , a candidate auxiliary relationship matrix is constructed to identify the auxiliary relationship between tasks; if ≥ , record =1, that is the value of the element in the first row and the first column of the matrix is marked as 1, indicating that the task can be a candidate auxiliary task of ; otherwise, if < , =0, that is the value of the element in the first row and the first column of the matrix is marked as 0; Step 4.2.3, for each target task , extract all values of 1 from the matrix , and , extract all values of 1 from the matrix , and , extract all values of 1 from the matrix .

5. The multi-sensor feature selection and co-optimization method for unmanned systems according to claim 4, wherein: Step 4.1 specifically includes, Step 4.1.1, initialization of parameters; current iteration number t = 1, maximum iteration number ; initialize table, all elements of the table are initialized to 0; Step 4.1.2: Initialize the population, setting the population size to N; evenly distribute the N individuals in the population to each task. Corresponding to a subpopulation The subpopulation size is N / M; and ={ ,..., ,..., }, express The g-th individual in the dataset; each individual is encoded using a real number. The position is ={ , ,...., ,..., },in Indicates the dimension of the dataset. Represents the g-th individual. The position of each feature, 0≤ ≤1, when When >0.6, it indicates that the first... One feature is selected and marked as '1', otherwise it is not selected and marked as '0'; Represents the g-th individual. The position of the last feature; Step 4.1.3, calculate the fitness function value; The Middle The error rate for each individual is: (3) Equation (3) represents the use of the error rate obtained when classifying using the feature selection scheme represented by represents the proportion of instances in class k that are correctly identified. The Middle The number of features for each individual is: = (4) Equation (4) represents the use of the number of features obtained when classifying using the feature selection scheme represented by The Middle The fitness function for each individual is: （5） Equation (5) indicates the use of The fitness function value obtained when classifying using the feature selection scheme represented by the example, where =0.999999, used to adjust the weight value between classification accuracy and the selected number of features; Use formula (3), formula (4), formula (5) to select the N / M feature selection schemes with the smallest fitness function value in each sub-population as the tth generation sub-population of each task, and combine the N feature selection schemes as the tth generation population to continue to participate in evolution. 6.A computer device, comprising a memory and a processor, wherein the memory stores a computer program, and the computer device is configured to perform the method according to any one of claims 1-5 when the computer program is executed by the processor. The computer program is executed by the processor, so that the processor executes the steps of the method as claimed in any one of claims 1 to 5.

Images