Typical task system performance evaluation method based on entropy weight method and graph convolutional network

Abstract

The invention relates to the technical field of evaluation and judgment, in particular to a typical task system performance evaluation method based on an entropy weight method and a graph convolutional network, which comprises the following steps of: weighting evaluation index data with a hierarchical structure by using the entropy weight method, and converting the traditional hierarchical index data into a data structure of a non-Euclidean space; and constructing a depth map convolutional network model, extracting deep features in an input sample, mining a coupling relationship between sensor data and task system performance, and realizing evaluation of the task system performance. The typical task system performance is preliminarily evaluated through the information entropy, and the relation between bottom layer single items can be well processed. According to the method, the graph network is applied to typical task system performance evaluation, the model can fully extract shallow and deep feature information of an input sample, the graph attention network is adopted to fuse and enhance the features, and it is guaranteed that the model has higher robustness and accuracy.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of task system performance evaluation and determination, and particularly relates to a typical task system performance evaluation method based on an entropy weight method and a graph convolution network. BACKGROUND

[0002] A typical task system is a core component of complex operations, including scheduling systems, control units, sensor networks, and decision support modules. Its tasks usually involve efficient resource scheduling, precise control and coordination, intelligent decision analysis, etc. In the fields of industrial production, aerospace, and automated logistics, the performance of a typical task system directly affects the overall operational efficiency and reliability of the system. With the increasing complexity and intelligence of industrial systems, typical task systems face higher performance demands. Therefore, it is of great significance to use advanced technologies to scientifically and accurately evaluate the performance of typical task systems to ensure system stability, improve operational efficiency, and optimize task strategies.

[0003] Currently, the methods for performance evaluation of typical task systems mainly include two types: knowledge-driven and traditional data statistical methods. Knowledge-driven methods rely on experts' domain knowledge and experience, such as the analytic hierarchy process (AHP) and fuzzy comprehensive evaluation method. The advantage of this type of method is that it can combine qualitative and quantitative analysis to effectively evaluate complex task systems. However, since the calculation of weights relies on experts' subjective scoring, the evaluation process has a high degree of subjectivity, especially when faced with changing task environments and complex operation scenarios, the evaluation results may be biased. Traditional data statistical methods such as entropy weight method and grey theory evaluation can better handle the uncertainty factors in the evaluation process and are suitable for multi-index and multi-factor comprehensive evaluation. However, this type of method usually only performs shallow feature extraction and is difficult to fully exploit deep information in task system sensor data, so its generalization and robustness are poor in complex environmental changes. When the task demand or environment changes dramatically, or the sensor signal is disturbed, the evaluation results are often unreliable.

[0004] Currently, the performance evaluation methods are mainly based on knowledge-driven and traditional data statistical methods. Knowledge-driven methods mainly rely on expert experience and subjective methods such as the analytic hierarchy process and fuzzy comprehensive evaluation method.

[0005] In the prior art, the paper "Performance Evaluation and Life Prediction of Power Transmission Equipment Based on Multi-physical Field Coupling" considers the multi-physical field model of power transmission equipment considering the coupling of electromagnetic field, thermal field and mechanical field. The final evaluation result is obtained by layer-by-layer calculation based on cloud rules. Based on the simulation results, a performance evaluation index system is constructed, and a comprehensive evaluation method is proposed.

[0006] The paper "Fuzzy Evaluation Method of Electronic Warfare System Effectiveness" combines fuzzy comprehensive evaluation and analytic hierarchy process to evaluate the combat effectiveness of electronic warfare system. Firstly, the paper determines the index set and of the evaluation method through analytic hierarchy process, and establishes the fuzzy distribution to form the fuzzy evaluation matrix, and then realizes the comprehensive evaluation of the typical task system.

[0007] The paper "Evaluation of Radar Seeker Jamming Effectiveness Based on Exponential Scale Analytic Hierarchy Process and Vague Set" realizes the evaluation of radar seeker jamming effectiveness based on exponential scale analytic hierarchy process and vague set. Firstly, the paper establishes the evaluation index system from the two stages of searching target and tracking target, and then determines the index weight through exponential scale analytic hierarchy process, and then determines the comprehensive decision vague set to realize the evaluation of radar seeker jamming effectiveness.

[0008] The advantage of this method is to combine qualitative and quantitative analysis, which is beneficial to the effectiveness calculation of complex system. However, the calculation of weight depends on the scoring of expert experience, which is difficult to realize, and the final evaluation result has strong subjectivity.

[0009] The evaluation method based on traditional data statistics includes entropy weight method and grey theory evaluation.

[0010] The paper "Health Status Evaluation of Electronic Countermeasure System Based on Combination Weighting and Normal Cloud Model" applies combination weighting method and normal cloud model to the health evaluation of electronic countermeasure system. The method realizes the weighting of evaluation index system through combination of expert weight and observable weight based on entropy weight method, and establishes the normal membership cloud model to realize the health evaluation of electronic countermeasure system.

[0011] The patent application No. CN202410888206.6, named "Load Switch Performance Detection Method and System, Electronic Equipment and Storage Medium", realizes the performance detection of load switch by combining kernel principal component analysis and reinforcement learning method. Firstly, the method uses kernel principal component analysis to extract features and high-order dimension reduction of multi-source detection data to obtain principal component features. Then, it uses reinforcement learning to evaluate the importance of feature data at each time point and dynamically updates the weight of performance evaluation model. Finally, through nonlinear feedback mechanism and current feature fusion, the performance prediction result of load switch is obtained, so as to realize the dynamic performance evaluation of load switch.

[0012] This kind of method can better solve the influence of uncertain factors and is suitable for comprehensive evaluation of complex system with multiple factors and indexes. However, for the current complex comprehensive task system, this kind of shallow evaluation method cannot extract deep information of sensor data, and the generalization and robustness of the method are poor. When the working condition of the aircraft changes or the sensor is disturbed, the evaluation result is often unreliable.

[0013] In recent years, with the development of big data, the large amount of data of the aircraft in the execution of the task is saved, and the improvement of the computing power makes the training of the neural network become more simple.

[0014] The paper "Electronic Warfare Effectiveness Evaluation Based on Two Type Fuzzy Neural Network" realizes the effectiveness evaluation of electronic warfare system based on two type fuzzy neural network. The method firstly establishes the electronic warfare evaluation index system, and then uses the self-adaptive training and mapping fuzzy neural network to model the effectiveness evaluation system of electronic warfare, and finally obtains the evaluation result. However, the deep learning method, especially the neural network evaluation method, can effectively model the nonlinear relationship between various factors in the complex system due to its strong autonomous learning and adaptive ability. However, although neural networks have been widely used in image recognition, natural language processing and other fields, their application in typical task system performance evaluation is still limited, and existing methods mostly use shallow neural networks, which have weak feature extraction ability for high-dimensional complex features and are difficult to capture deep information hidden in the running process of the task system. SUMMARY

[0015] To solve the above problems, the present application provides a typical task system performance evaluation method based on entropy weight method and graph convolution network. The evaluation method can fully utilize the characteristics of deep graph convolution network, efficiently extract deep information from task system running data, and improve the accuracy and stability of the evaluation result.

[0016] To achieve the above-mentioned purposes, the technical solutions provided by the present application are as follows: A typical task system performance evaluation method based on entropy weight method and graph convolution network, first, the entropy weight method is used to weight the evaluation index data with hierarchical structure, and the traditional hierarchical index data is converted into a data structure in non-Euclidean space, then a deep graph convolution network model is constructed to extract deep features in the input sample, and the coupling relationship between the sensor data and the task system performance is deeply mined, and finally the performance of the task system is evaluated.

[0017] Specifically, the following steps are included: Step S1. Construction of typical task system index system; The construction of the performance index system of the typical task system is divided into three layers: the top layer is the overall performance index, which is used to reflect the overall performance of the task system; the middle layer is the functional performance index, which is calculated according to the basic index and can describe the performance of the device function; the bottom layer is the basic index, which covers the data directly measured by the sensor, and is the basis of the evaluation system.

[0018] Step S2. Construction of input sample based on entropy weight method; The elements of the original one-dimensional feature vector are mapped onto the nodes of the graph according to the typical task system evaluation index system structure, so as to establish a node feature matrix, and an adjacency matrix is constructed according to the connection relationship in the evaluation index system.

[0019] The bottom layer features are weighted by using the entropy weight method to obtain intermediate feature values.

[0020] Further, the specific process of the entropy weight method is as follows: First, an evaluation matrix is constructed, assuming that there are evaluation objects, evaluation indexes, and the evaluation matrix is constructed , wherein represents the value of the evaluation object on the index; Then, the evaluation matrix is standardized; Further, the standardization method is range standardization: (1) , wherein and are the minimum value and the maximum value of the index, respectively; Next, the information entropy of each index is calculated. First, the proportion of the evaluation object on the index is calculated: (2) Then, the information entropy of the index is calculated: (3) , wherein is the number of evaluation objects, is a logarithmic function used to ensure that the value of the entropy is between ; Next, the entropy weight is calculated. The weight of the index is calculated according to the information entropy : (4) , wherein is the number of evaluation indexes.

[0021] Finally, the standardized values of each index are weighted and summed using the calculated weights to obtain the comprehensive score of each evaluation object: (5) , wherein, For the first The overall score of each evaluated object.

[0022] The entropy weight method described above can be used to calculate the complete system architecture diagram of the performance evaluation index for typical task systems, as well as the initial weights obtained by the entropy weight method.

[0023] Step S3. Performance evaluation of typical task systems based on graph convolutional networks; (1) Graph Convolutional Network For graph networks, the input is a graph. , Represented as ,in This represents the basic single-item indicators and the intermediate-level and top-level indicators calculated using the entropy weight method; This represents the weight of each indicator calculated using the entropy weight method.

[0024] Specifically, for node characteristics, each node It may contain feature vectors , representing the attributes or states of a node, node feature matrix It can be represented as ,in It is the number of nodes. This represents the number of features for each node; since each node represents a fixed performance metric, therefore... It is always 1, while for the adjacency matrix Used to represent the structure of a graph, where the adjacency matrix A In Represents a node and The connection relationship between nodes, if there are edges between nodes, then The weight of the corresponding connection; if it does not exist, then .

[0025] Furthermore, the computation process of graph convolutional networks integrates graph structure information into node representations layer by layer through multiple aggregation and update operations. Each aggregation operation aggregates the local structural information of a node into that node, while the update operation updates the node representation through linear transformations and nonlinear activations.

[0026] Aggregation operations are used to collect information from a node's neighbors. Neighbor aggregation is performed using the following formula: (6) in, Indicates the first Nodes in a layered network Embedded representation, Represents a node The set of neighboring nodes, The aggregation function is represented by the maximum relative graph convolution method, and the node feature aggregation and update operations adopt the maximum relative graph convolution method. The aggregation process can then be re-expressed as follows: (7) The MAX(.) operation selects the maximum eigenvalue of each neighbor node in each dimension element-wise. Specifically, if... It is the first Neighbor nodes in the layer Feature representation, assuming there are The first feature dimension, then the second feature dimension... Layer nodes The 3D features Represented as: (8) Represents a node In the The first layer Dimensional features.

[0027] After the aggregation operation, the feature representation of the node needs to be updated. The update operation uses the following formula: (9) in, It is the first The weight matrix of the layer, It is a bias term. It is a non-linear activation function. It is a node in the upper level. The expression .

[0028] Use the following formula to uniformly express neighbor aggregation and node update operations: (10) in, It is an adjacency matrix with self-loops added; yes The degree matrix, where ; It is the first Feature matrix of layer nodes It is the first Layer weight matrix.

[0029] Further, and in order to improve the further extraction ability of the model to the features, a feed-forward network (FFN) is added after the graph convolution module, the feed-forward network is composed of two fully connected layers and a residual layer, and an activation function is used to increase its nonlinear expression ability, and the calculation process is as shown in the following formula: (11) wherein, represents the output feature after the feed-forward network calculation, and are the learnable weights of the two fully connected layers of the feed-forward network, is a nonlinear activation function, is a bias term, and FFN(.) represents the feed-forward network calculation; The calculation process of the graph convolution module can be represented as follows: (12) wherein, GraphBlock(.) represents the graph convolution module calculation, represents the module output, and Grapher(.) represents the graph convolution calculation.

[0030] (2) Graph attention mechanism The graph attention mechanism first performs linear transformation on the feature vector of each node to generate the representation of the node. Let the transformation matrix be W , then the transformed node representation is: (13) wherein, is the feature vector of each node , , is the feature dimension, and is the new feature dimension.

[0031] Then the attention coefficient is calculated, for each pair of adjacent nodes and , the attention coefficient between them is calculated, which reflects the attention degree of node to the information of node . First, the unnormalized attention coefficient is calculated: (14) wherein, is a learned weight vector, represents the vector connection operation, and LeakyReLU is a ReLU activation function with leakage; Then the softmax function is used to normalize the attention coefficients: (15) where denotes the set of neighbor nodes of node . The feature representation of the node is then updated by weighting and summing the features of the neighbor nodes using the computed attention coefficients: (16) where is a nonlinear activation function (e.g., ReLU).

[0032] Furthermore, a multi-head attention mechanism is used, with separate attention heads applied simultaneously, and their outputs concatenated (or averaged) at the end: (17) In summary, the graph attention network can be written as follows: (18) (3) Performance evaluation method of typical task system based on graph convolutional network The performance evaluation method of the typical task system consists of three parts: the first part is feature extraction based on cascaded graph convolutional network, the second part is feature fusion and enhancement based on graph attention network, and the third part is performance prediction of feature mapping output: First, the input graph is input into the network. The network first uses four identical graph convolutional modules to extract features from the input information. The feature extraction module based on the graph network can be represented by the following equation: (19) where is the output of the th graph convolutional module, , denotes that the input of this graph convolutional module is the output of the previous graph convolutional module, denotes the original input graph; Then the model cascades the outputs of the four graph convolutional modules to obtain a new feature graph , which is represented as: (20) The graph attention network is used to further fuse and extract features from the feature graph, obtaining the final feature information : (21)​ wherein, is the feature vector of each node ; is the attention coefficient.

[0033] Then the features are processed by using the feature concatenation method, and the feature vector of each node is concatenated into a vector, wherein is the sum of the lengths of all node feature vectors, and the specific process can be represented as: (22) wherein, is the vector of each node of the new feature map output by the graph attention network; is the new feature vector after concatenation.

[0034] Finally, the features are mapped by a multilayer perceptron to obtain the final performance evaluation result, specifically, assuming that the MLP model has L layers, the output of the layer can be represented as: (23) wherein, is the original input, is the weight matrix of the layer, is the bias vector, is the output of the layer, and s is the activation function.

[0035] And the output dimension of the last layer of the multilayer perceptron should be 1, that is, the final performance evaluation result needs to be obtained.

[0036] In the training stage, the existing data in the database, including various indicators and evaluation values, are used to train the model, and the evaluation values come from the results of historical expert evaluation.

[0037] The advantages of the present application are: 1. By analyzing various indicators of a typical task system, a performance evaluation index system of the typical task system is constructed, and the entropy weight method and the evaluation index system are used to construct the index features into samples suitable for the input of the graph network, and the graph convolution network is used to quickly and efficiently evaluate the performance of the typical task system.

[0038] 2、Step 1, through the analysis of typical task system, the complex index is constructed into a hierarchical index system, which is convenient for subsequent performance evaluation of typical task system. Step 2, the intermediate layer and top layer indexes in the index system are calculated by entropy weight method, and the one-dimensional evaluation index is constructed into a sample suitable for the input of graph network according to the typical task system evaluation index system. Step 3, the shallow and deep feature information of the input sample is extracted through the cascaded graph convolution network, and then the feature information at different levels is further enhanced and fused through the graph attention network, and finally the performance evaluation value is output through the feature cascade and multilayer perception, which has good rapidity and accuracy.

[0039] 3、The one-dimensional index feature is constructed into a hierarchical feature graph structure suitable for the input of graph network through entropy weight method, which preliminarily evaluates the performance of typical task system through information entropy, and can better handle the relationship between the bottom single items.

[0040] 4、Since the structure of the performance evaluation index system of the typical task system is highly consistent with the input of the graph network, the graph network is applied to the performance evaluation of the typical task system in the patent, the model can fully extract the shallow and deep feature information of the input sample, and the graph attention network is used to fuse and enhance the features, so that the model has higher robustness and accuracy. BRIEF DESCRIPTION OF DRAWINGS

[0041] Fig. 1 is a flow chart of the performance evaluation method of typical task system based on entropy weight method and graph convolution network.

[0042] Figure 2 Fig. 2 is a structure diagram of the performance evaluation index system of typical task system.

[0043] Figure 3 Fig. 3 is a flow chart of constructing input sample based on entropy weight method.

[0044] Figure 4 Fig. 4 is a flow chart of the performance evaluation method of typical task system based on graph network. DETAILED DESCRIPTION

[0045] In order to make the purpose, technical scheme and advantages of the present application clearer, the technical scheme in the embodiments of the present application will be described clearly and completely below in combination with the drawings in the embodiments of the present application. Obviously, the described embodiments are for explaining the present application but not limiting the present application. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor fall within the scope of protection of the present application.

[0046] The specific implementation method of the present application will be described below in combination with the drawings and examples, and the present application is not limited to the embodiments.

[0047] Example 1 A typical task system performance evaluation method based on entropy weight method and graph convolution network, first, the entropy weight method is used to weight the evaluation index data with hierarchical structure, and the traditional hierarchical index data is converted into non-Euclidean space data structure, which prepares for the subsequent graph convolution network input. Then, a deep graph convolution network model is constructed to effectively extract deep features in the input sample, and deeply mine the complex coupling relationship between sensor data and task system performance, and finally realize the accurate evaluation of the task system performance. This method can utilize the correlation characteristics of sensor data to provide more detailed and reliable analysis of the device state.

[0048] Specifically, the following steps are included: Step S1. Construction of typical task system index system; To realize the accurate evaluation of the performance of the typical task system, it is a key prerequisite to construct a reasonable evaluation index system. Since the functional performance of the typical task system is complex and diverse, it covers motor control, sensing and monitoring, fault diagnosis and other aspects, therefore, performance evaluation usually needs to rely on multiple basic indexes. The basic single index is used to evaluate the performance of the task system under specific performance requirements, and these indexes are the basic components of performance evaluation, which can usually be obtained by sensor data acquisition and calculation. The complex index system will increase the difficulty of evaluation, therefore, before evaluation, the basic single index needs to be reasonably refined and classified according to the structural characteristics of the task system.

[0049] The construction of the performance index system of the typical task system is divided into three layers: the top layer is the overall performance index, which is used to reflect the overall performance of the task system; the middle layer is the functional performance index, which is calculated according to the basic index, and can describe the performance of the main function of the device; the bottom layer is the basic index, which covers various data measured directly by the sensor, and is the basis of the evaluation system. This hierarchical structure can effectively simplify the evaluation process and improve the accuracy and reliability of the evaluation.

[0050] Step S2. Construction of input sample based on entropy weight method; The input of the graph convolution network is a node feature matrix and an adjacency matrix, while the features for evaluating the performance of a typical task system are one-dimensional vectors, which cannot be directly input into the graph convolution network. Therefore, the input data needs to be transformed before building the evaluation model, and samples suitable for the input of the graph convolution network are constructed. The performance evaluation index system of the typical task system constructed in the previous step has a good hierarchical relationship and is highly consistent with the input structure of the graph network. Therefore, the elements of the original one-dimensional feature vector can be mapped to the nodes of the graph according to the structure of the typical task system evaluation index system, to establish a node feature matrix, and an adjacency matrix can be constructed according to the connection relationship in the evaluation index system.

[0051] The underlying basic single indicators have been calculated by formula, while the intermediate features (sub-indicator performance) and the top-level features (typical task system performance) need to be further calculated from the basic single indicators. Therefore, the entropy weight method is used to weight the bottom-level features to obtain the intermediate feature values.

[0052] The entropy weight method (EWM) is an objective weighting method used to determine the weights of each evaluation index in comprehensive evaluation. This method is based on the concept of information entropy and uses the amount of information of each evaluation index to determine the weight. The greater the information entropy, the higher the uncertainty of the index and the less the amount of information, so the weight is lower; on the contrary, the smaller the information entropy, the greater the weight. The core idea of the entropy weight method is to reflect the importance of each index to the comprehensive evaluation through its dispersion. Specifically, an index with greater dispersion provides more information, and the weight should be larger; an index with smaller dispersion provides less information, and the weight should be smaller.

[0053] Further, the specific process of the entropy weight method is as follows: First, an evaluation matrix is constructed, assuming that there are evaluation objects, evaluation indexes, and an evaluation matrix is constructed, where represents the value of the evaluation object on the index; Then, the evaluation matrix is standardized to eliminate the dimensional influence; Further, the standardization method is range standardization: (1) where and are the minimum and maximum values of the index, respectively; Then, the information entropy of the first index is calculated as follows: (2) (3) (4)

[0054] Finally, the comprehensive score of each evaluation object is obtained by weighting and summing the standardized values of each index using the calculated weights: (5)

[0055] After the above entropy weight method calculation, the complete typical task system performance evaluation index system structure diagram and the initial weight obtained by the entropy weight method can be obtained.

[0056] Step S3. Performance evaluation of typical task system based on graph convolutional network After constructing the sample suitable for the input of the graph convolutional network, the graph convolutional network is constructed to evaluate the performance of the typical task system. Graph Convolutional Network (GCN) is a class of deep learning models for processing graph structured data, especially suitable for non-Euclidean space data. Unlike traditional Convolutional Neural Network (CNN), the convolution operation of GCN is defined on the graph through adjacency relationship. GCN can be applied to social networks, molecular structures, bioinformatics, etc.

[0057] (1) Graph Convolutional Network For a graph network, its input is a graph A graph is usually represented as where ​​​​​​​​​​​​​​​​​​​is a set of nodes or vertices, representing points in the graph, in this patent these points are the basic single indicators and the intermediate layer indicators and top layer indicators calculated by entropy weight method; is a set of edges, representing the connection relationship between nodes, in this patent it is the weight of each indicator calculated by entropy weight method.

[0058] Specifically, for node features, each node may have a feature vector , representing the attributes or states of the node, the node feature matrix can be represented as , where is the number of nodes, is the number of features of each node; in this patent, since each node is a fixed performance indicator, is always 1, and for the adjacency matrix is usually used to represent the structure of the graph, where A in the adjacency matrix represents the connection relationship between nodes and , if there is an edge between the nodes, is the weight of the corresponding connection; if not, .

[0059] Further, the calculation process of the graph convolution network is to integrate the graph structure information into the node representation layer by layer through multi-layer aggregation and update operations, and each layer of aggregation operation aggregates the local structure information (i.e. neighbor information) of the node into the node, and the update operation updates the representation of the node through linear transformation and nonlinear activation.

[0060] The aggregation operation is used to collect information from the neighbors of the node, and for neighbor aggregation of node , the following formula is used: (6) , where represents the embedding representation of node in the layer network, represents the neighbor node set of node , and represents the aggregation function, in order to simplify and improve the efficiency of graph convolution calculation, the node feature aggregation and update operation adopts the maximum relative graph convolution method, that is, by calculating the relative difference between the node and its neighbor nodes, and selecting the maximum value in these differences to update the node feature, then the aggregation process can be represented as follows: (7) where the MAX(.) operation selects the largest feature value of the neighbor nodes in each dimension element-wise, specifically, if is the feature representation of the neighbor nodes in the th layer, assuming there are feature dimensions, then the th feature of the node in the th layer is represented as: (8) represents the th feature of the node in the th layer, by selecting the maximum value of the th feature among all neighbor nodes as the th feature of the node , the max-pooling can extract the strongest part of the features among the neighbor nodes.

[0061] After the aggregation operation, the feature representation of the node needs to be updated, the update operation uses the following formula: (9) where is the weight matrix of the th layer, is the bias term, is a nonlinear activation function (such as ReLU), is the representation of the node in the previous layer.

[0062] Then the following formula is usually used in the graph convolution network to uniformly express the neighbor aggregation and node update operations: (10) where is the adjacency matrix plus the self-loop; is the degree matrix of , where ; is the feature matrix of the nodes in the th layer, is the weight matrix of the th layer.

[0063] Through this formula, the graph convolution network aggregates neighbor information to the node in each layer, and uses the weight matrix to linearly transform the features, and then performs nonlinear transformation through the activation function .

[0064] ​​Further, and in order to improve the further extraction ability of the model to the features, a feed-forward network (FFN) is added after the graph convolution module, the feed-forward network is composed of two fully connected layers and a residual layer, and an activation function is used to increase its nonlinear expression ability, and the calculation process is as follows: (11) wherein, represents the output feature after the feed-forward network calculation, and are the learnable weights of the two fully connected layers of the feed-forward network, is a nonlinear activation function, is a bias term, and FFN(.) represents the feed-forward network calculation. The calculation process of the graph convolution module can be represented as follows: (12) wherein, GraphBlock(.) represents the graph convolution module calculation, represents the module output, and Grapher(.) represents the graph convolution calculation.

[0065] (2) Graph attention mechanism The core idea of the graph attention network GAT is to calculate the attention weight for each node according to the importance of its neighbor nodes, and then perform weighted summation on the features of the neighbor nodes to update the feature of the node. Unlike traditional graph convolution networks, GAT can adaptively learn the importance of neighbor nodes without relying on fixed graph structures.

[0066] The graph attention mechanism first performs a linear transformation on the feature vector of each node to generate the representation of the node. Let the transformation matrix be W , then the transformed node representation is: (13) wherein, is the feature vector of each node , , is the feature dimension, and the new feature dimension is

[0067] Then calculate the attention coefficient, for each pair of adjacent nodes and , , calculate the attention coefficient between them, which reflects the degree of attention of node to the information of node , first calculate the unnormalized attention coefficient: (14) where, is a learned weight vector, denotes vector concatenation operation, LeakyReLU is a ReLU activation function with a leak; Then, the softmax function is used to normalize to obtain the final attention coefficient: (15) where denotes the neighbor node set of node ; Next, the features of the neighbor nodes are weighted and summed using the calculated attention coefficient to update the feature representation of the node: (16) where is a nonlinear activation function (such as ReLU).

[0068] Furthermore, to improve the expression ability of the model, a multi-head attention mechanism can be used, with independent attention heads applied simultaneously, and their outputs concatenated (or averaged) at the end: (17) In summary, the graph attention network can be written as follows: (18) (3) Performance evaluation method of typical task system based on graph convolutional network The performance evaluation method of the typical task system mainly consists of three parts: the first part is the feature extraction module based on cascaded graph convolutional network, the second part is the feature fusion enhancement module based on graph attention network, and the third part is the feature mapping output performance prediction module. Specifically, first, the model will construct the input graph and input it into the network. The network first uses four identical graph convolutional modules to extract features from the input information. As the network deepens, each module will output different feature information. The low-level modules can capture local and fine-grained features, while the higher-level modules can capture more global and abstract features. The feature extraction module based on graph network can be represented by the following equation: (19) where, is the output of the th graph convolutional module, , The input of this graph convolution module is the output of the previous graph convolution module, The original input graph is represented as: The model then concatenates the outputs of the four graph convolution modules to obtain a new feature map, which contains both fine-grained shallow features and abstract deep features. The new feature map is represented as: (20) The new feature map contains a large amount of feature information, and the graph attention network is used to further fuse and extract features from the feature map to obtain the final feature information : (21) where, is the feature vector of each node ; is the attention coefficient.

[0069] Then, the features are processed using the feature concatenation method to facilitate the subsequent input of the multilayer perceptron. Specifically, the feature vector of each node is concatenated into a vector, where is the sum of the lengths of all node feature vectors, and the specific process can be represented as: (22) where, is the new feature map output by the graph attention network of each node; is the new feature vector after concatenation.

[0070] Finally, a multilayer perceptron (MLP) is used to map the features to obtain the final performance evaluation results. The MLP model is usually composed of multiple fully connected layers, each of which can perform feature transformation and mapping on the feature vector fused in step one, and then output to the next layer of neurons for processing.

[0071] Specifically, assuming that the MLP model has L layers, the output of the layer can be represented as: (23) where, is the original input, is the weight matrix of the layer, is the bias vector, is the output of the layer, and s is the activation function.

[0072] And the output dimension of the last layer of the multi-layer perception should be 1, that is, the performance evaluation result finally needs to be obtained, and the performance evaluation result of the typical task system can be obtained through the above steps.

[0073] In the training stage, the existing data in the database, including various indicators and evaluation values, are used to train the model, and the evaluation values come from the results of historical expert evaluation.

[0074] Embodiment 2 The typical task system performance evaluation method based on entropy weight method and graph convolution network mainly includes two parts. First, the input sample construction based on entropy weight method is performed to convert the index data with obvious hierarchical relationship into non-Euclidean data, so as to be used as the input of the graph convolution network. Then, the deep graph convolution network is constructed to extract the deep feature information in the input sample, mine the correlation and coupling relationship between the sensor data and the performance of the typical task system, and finally realize the evaluation of the performance of the typical task system. The implementation mode is as shown in Figure 1 .

[0075] The specific steps are as follows: (1) Constructing the index system of the typical task system First, the basic single index of the typical task system is classified according to the structure of the typical task system. At present, the construction of the performance index system of the typical task system mainly includes three layers, which are the top layer, i.e. the performance of the typical task system, the middle layer, i.e. the performance evaluation index of the main function of the typical task system obtained from the basic index, and the bottom layer, i.e. the basic index. The typical task system evaluation index system structure diagram is as shown in Figure 2 . The performance index should be modified according to the actual structure of the typical task system. The construction of the performance evaluation index system of the typical task system will help to clarify whether the evaluation performance is comprehensive and perfect and its credibility, and will help the subsequent step to convert the evaluation index into the graph structure data required by the graph network input.

[0076] (2) Input sample construction based on entropy weight method Then the one-dimensional evaluation index needs to be converted into graph structure data. For the middle layer index and top layer index of the above typical task system performance evaluation index system, the entropy weight method is used to initialize the unknown index value. Then the index value is mapped into the corresponding input matrix point by point according to the index system structure, and the specific process is as shown in Figure 3 .

[0077] The entropy weight method process is as follows: first, an evaluation matrix is constructed, assuming that there are evaluation objects and evaluation indexes, and the evaluation matrix is constructed , wherein Indicates the first The evaluation object is in the first The values ​​for each indicator.

[0078] The evaluation matrix is ​​then standardized to eliminate the influence of dimensions. A commonly used standardization method is range standardization. (1) Among them, among them, and The first The minimum and maximum values ​​of each indicator.

[0079] Next, the information entropy of each indicator is calculated. First, the information entropy of the first indicator is calculated. The evaluation object is in the first The proportion of each indicator : (2) Then, calculate the first... Information entropy of each indicator : (3) in, For the number of evaluation objects, It is a logarithmic function, used to ensure that the value of entropy is within a certain range. between.

[0080] Next, the entropy weight is calculated based on the information entropy. Calculate the first Weight of each indicator : (4) in, The number of evaluation indicators.

[0081] Finally, the standardized values ​​of each indicator are weighted and summed using the calculated weights to obtain the comprehensive score for each evaluation object: (5) in, For the first The overall score of each evaluated object.

[0082] The entropy weight method described above yields a complete architecture diagram of the performance evaluation index for a typical task system, along with its initial weights. Then, based on the index architecture, the index values ​​and their relationships are mapped to the feature matrix required for the graph network input. and adjacency matrix In this way, the input samples of the graph convolutional network can be obtained. .

[0083] (3) Performance evaluation of typical task systems based on graph convolutional networks The typical task system performance evaluation method in this section is mainly divided into three parts: a feature extraction module based on cascaded graph convolutional networks, a feature fusion and enhancement module based on graph attention networks, and a feature mapping output performance prediction module. The specific structure of the method is as follows: Figure 4 As shown. Specifically, the model first constructs the input graph. The input network first uses four identical graph convolutional modules to extract features from the input information. As the network deepens, each module outputs different feature information. Lower-level modules can capture local, fine-grained features, while higher-level modules can capture more global, abstract features. Therefore, the feature extraction module based on the graph network can be represented by the following equation: (19) in, For the first The output of the graph convolution module, , This indicates that the input to this graph convolution module is the output of the previous graph convolution module. This represents the original input image.

[0084] The model then concatenates the outputs of the four graph convolutional modules to obtain a new feature map. This feature map contains both fine-grained shallow features and abstract deep features. Represented as: (20) At this point, the new feature map Containing a large amount of feature information, a graph attention network is then used to further fuse and extract features from the feature map, yielding the final feature information. : (twenty one) in, It is each node eigenvectors; This represents the attention coefficient.

[0085] Next, this patent employs a feature concatenation method to process the features for subsequent input into a multilayer perceptron. Specifically, it requires concatenating the feature vectors of each node into a single 1-bit multilayer perceptron. The vector of L, here The sum of the lengths of the feature vectors of all nodes can be represented as follows: (twenty two) in, This is a new feature map output by a graph attention network. The vector of each node; It is the new feature vector after cascading.

[0086] Finally, an MLP is used to map the features to obtain the final performance evaluation result. The MLP model is usually composed of multiple fully connected layers. Each fully connected layer can transform and map the features obtained by fusing in step one, and then output them to the next layer of neurons for processing. Specifically, assuming our MLP model has L layers, then the... The output of the layer can be represented as: (twenty three) in, That is, the original input. It is the first The weight matrix of the layer, It is a bias vector. It is the first The output of the layer, This is the activation function.

[0087] Furthermore, the output dimension of the last layer of the multilayer perceptron should be 1, which is the final performance evaluation result to be obtained. Following the above steps, the performance evaluation result of a typical task system can be obtained.

Claims

1. A typical task system performance evaluation method based on entropy weight method and graph convolution network, characterized in that, The entropy weight method is used for weighted processing of evaluation index data with hierarchical structure, the hierarchical index data is converted into data structure in non-Euclidean space, a deep graph convolution network model is constructed, deep features in the input sample are extracted, coupling relationship between sensor data and task system performance is mined, and evaluation of the task system performance is realized.

2. The system performance evaluation method for typical tasks based on entropy weight method and graph convolution network according to claim 1, characterized in that, Specifically, the following steps are included: Step S1. Construction of typical task system index system; Through analysis of the typical task system, complex indexes are constructed into a hierarchical index system; the construction of the typical task system performance index system is divided into three layers: the top layer is the overall performance index; the middle layer is the functional performance index; and the bottom layer is the basic index; Step S2. Construction of input sample based on entropy weight method; The middle layer and top layer indexes in the index system are obtained through entropy weight method, and one-dimensional evaluation indexes are constructed into samples suitable for graph network input according to the typical task system evaluation index system; The elements of the original one-dimensional feature vector are mapped to the nodes of the graph according to the structure of the typical task system evaluation index system, so as to establish a node feature matrix, and an adjacency matrix is constructed according to the connection relationship in the evaluation index system; wherein the bottom layer features are weighted by the entropy weight method to obtain the middle feature values; Step S3. Performance evaluation of the typical task system based on graph convolution network; The shallow and deep feature information of the input sample is extracted through the cascaded graph convolution network, then the feature information at different levels is further enhanced and fused through the graph attention network, and finally the performance evaluation value is output through the feature concatenation and multilayer perceptron.

3. The system performance evaluation method for typical tasks based on entropy weight method and graph convolution network according to claim 2, characterized in that, In step S2, the specific process of the entropy weight method is as follows: First, construct the evaluation matrix, assuming there are evaluation objects, evaluation indexes, construct the evaluation matrix , where represents the value of the evaluation object on the index.

4. The system performance evaluation method for typical tasks based on entropy weight method and graph convolution network according to claim 3, characterized in that, The evaluation matrix is standardized by the range standardization method: （1） wherein, and are the minimum and maximum values of the first th index, respectively. Then, the information entropy of each index is calculated. First, the proportion of the first evaluation object on the first index is calculated : 0.2 . Then, the information entropy of each index is calculated. First, the proportion of the first evaluation object on the first index is calculated : 0.2 （2） Then, the information entropy of the first index is calculated : （3） wherein is the number of evaluation objects, is a logarithmic function to ensure that the value of the entropy is in between 0 and 1. Then the entropy weight is calculated according to the information entropy The weight of the first index is calculated : （4） wherein, the number of evaluation indices; Finally, the weights obtained by calculation are used to weight and sum the standardized values of each index to obtain the comprehensive score of each evaluation object: （5） wherein, is the overall score for the th evaluation object; The complete typical task system performance evaluation index system structure diagram and the initial weights obtained by the entropy weight method can be obtained through the above entropy weight method calculation.

5. The system performance evaluation method for typical tasks based on entropy weight method and graph convolution network according to claim 2, characterized in that, For the graph convolution network in step S3, the input is a graph , is expressed as wherein represents the basic single-item indicator and the intermediate layer indicator and the top layer indicator calculated by the entropy weight method; is expressed as the weight of each indicator calculated by the entropy weight method; For node characteristics, each node With feature vectors It represents the attributes or state of a node, the node feature matrix. Represented as ,in It is the number of nodes. This represents the number of features for each node; since each node represents a fixed performance metric, therefore... It is always 1, while for the adjacency matrix Used to represent the structure of a graph, where the adjacency matrix A In Represents a node and The connection relationship between nodes, if there are edges between nodes, then The weights of the corresponding connections; If not present, then .

6. The system performance evaluation method for typical tasks based on entropy weight method and graph convolution network according to claim 5, characterized in that, The calculation process of the graph convolution network is to integrate the graph structure information into the node representation layer by layer through the aggregation and update operations, and each layer of aggregation operation aggregates the local structure information of the node into the node, and the update operation updates the representation of the node through linear transformation and nonlinear activation; The aggregation operation is used to collect information from the node's neighbors, for the node neighbor aggregation, the following formula is used: （6） wherein, represents the i-th layer network in the embedding representation of the i-th node in the i-th layer network, represents the set of neighbor nodes of the i-th node in the i-th layer network, represents an aggregation function, and the node feature aggregation and update operation adopts the max relative graph convolution method, and the aggregation process can be re-represented as follows: （7） The MAX(.) operation selects the maximum eigenvalue of each neighbor node in each dimension element by element. It is the first Neighbor nodes in the layer Feature representation, assuming there are The first feature dimension, then the second feature dimension... Layer nodes The 3D features Represented as: （8） representing nodes In a first layer of the first dimensional features.

7. The system performance evaluation method for typical tasks based on entropy weight method and graph convolution network according to claim 6, characterized in that, After the aggregation operation, the feature representation of the node needs to be updated, and the update operation uses the following formula: （9） wherein, is the weight matrix of the layer, is the bias term, is the non-linear activation function, is the representation of the node in the previous layer; The following formula is used to uniformly express the neighbor aggregation and node update operation: （10） wherein, is the adjacency matrix plus self-loops; is the degree matrix of ; is the feature matrix of the nodes in the layer, is the weight matrix of the layer.

8. The system performance evaluation method for typical tasks based on entropy weight method and graph convolution network according to claim 7, characterized in that, And in order to improve the further extraction ability of the model to the features, a feedforward network is added after the graph convolution module, the feedforward network is composed of two fully connected layers and a residual layer, and an activation function is used to increase its nonlinear expression ability, and the calculation process is as follows: （11） wherein, represents the output features computed by the feed-forward network, and are the learnable weights of two fully connected layers of the feed-forward network, is a non-linear activation function, is a bias term, and FFN(.) represents the feed-forward network computation; Then the calculation process of the graph convolution module can be represented as follows: （12） wherein GraphBlock(.) represents a graph convolution module computation, denotes the module output, and Grapher(.) denotes a graph convolution computation.

9. The system performance evaluation method for typical tasks based on entropy weight method and graph convolution network according to claim 8, characterized in that, For the graph attention mechanism in step S3, first, the feature vector of each node is linearly transformed to generate the representation of the node; Let the transformation matrix be W The transformed node is represented as: （13） wherein, is a feature vector of each node , , is a feature dimension, is a new feature dimension; Then the attention coefficient is calculated, for each pair of adjacent nodes and , , the attention coefficient between them is calculated , which reflects the degree of attention of the node to the information of the node , first calculate the unnormalized attention coefficient: （14） wherein, is a learned weight vector, denotes a vector concatenation operation, and LeakyReLU is a ReLU activation function with a leak. The softmax function is then used to normalize the final attention coefficients: （15） wherein represents a set of neighbor nodes of the node ; Then the calculated attention coefficients are used to weight and sum the features of the neighbor nodes, so as to update the feature representation of the node: （16） wherein is a non-linear activation function.

10. The system performance evaluation method for typical tasks based on entropy weight method and graph convolution network according to claim 9, characterized in that, Using multi-head attention mechanism, while applying separate attention heads, and finally concatenating their outputs: （17） In summary, the graph attention network can be written as follows: （18）。 11. The system performance evaluation method for typical tasks based on entropy weight method and graph convolution network according to claim 10, characterized in that, The typical task system performance evaluation method is three parts, one is based on the feature extraction of cascade graph convolution network, the second part is based on the feature fusion enhancement of graph attention network, the third part is the feature mapping output performance prediction, as follows: Firstly, the input graph constructed by the model The input network firstly adopts four identical graph convolution modules to extract features from the input information. The feature extraction module based on the graph network can be represented by the following equation: （19） wherein, is the output of the th graph convolution module, , denotes that the input of this graph convolution module is the output of the previous graph convolution module, denotes the original input graph; The model then concatenates the outputs of the four graph convolution modules to obtain a new feature map, and the new feature map is represented as: （20） The graph attention network is used for further feature fusion and feature extraction on the feature map to obtain final feature information : （21） wherein, is a feature vector of each node ; is an attention coefficient.

12. The system performance evaluation method for typical tasks based on entropy weight method and graph convolution network according to claim 11, characterized in that, The features are processed by using a feature concatenation method, and a feature vector of each node is concatenated into a vector, where is the sum of lengths of feature vectors of all nodes, and the specific process can be represented as: （22） wherein, is a new feature map output by the graph attention network is a vector for each node of the graph; is the concatenated new feature vector.

13. The system performance evaluation method for typical tasks based on entropy weight method and graph convolution network according to claim 12, characterized in that, The final performance evaluation result is obtained by mapping the features through a multi-layer perception machine. Assuming that the MLP model has L layers, the output of the Lth layer can be represented as: ​ （23） where, , that is, the original input, is the weight matrix of the layer, is the bias vector, is the output of the layer, s is the activation function, and the output dimension of the last layer of the multi-layer perception should be 1, that is, the performance evaluation result finally needs to be obtained.

Images