Typical task system comprehensive performance evaluation method

Abstract

The invention relates to the technical field of task system performance evaluation and judgment, in particular to a typical task system comprehensive performance evaluation method based on weight analysis and a space-time diagram neural network, which comprises the following steps of: performing weight analysis on system state monitoring data by using a principal component analysis method, and weighting the data according to different importance degrees of different indexes; simultaneously extracting features of two dimensions of time and space by using a space-time diagram neural network, and generating a time feature matrix and a space feature matrix; fusing the time features and the space features by using a cross attention mechanism to generate a fused feature matrix; and predicting by using the fused feature matrix to obtain a comprehensive performance evaluation result of the system. According to the method, weight analysis based on principal component analysis is introduced, quantitative evaluation is performed on different performance indexes, the importance of each index is determined, and a corresponding weight is given, so that a key index can better guide the learning process of a model in subsequent time-space diagram neural network processing, and the performance of performance evaluation is improved.

Description

Technical Field

[0001] This invention relates to the field of task system performance evaluation and judgment technology, specifically to a comprehensive performance evaluation method for typical task systems based on weight analysis and spatiotemporal graph neural networks. Background Technology

[0002] A typical task system is a system that executes a series of predefined tasks in a specific application scenario. These tasks have defined operational logic and objectives. Typical task systems are widely used in fields such as industrial production, transportation, and energy dispatch. Performance evaluation of a typical task system requires analyzing key indicators such as system efficiency, stability, and resource utilization during task execution. The significance of performance evaluation lies in identifying potential problems through quantitative analysis of system operation, improving system operating efficiency, reducing maintenance costs, and ensuring the reliability and stability of the system in critical tasks.

[0003] Commonly used methods for performance evaluation of typical task systems include statistical analysis, signal processing, simulation testing, and traditional machine learning. Statistical analysis methods analyze system performance data based on statistics to evaluate system performance. This approach relies on large amounts of historical data and primarily targets known threats, lacking sensitivity and predictive ability for dynamic environments. Signal processing techniques utilize spectral analysis, filter design, and other methods to process and analyze system state signal data to evaluate system response. Signal processing is typically targeted at specific signal types and environments, lacking a global perspective and making it difficult to perform holistic evaluations of complex systems. Simulation testing constructs a virtual operating environment to simulate and test the system under various hypothetical scenarios. For example, Chinese invention patent application CN119513575A, entitled "A Method and System for Predicting Power Equipment Faults," discloses a method of multimodal fusion of acquired historical power equipment data to obtain feature vectors for each modality; fusion of these feature vectors using principal component analysis to obtain a fusion vector; construction of the fusion vector into a spatiotemporal graph; learning spatial information and extracting spatial features from the data using graph convolution to obtain a spatial feature representation vector; fusion of the spatial feature representation vector with attention weights to generate an attention fusion vector; learning temporal feature representation vectors using a gated recurrent unit; weighted feature fusion of the spatial and temporal feature representation vectors to obtain a spatiotemporal fusion representation vector; and processing the spatiotemporal fusion representation vector using a multilayer perceptron to output a power equipment fault prediction. However, simulation testing relies on preset environments and parameters, making it difficult to fully cover all possibilities in reality, and the realism of the simulation is limited by the accuracy of the modeling. When using traditional machine learning algorithms to evaluate the performance of a task system, feature extraction and classification of system state monitoring data are required to assess system performance. However, traditional machine learning-based methods for evaluating the overall performance of typical task systems often only consider the time or space dimension, ignoring the spatiotemporal relationship. Summary of the Invention

[0004] To address the aforementioned problems, this invention provides a comprehensive performance evaluation method for typical task systems based on weight analysis and spatiotemporal graph neural networks.

[0005] To achieve the above-mentioned objectives, the technical solution provided in this application is as follows: A typical method for comprehensive performance evaluation of a task system is proposed. Principal component analysis is used to perform weight analysis on system state monitoring data, and the data is weighted according to the different importance of different indicators. Spatiotemporal graph neural network is used to extract features in both time and space dimensions simultaneously, generating time feature matrices and space feature matrices. Cross-attention mechanism is used to fuse the time and space features to generate a fused feature matrix. The fused feature matrix is ​​then used for prediction to obtain the comprehensive performance evaluation result of the system.

[0006] Furthermore, the specific steps include the following: Step S1: First, collect data on relevant indicators of typical task systems, use them as input, and analyze the data using principal component analysis to obtain corresponding weights according to the different importance of the indicators. Step S2: Construct a time dataset and a graph dataset using the weighted data, and input them into a spatiotemporal graph neural network. Use the spatiotemporal graph neural network to process the time dataset and the spatial dataset. The weighted data is processed into a time dataset and a graph dataset to obtain a time feature matrix and a spatial feature matrix. Use a cross-attention mechanism to fuse the time features and spatial features. Step S3: Using the fused feature matrix as input, flatten the fused features, and then predict the fused feature matrix to obtain the final system performance evaluation result.

[0007] Furthermore, the specific steps in step S1 are as follows: The original data is a collection There are 1 time points, and each time point has 1 Feature data matrix Each row represents a sample, and each column represents a feature. First, standardize the data so that each feature has the same mean and variance: (1) (2) (3) in, It is the mean vector of the features. It is the standard deviation vector of the features. Let be the data matrix at time point i. The standardized data matrix is: .

[0008] Furthermore, after data standardization, the covariance matrix of the standardized data matrix is ​​calculated. : (4) For covariance matrix Eigenvalue decomposition yields eigenvalues ​​and eigenvectors: (5) in, It is an eigenvector matrix, where each column is an eigenvector; It is a diagonal matrix, where the elements on the diagonal are eigenvalues, representing the variance of the data explained by each principal component. The larger the eigenvalue, the larger the variance of the data explained by the principal component represented by the corresponding eigenvector. The principal components are sorted according to the size of their eigenvalues, and the top ones are selected. One principal component is used for dimensionality reduction: (6) It corresponds to the first Eigenvectors with large eigenvalues.

[0009] Furthermore, after selecting the principal components, for each feature, use its preceding... The weights are calculated by projecting each feature onto the principal components, and the cumulative contribution of each feature to these principal components is calculated: (7) It is the first The feature in the first The projections along the directions of the principal components, i.e., the eigenvector matrix. Element; These are the corresponding eigenvalues, representing the importance of the principal component; To ensure that the sum of the weight vectors is 1, the weight vectors are normalized: (8) Final weight vector Each component Indicates the first The importance of each original feature; Weights are directly assigned to the original data, and the input feature matrix is ​​multiplied element-wise by the weight vector to generate a weighted data matrix: (9) in, This indicates element-wise multiplication.

[0010] Furthermore, step S2 includes Step S21: Extraction of temporal features; First, a multi-scale temporal convolutional network is used to extract temporal features. By using different time windows for convolution, temporal features at multiple scales are extracted and finally combined into a temporal feature matrix. Step S22: Spatial feature extraction; After the temporal features are extracted, a graph convolutional network is used to extract the spatial features. Step S23: Spatiotemporal feature fusion; The spatiotemporal graph neural network extracts features from the data to obtain temporal and spatial feature matrices, and then uses cross-attention to fuse the spatiotemporal features.

[0011] Furthermore, the specific steps of step S21 are as follows: The input time dataset is a time series input matrix. ,in It is the number of time steps. It is the feature dimension at each time step. Multiple convolutional kernels with different time windows are used to convolve the input, each kernel having a different time window. Different convolution kernels The size of each convolutional kernel is , ,in It is the first The number of channels output by each convolution kernel, and the different convolution kernels Corresponding to different time windows It is used to capture features at different time scales.

[0012] Furthermore, for each convolutional kernel For the input matrix Perform a one-dimensional convolution operation, convolving along the time axis: (10) in, It is the first The output feature matrix of each convolutional kernel represents the temporal features within that time window. Indicates the time step. It is an activation function; For each time step convolution kernel The extracted features are represented as By using multiple convolutional kernels, feature representations at different time scales can be obtained; The features extracted from different time windows are concatenated together, and the output feature matrices of all convolutional kernels are concatenated along the channel dimension to obtain a comprehensive feature matrix: (11) in, It is the concatenated feature matrix.

[0013] To standardize the number of time steps in the feature matrix, pooling operations are used to downsample along the time dimension: (12) in, , is the time feature matrix after pooling; Furthermore, in order to process and integrate the time feature matrix, a linear transformation is applied to obtain the final time feature representation: (13) in, It is a learnable weight matrix used to transform the pooled features to the desired output dimension. , It is the final time feature matrix.

[0014] Furthermore, the steps for generating the spatial feature matrix using a graph convolutional network will be detailed below. Step S22 is as follows: The weighted data matrix The data is divided according to time, and the d data points at each time point form a graph. By connecting all the nodes in the graph, a data matrix can be constructed that contains... A graph dataset of graphs; The input to the graph convolutional network is as follows: Adjacency Matrix , used to represent the structure of a graph, where It is the number of nodes in the graph. Represents a node and nodes There are edges connecting them, and This indicates that no edge connection exists; Node feature matrix ,in It is the number of nodes in the graph; The adjacency matrix representation using symmetric normalization is as follows: (14) in, It is a node degree matrix, and its diagonal elements Represents a node The degree; It is an identity matrix, used to add self-loops to the adjacency matrix to ensure that each node contains its own information; Graph convolution is an operation that converts the feature matrices of nodes into graph convolutional functions. With the normalized adjacency matrix Convolution is performed, and a linear transformation is applied using a learnable weight matrix. The calculation process is as follows: (15) in, It is the first The input feature matrix of the layer, It is the first The feature dimensions of the layer It is the first The learnable weight matrix of the layer, It is a non-linear activation function. It is the feature matrix of the input nodes.

[0015] Furthermore, through multi-layer graph convolution operations, deeper spatial features are extracted layer by layer. After the final graph convolution, the final spatial feature matrix is ​​obtained: (16) in, It is the final spatial feature matrix. This is the feature dimension of the last convolutional layer, set here. ,pass Layered graph convolution extracts spatial features that combine neighborhood information and graph structure.

[0016] Furthermore, step S23 specifically includes: In the cross-attention mechanism, the temporal and spatial features first need to be converted into queries, keys, and values. Calculate the query matrix based on time features Bond matrix : (17) (18) in, and It is a learnable weight matrix. It is the dimension of keys and queries; Calculate the value matrix of spatial features : (19) in, It is a learnable weight matrix. It is the dimension of the value.

[0017] Furthermore, a query matrix using time features. Value matrix of spatial features To calculate the attention score: (20) in, It is an attention matrix. It is a scaling factor used to stabilize the gradient and control the range of the inner product value; Transpose the key matrix of time features Multiplying by the attention score matrix generates a new weighted combination representation of the temporal features: (twenty one) It is the fused temporal feature, and the feature of each time step is a weighted combination of spatial features; To further process the fused features A linear transformation is applied to obtain the final output features: (twenty two) in, It is a learnable weight matrix. It is the final output feature matrix.

[0018] Furthermore, step 3 specifically includes: Output feature matrix Flatten into a one-dimensional vector: (twenty three) The flattened vector is mapped to an output value through a fully connected layer: (twenty four) in, It is a weight matrix. It's a bias. It is an activation function; Data labels are created using prior system state monitoring data and performance evaluation results, and the network model is trained. After training, the test data is input into the network model, and the final output is the evaluation result of the comprehensive performance of the typical task system.

[0019] The advantages of this invention are: 1. This invention analyzes data from various indicators of a typical task system, uses principal component analysis (PCA) for weight analysis, then employs a spatiotemporal graph neural network to extract temporal and spatial features, and finally uses cross-attention to fuse the temporal and spatial features to obtain the final performance evaluation result. By introducing PCA-based weight analysis to quantitatively evaluate different performance indicators, determine the importance of each indicator and assign corresponding weights, key indicators can better guide the model's learning process in subsequent spatiotemporal graph neural network processing, thereby improving the performance evaluation.

[0020] 2. This invention uses a spatiotemporal graph neural network to model the performance index data of typical task systems. It uses a multi-scale temporal convolutional network to extract temporal information and a graph convolutional network to extract spatial features. By making full use of the spatial and temporal characteristics between the indicators, the model can capture potential performance patterns in complex data structures.

[0021] 3. This invention uses a cross-attention mechanism to fuse features, organically combining temporal and spatial features to capture the complex interaction relationships between various feature dimensions, thereby enhancing the model's ability to express high-dimensional data and obtaining more accurate performance evaluation results. Attached Figure Description

[0022] Figure 1 This is a flowchart for evaluating the comprehensive performance of a typical task system based on weight analysis and spatiotemporal graph neural networks.

[0023] Figure 2 This is a flowchart of weight analysis based on principal component analysis.

[0024] Figure 3 This is for data feature extraction based on spatiotemporal graph neural networks.

[0025] Figure 4 This is a schematic diagram of a typical task system comprehensive performance evaluation method based on spatiotemporal feature fusion. Detailed Implementation

[0026] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of the embodiments of this invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are for explaining the invention and not for limiting it. All other embodiments obtained by those skilled in the art based on the embodiments of this invention without creative effort are within the scope of protection of this invention.

[0027] The specific implementation method of the present invention will be described below with reference to the accompanying drawings and examples. The present invention is not limited to this embodiment.

[0028] Example 1 A typical method for comprehensive performance evaluation of a task system is proposed. Principal component analysis (PCA) is used to perform weighted analysis on system state monitoring data, weighting the data according to the importance of different indicators. A spatiotemporal graphical neural network (SPLM) is used to simultaneously extract features from both temporal and spatial dimensions, generating temporal and spatial feature matrices. A cross-attention mechanism is then used to fuse the temporal and spatial features, generating a fused feature matrix. Finally, the fused feature matrix is ​​used for prediction to obtain the comprehensive performance evaluation result of the system. Since task system state monitoring data has a high degree of spatiotemporal dependence, and SPLM can simultaneously consider the relationship between time and space dimensions, it provides more comprehensive evaluation results. Therefore, using SPLM for system performance evaluation can effectively address the shortcomings of current methods and provide more in-depth and comprehensive evaluation results.

[0029] Furthermore, the specific steps include the following: Step S1: First, collect data on relevant indicators of typical task systems, use them as input, and analyze the data using principal component analysis to obtain corresponding weights according to the different importance of the indicators. Step S2: Construct a time dataset and a graph dataset using the weighted data, and input them into a spatiotemporal graph neural network. Use the spatiotemporal graph neural network to process the time dataset and the spatial dataset. The weighted data is processed into a time dataset and a graph dataset to obtain a time feature matrix and a spatial feature matrix. Use a cross-attention mechanism to fuse the time features and spatial features. Step S3: Using the fused feature matrix as input, flatten the fused features, and then predict the fused feature matrix to obtain the final system performance evaluation result. The method block diagram of the comprehensive performance evaluation technique for typical task systems based on weight analysis and spatiotemporal graph neural networks is shown below. Figure 1 As shown.

[0030] Furthermore, in step S1, based on principal component analysis (PCA) for weight analysis, a comprehensive performance evaluation of a typical task system requires the collection of multiple relevant evaluation indicators. Different indicators have varying degrees of importance to system performance. The data is analyzed to determine the importance of each indicator, and corresponding weights are assigned for subsequent analysis. These weights reflect the importance of each indicator in the comprehensive performance evaluation, ensuring that key indicators have a greater impact on the model in subsequent analyses, thereby improving the accuracy and reliability of the evaluation results. The weight analysis process based on PCA is as follows: Figure 2 As shown, the input consists of data on multiple indicators of a typical task system. Principal component analysis is used for analysis, and the output is a weight vector, which reflects the importance of different indicators.

[0031] The steps of the weight analysis method based on principal component analysis will be described in detail below.

[0032] The original data is a collection There are 1 time points, and each time point has 1 Feature data matrix Each row represents a sample, and each column represents a feature. First, to ensure that the weight calculation of each feature is not affected by the original scale, the data needs to be standardized so that each feature has the same mean and variance: (1) (2) (3) in, It is the mean vector of the features. It is the standard deviation vector of the features. Let be the data matrix at time point i. The standardized data matrix is: .

[0033] Furthermore, after data standardization, the covariance matrix of the standardized data matrix is ​​calculated. The covariance matrix can reflect the correlation between various features: (4) For covariance matrix Eigenvalue decomposition yields eigenvalues ​​and eigenvectors: (5) in, It is an eigenvector matrix, where each column is an eigenvector; It is a diagonal matrix, where the elements on the diagonal are eigenvalues, representing the variance of the data explained by each principal component. The larger the eigenvalue, the larger the variance of the data explained by the principal component represented by the corresponding eigenvector. The principal components are sorted according to the size of their eigenvalues, and the top ones are selected. One principal component is used for dimensionality reduction: (6) It corresponds to the first Eigenvectors with large eigenvalues.

[0034] Furthermore, after selecting the principal components, for each feature, use its preceding... The weights are calculated by projecting each feature onto the principal components, and the cumulative contribution of each feature to these principal components is calculated: (7) It is the first The feature in the first The projections along the directions of the principal components, i.e., the eigenvector matrix. Element; These are the corresponding eigenvalues, representing the importance of the principal component; To ensure that the sum of the weight vectors is 1, the weight vectors are normalized: (8) Final weight vector Each component Indicates the first The importance of each original feature; Weights are directly assigned to the original data, and the input feature matrix is ​​multiplied element-wise by the weight vector to generate a weighted data matrix: (9) in, This indicates element-wise multiplication.

[0035] Furthermore, spatiotemporal graph neural networks (SPNNs) are deep learning models that combine graph-structured data with temporal information. They capture temporal features through multi-scale temporal convolutional networks; simultaneously, they construct a spatiotemporal graph, treating system performance metrics as nodes and relationships between nodes as edges, using graph convolution to fully capture the spatial features between these metrics. This method can efficiently integrate multi-dimensional spatiotemporal information, accurately identify potential patterns in complex systems, and lay a solid foundation for the accurate evaluation of the comprehensive performance of typical task systems. The data feature extraction process based on spatiotemporal graph neural networks is as follows: Figure 3 As shown, the input is the dataset after weight analysis. Using this dataset, a temporal dataset and a graph dataset are constructed. These datasets are then input into a multi-scale temporal convolutional network and a graph convolutional network, respectively, for feature extraction. Finally, a cross-attention mechanism is used to fuse the temporal and spatial features to obtain the final output features.

[0036] Step S2 includes Step S21: Extraction of temporal features; First, a multi-scale temporal convolutional network is used to extract temporal features. The multi-scale temporal model is a method used to capture short-term and long-term trends in time series data. By using different time windows for convolution, temporal features at multiple scales can be extracted and finally combined into a temporal feature matrix. This method can capture local and global patterns in the data and improve the model's ability to represent temporal features.

[0037] Step S22: Spatial feature extraction; After the temporal features are extracted, a graph convolutional network is used to extract the spatial features. Step S23: Spatiotemporal feature fusion; The spatiotemporal graph neural network extracts features from the data to obtain temporal and spatial feature matrices, and then uses cross-attention to fuse the spatiotemporal features.

[0038] Furthermore, the specific steps of step S21 are as follows: The input time dataset is a time series input matrix. ,in It refers to the number of time steps, i.e., the length of the time series. This refers to the feature dimension at each time step. To capture features at different time scales, multiple convolutional kernels with different time windows are used to convolve the input. Each convolutional kernel has a different time window. Different convolution kernels The size of each convolutional kernel is , ,in It is the first The number of channels output by each convolution kernel, and the different convolution kernels Corresponding to different time windows It is used to capture features at different time scales.

[0039] Furthermore, for each convolutional kernel For the input matrix Perform a one-dimensional convolution operation, convolving along the time axis: (10) in, It is the first The output feature matrix of each convolutional kernel represents the temporal features within that time window. Indicates the time step. It is an activation function used to introduce nonlinearity.

[0040] For each time step convolution kernel The extracted features are represented as By using multiple convolutional kernels, feature representations at different time scales can be obtained; The features extracted from different time windows are concatenated together, and the output feature matrices of all convolutional kernels are concatenated along the channel dimension to obtain a comprehensive feature matrix: (11) in, It is the concatenated feature matrix.

[0041] To standardize the number of time steps in the feature matrix, pooling operations are used to downsample along the time dimension: (12) in, , is the time feature matrix after pooling; Furthermore, in order to process and integrate the time feature matrix, a linear transformation is applied to obtain the final time feature representation: (13) in, It is a learnable weight matrix used to transform the pooled features to the desired output dimension. , It is the final time feature matrix.

[0042] Furthermore, the steps for generating the spatial feature matrix using a graph convolutional network will be detailed below. Step S22 is as follows: The weighted data matrix The data is divided according to time, and the d data points at each time point form a graph. By connecting all the nodes in the graph, a data matrix can be constructed that contains... A graph dataset of graphs; The input to the graph convolutional network is as follows: Adjacency Matrix , used to represent the structure of a graph, where It is the number of nodes in the graph. Represents a node and nodes There are edges connecting them, and This indicates that no edge connection exists; Node feature matrix ,in It is the number of nodes in the graph; To make graph convolution operations more stable and efficient, the adjacency matrix is ​​normalized. The symmetrically normalized adjacency matrix is ​​represented as follows: (14) in, It is a node degree matrix, and its diagonal elements Represents a node The degree; It is an identity matrix, used to add self-loops to the adjacency matrix to ensure that each node contains its own information; Graph convolution is an operation that converts the feature matrices of nodes into graph convolutional functions. With the normalized adjacency matrix Convolution is performed, and a linear transformation is applied using a learnable weight matrix. The calculation process is as follows: (15) in, It is the first The input feature matrix of the layer, It is the first The feature dimensions of the layer It is the first The learnable weight matrix of the layer, It is a non-linear activation function. It is the feature matrix of the input nodes.

[0043] Furthermore, through multi-layer graph convolution operations, deeper spatial features are extracted layer by layer. After the final graph convolution, the final spatial feature matrix is ​​obtained: (16) in, It is the final spatial feature matrix. This is the feature dimension of the last convolutional layer, set here. ,pass Layered graph convolution extracts spatial features that combine neighborhood information and graph structure.

[0044] Furthermore, step S23 specifically includes: In the cross-attention mechanism, the temporal and spatial features first need to be converted into queries, keys, and values. Calculate the query matrix based on time features Bond matrix : (17) (18) in, and It is a learnable weight matrix. It is the dimension of keys and queries; Calculate the value matrix of spatial features : (19) in, It is a learnable weight matrix. It is the dimension of the value.

[0045] Furthermore, a query matrix using time features. Value matrix of spatial features To calculate the attention score: (20) in, It is an attention matrix. It is a scaling factor used to stabilize the gradient and control the range of the inner product value; Transpose the key matrix of time features Multiplying by the attention score matrix generates a new weighted combination representation of the temporal features: (twenty one) It is the fused temporal feature, and the feature of each time step is a weighted combination of spatial features; To further process the fused features A linear transformation is applied to obtain the final output features: (twenty two) in, It is a learnable weight matrix. It is the final output feature matrix.

[0046] Furthermore, step 3 specifically involves: the process of a typical task system performance evaluation method based on spatiotemporal feature fusion as follows: Figure 4 As shown, the input is the output feature matrix of the spatiotemporal graph neural network, which is flattened into a one-dimensional vector and then mapped into an output value, which is the performance evaluation result.

[0047] Output feature matrix Flatten into a one-dimensional vector: (twenty three) The flattened vector is mapped to an output value through a fully connected layer: (twenty four) in, It is a weight matrix. It's a bias. It is an activation function; Data labels are created using prior system state monitoring data and performance evaluation results, and the network model is trained. After training, the test data is input into the network model, and the final output is the evaluation result of the comprehensive performance of the typical task system.

[0048] Example 2 The performance evaluation method for typical task systems based on weighted analysis and spatiotemporal graph convolutional networks mainly consists of three parts. First, principal component analysis is used to quantify and weight different performance indicators to ensure that key indicators are fully considered in the evaluation. Second, temporal and graph datasets are constructed, and spatial and temporal features are extracted using multi-scale temporal convolutional networks and graph convolutional networks, respectively. Finally, the spatiotemporal features are integrated through a cross-attention mechanism to generate a comprehensive and accurate evaluation result. The implementation method is as follows: Figure 1 As shown. The specific steps are as follows: (1) Weight analysis of index data of typical task system based on principal component analysis First, the data for multiple indicators in the original system are standardized. After standardization, the covariance matrix of the standardized data matrix is ​​calculated. And perform eigenvalue decomposition on it: (4) (5) in, It is an eigenvector matrix, where each column is an eigenvector; It is a diagonal matrix, and the elements on the diagonal are eigenvalues, representing the variance of the data explained by each principal component.

[0049] Sort the principal components by the magnitude of their eigenvalues ​​and select the top ones. One principal component is used for dimensionality reduction: (6) in, It corresponds to the first Eigenvectors with large eigenvalues.

[0050] After selecting the principal components, for each feature, use its preceding... The weights are calculated by projecting each feature onto the principal components, and the cumulative contribution of each feature to these principal components is calculated: (7) It is the first The feature in the first The projections along the directions of the principal components, i.e., the eigenvector matrix. Element; These are the corresponding eigenvalues, representing the importance of the principal component.

[0051] Weights are directly assigned to the original data, and the input feature matrix is ​​multiplied element-wise by the weight vector to generate a weighted data matrix: (9) in, This indicates element-wise multiplication.

[0052] The process of weighting index data for typical task systems based on principal component analysis is as follows: Figure 2 As shown, the original data underwent weight analysis using principal component analysis, assigning different weights to different indicators based on their importance, thus highlighting the importance of key indicators.

[0053] (2) Data feature extraction based on spatiotemporal graph convolutional network Data feature extraction based on spatiotemporal graph convolutional networks includes temporal feature extraction, spatial feature extraction, and spatiotemporal feature fusion, and its specific implementation is as follows.

[0054] To extract time features, a time dataset needs to be constructed. ,in It refers to the number of time steps, i.e., the length of the time series. This represents the feature dimension at each time step. The input is convolved using multiple convolutional kernels with different time windows. Each convolutional kernel has a different time window. Different convolution kernels The size of each convolutional kernel is , ,in It is the first The number of channels output by each convolution kernel, and the different convolution kernels Corresponding to different time windows It is used to capture features at different time scales.

[0055] For each convolution kernel For the input matrix Perform a one-dimensional convolution operation, convolving along the time axis: (10) in, It is the first The output feature matrix of each convolutional kernel represents the temporal features within that time window. Indicates the time step. It is an activation function used to introduce nonlinearity.

[0056] Features are extracted using a convolutional network with multiple windows, and the features extracted from different time windows are concatenated together to obtain a comprehensive feature matrix. (11) in, It is the concatenated feature matrix Use pooling operations to perform downsampling over time: (12) in, , is the time feature matrix after pooling.

[0057] The final time feature representation is obtained using a linear transformation: (13) in, It is a learnable weight matrix used to transform the pooled features to the desired output dimension. , It is the final time feature matrix.

[0058] For spatial feature extraction, first construct a graph dataset, and then weight the data matrix. The data is divided according to time, and the d data points at each time point form a graph. By connecting all the nodes in the graph, a data matrix can be constructed that contains... A graph dataset of graphs.

[0059] Subsequently, a graph convolutional network is used for feature extraction, and the adjacency matrix is ​​normalized. The symmetrically normalized adjacency matrix is ​​represented as follows: (14) in, It is a node degree matrix, and its diagonal elements Represents a node The degree; It is an identity matrix, used to add self-loops to the adjacency matrix to ensure that each node contains its own information.

[0060] The process of graph convolution calculation is as follows: (15) in, It is the first The input feature matrix of the layer, It is the first The feature dimensions of the layer It is the first The learnable weight matrix of the layer, It is a non-linear activation function. It is the feature matrix of the input nodes.

[0061] By performing multi-layer graph convolution operations, deeper spatial features are extracted layer by layer. After the final graph convolution, the final spatial feature matrix is ​​obtained. (16) in, It is the final spatial feature matrix. This refers to the feature degree of the last convolutional layer, which is set here. ,pass Layered graph convolution extracts spatial features that combine neighborhood information and graph structure.

[0062] After the spatiotemporal features are extracted, a cross-attention mechanism is used to fuse the spatiotemporal features.

[0063] Calculate the query matrix based on time features Bond matrix : (17) (18) in, and It is a learnable weight matrix. It is the dimension of keys and queries.

[0064] Calculate the value matrix of spatial features : (19) in, It is a learnable weight matrix. It is the dimension of the value.

[0065] Query matrix using time features Value matrix of spatial features To calculate the attention score: (20) in, It is an attention matrix. It is a scaling factor used to stabilize the gradient and control the range of inner product values.

[0066] Transpose the key matrix of time features Multiplying by the attention score matrix generates a new weighted combination representation of the temporal features: (twenty one) It is the fused temporal feature, and the feature of each time step is a weighted combination of spatial features.

[0067] To further process the fused features A linear transformation is applied to obtain the final output features: (twenty two) in, It is a learnable weight matrix. It is the final output feature matrix.

[0068] The process of data feature extraction is shown in the appendix. Figure 3 As shown, using the multi-indicator data of the system to be analyzed, a temporal dataset and a graph dataset are constructed. Multi-scale temporal convolutional networks and graph convolutional networks are used to extract the temporal and spatial features of the data, respectively. Then, a cross-attention mechanism is used to fuse the features to obtain the output features.

[0069] (3) Comprehensive performance evaluation of typical task systems The process of typical task system comprehensive performance evaluation method is attached. Figure 4As shown, the input is the output feature matrix of the spatiotemporal graph neural network, which is flattened into a one-dimensional vector and then mapped to an output value. The output value is the performance evaluation result.

[0070] Output feature matrix Flatten into a one-dimensional vector: (twenty three) The flattened vector is mapped to an output value through a fully connected layer: (twenty four) in, It is a weight matrix. It's a bias. It is an activation function Data is generated using existing prior system state monitoring data and performance evaluation results. Labels are assigned to each training sample, reflecting the sample's performance level. The network is trained using this labeled sample data, and the network parameters are optimized by minimizing the error between the predicted results and the actual labels. The trained network can generate predictions for new data based on the learned feature relationships, providing an evaluation of the overall system performance for typical tasks.

Claims

1. A typical task system comprehensive performance evaluation method, characterized in that, The principal component analysis method is used for weight analysis of system state monitoring data, and the data is weighted according to the importance of the indexes; the space-time graph neural network is used to extract the features in the time and space dimensions simultaneously, to generate a time feature matrix and a space feature matrix; the cross-attention mechanism is used to fuse the time features and the space features, to generate a feature matrix after fusion; the feature matrix after fusion is used for prediction, to obtain the result of system comprehensive performance evaluation.

2. The method of claim 1, wherein, Specifically, the following steps are included: Step S1, first, the data of the related indexes of the typical task system are collected as input, the principal component analysis method is used for data analysis, and the corresponding weights are obtained according to the different importance of the indexes; Step S2, the weighted data is used to construct a time data set and a graph data set, and is input into the space-time graph neural network, the space-time graph neural network is used to process the time data set and the space data set, the weighted data is processed into a time data set and a graph data set, a time feature matrix and a space feature matrix are obtained, and the cross-attention mechanism is used to fuse the time features and the space features; Step S3, the feature matrix after fusion is used as input, the features after fusion are flattened, and then the feature matrix after fusion is predicted, to obtain the final system performance evaluation result.

3. The method of claim 2, wherein, The specific steps in step S1 are as follows: The original data is a data matrix containing a time point, each time point having a feature , wherein each row represents a sample and each column represents a feature; First, the data is standardized, so that each feature has the same mean and variance: （1） （2） （3） wherein, is the mean vector of the features, is the standard deviation vector of the features, is the data matrix at the i-th time point, and the normalized data matrix is .

4. The typical task system integrated performance evaluation method according to claim 3, characterized in that, After data standardization, a covariance matrix of the standardized data matrix is calculated : （4） performing eigenvalue decomposition on the covariance matrix to obtain eigenvalues and eigenvectors: （5） wherein, is a feature vector matrix, each column is a feature vector; is a diagonal matrix, the elements on the diagonal are eigenvalues, indicating the data variance explained by each principal component, the larger the eigenvalue, the greater the data variance explained by the principal component represented by the corresponding feature vector, and the principal components are sorted according to the size of the eigenvalue, and the first principal components are selected for dimension reduction: （6） is the eigenvector corresponding to the largest eigenvalue.

5. The typical task system integrated performance evaluation method according to claim 4, characterized in that, After the principal components are selected, for each feature, its weight is computed using its projection in the first principal components, and the cumulative contribution of each feature in these principal components is computed: （7） is the projection of the i-th feature on the direction of the j-th principal component, i.e. the elements of the eigenvector matrix is the corresponding eigenvalue, indicating the importance of the principal component.​​​ In order to ensure that the sum of the weight vector is 1, the weight vector is normalized: （8） final weight vector each component of the vector represents the importance of the original feature The original data is directly assigned a weight, the input feature matrix is multiplied by the weight vector element by element, and a weighted data matrix is generated: （9） wherein represents an element-wise multiplication.

6. The method of claim 1, wherein, Step S2 includes Step S21, time feature extraction; First, the multi-scale time sequence convolutional network is used to extract the time features, different time windows are used for convolution, and multiple scales of time features are extracted, which are finally combined into a time feature matrix; Step S22, space feature extraction; After the time feature extraction is completed, the graph convolutional network is used to extract the space features; Step S23, space-time feature fusion; After the data is processed by the space-time graph neural network, the time feature matrix and the space feature matrix are obtained, and the cross-attention is used to fuse the space-time features.

7. The method of claim 6, wherein, The specific steps of step S21 are as follows: The input time data set is a time series input matrix wherein is the number of time steps, is the feature dimension of each time step, the input is convolved using multiple convolution kernels with different time windows, each convolution kernel has a different time window, using different convolution kernels each with a size of , wherein is the number of channels of the output of the th convolution kernel, different convolution kernels correspond to different time windows to capture features of different time scales.

8. The typical task system integrated performance evaluation method according to claim 7, characterized in that, For each convolution kernel , a one-dimensional convolution operation is performed on the input matrix , and the convolution is performed along the time axis: （10） wherein, is the output feature matrix of the th convolution kernel, representing the time feature under the time window, represents the time step, is an activation function; For each time step , the convolution kernel extracts the feature representation By using multiple convolution kernels, feature representations at different time scales can be obtained; The features extracted by different time windows are spliced together, the output feature matrices of all convolution kernels are spliced along the channel dimension, and a comprehensive feature matrix is obtained: （11） wherein, is the concatenated feature matrix; In order to unify the time step number of the feature matrix, a pooling operation is used for downsampling in the time dimension: （12） wherein, is the time feature matrix after pooling.

9. The typical task system integrated performance evaluation method according to claim 8, characterized in that, In order to process and integrate the time feature matrix, a linear transformation is applied to obtain the final time feature representation: （13） wherein, is a learnable weight matrix used to convert the pooled features to the desired output dimension , is the final temporal feature matrix.

10. The method of claim 9, wherein, The specific steps of step S22 are as follows: The weighted data matrix According to time, d data points at each time point constitute a graph, and all nodes in the graph are connected, and the data matrix can construct a graph data set containing graphs The input of the graph convolutional network is as follows: adjacency matrix , for representing the structure of a graph, where is the number of nodes in the graph, denotes that there is an edge connection between node and node , while denotes that there is no edge connection; node feature matrix wherein is the number of nodes in the graph; The symmetric normalized adjacency matrix is represented as: （14） wherein, is the node degree matrix, whose diagonal elements represent the degree of node ; is the identity matrix, used to add self-loops in the adjacency matrix, ensuring that each node contains information about itself; The operation of graph convolution is to perform convolution on the feature matrix of nodes with the normalized adjacency matrix and combine a linear transformation with a learnable weight matrix, whose calculation process is as follows: （15） wherein, is the input feature matrix of the layer, is the feature dimension of the layer, is the learnable weight matrix of the layer, is a non-linear activation function, is the input node feature matrix.

11. The typical task system integrated performance evaluation method according to claim 10, characterized in that, Through multi-layer graph convolution operations, deeper spatial features are extracted layer by layer, and after the last layer of graph convolution, the final spatial feature matrix is obtained: （16） wherein, is the final spatial feature matrix, is the feature dimension of the last graph convolution layer, which is set to , through layer graph convolution, the spatial features that integrate neighborhood information and graph structure are extracted.

12. The method of claim 11, wherein, The step S23 is specifically: In the cross-attention mechanism, the time and space features need to be converted into query, key and value first; Query matrix for time feature computation and key matrix : （17） （18） wherein, and are learnable weight matrices, is the dimension of the key and query. a matrix of spatial feature computed values : （19） wherein, is a learnable weight matrix, is the dimension of the values.

13. The typical task system integrated performance evaluation method according to claim 12, characterized in that, Query matrix using temporal features and value matrix of spatial features to compute attention scores: （20） wherein, is an attention matrix, is a scaling factor to stabilize the gradient and control the range of inner product values; transpose of the key matrix of temporal features multiplied with the attention score matrix, generating a new weighted combination representation of temporal features: （21） is the fused temporal feature, the feature of each time step is a weighted combination of spatial features; To further process the fused features , a linear transformation is applied to obtain the final output features: （22） wherein, is a learnable weight matrix, is the final output feature matrix.

14. The method of claim 13, wherein, The step 3 is specifically: The output feature matrix is flattened into a one-dimensional vector: （23） The flattened vector is mapped to an output value by a fully connected layer: （24） wherein, is a weight matrix, is a bias, is an activation function; The network model is trained using prior system state monitoring data and performance evaluation results to make data labels, and after training, the test data is input into the network model, and the final output is the evaluation result of the typical task system comprehensive performance.

Images