Power grid load situation awareness method and system based on load situation time sequence energy value

Abstract

This invention discloses a method and system for determining the type of power grid load situation awareness based on the energy value of load situation time series. The method includes: preprocessing the original collected load situation time series to obtain a standardized time series; performing CEEMDAN decomposition on the standardized time series to obtain multiple IMF components; extracting features from each IMF component and forming a comprehensive feature vector from all features; setting a set of situation awareness types and calculating the importance weight of each feature through an attention mechanism to generate attention-enhanced feature vectors; using a multi-level classification architecture to perform coarse-grained and fine-grained classification on the attention-enhanced feature vectors to obtain the probability that the feature vector belongs to a situation awareness type. The classification of power grid load situation awareness methods by the method provided by this invention enables power grid operation and maintenance and dispatching personnel to accurately grasp the load status, changing trends and potential risks from different perspectives, and thus formulate targeted control strategies.

Description

Technical Field

[0001] This invention belongs to the field of power load situation awareness technology, specifically relating to a method and system for power grid load situation awareness based on the energy value of load situation time series. Background Technology

[0002] With the increasing demands for real-time monitoring and intelligent decision-making in new power systems, traditional load situation awareness technologies face bottlenecks such as low data quality, incomplete feature representation, and insufficient classification accuracy. Existing methods, when processing load time series, often employ static threshold anomaly detection and simple interpolation strategies, which are ill-suited to the non-stationary nature of load fluctuations. Furthermore, feature extraction is often limited to single-scale statistics or frequency domain information, lacking in-depth analysis of multimodal energy distribution and complex time-series patterns. Simultaneously, the static weight allocation mechanism of traditional classification models fails to effectively distinguish between key situation features and redundant information, making classification results susceptible to noise interference.

[0003] In summary, overcoming the limitations of traditional single-classification models and providing key technical support for accurate perception and risk warning of power system operation status is an urgent problem to be solved. Summary of the Invention

[0004] Therefore, the purpose of this invention is to provide a method and system for sensing the power grid load situation based on the energy value of the load situation time series, so as to achieve more accurate and reliable sensing and classification of the power grid load situation.

[0005] The technical solution provided by this invention is as follows: Firstly, this invention provides a method for determining the type of power grid load situation perception based on the energy value of a time series load situation, comprising:

[0006] The original load situation time series was preprocessed to obtain a standardized time series;

[0007] The standardized time series was decomposed using CEEMDAN to obtain multiple IMF components;

[0008] Features are extracted from each IMF component, and the mean, variance, skewness, dominant frequency, multi-scale energy entropy, and deep temporal series features of each IMF are calculated. All features are then combined into a comprehensive feature vector.

[0009] Based on the comprehensive feature vector, a set of situational awareness types is set, and the importance weight of each feature is calculated through an attention mechanism to generate an attention-enhanced feature vector.

[0010] A multi-level classification architecture is adopted to perform coarse-grained and fine-grained classification on the attention-enhanced feature vectors to obtain the probability that the feature vectors belong to the situational awareness type.

[0011] The situational awareness type with the highest probability is selected as the final result.

[0012] Preferably, load status time series ,in The total length of the sequence. For the first The original energy value at each point in time;

[0013] Outliers in the original load situation time series were removed using a quantile-based dynamic thresholding method, resulting in a preliminarily cleaned series. ;

[0014] Linear interpolation is used to process the pre-cleaned sequence. After removing the missing values, the complete sequence was obtained. :

[0015]

[0016] In the formula, This represents the energy value calculated by interpolation at the missing position j. The nearest non-missing value preceding the missing position j; The nearest non-missing value after the missing position j; The index position of the missing point in the time series that needs to be interpolated;

[0017] Based on complete sequence Perform data normalization to obtain a normalized sequence. :

[0018]

[0019] In the formula, This is the normalized energy value at the i-th time point; Let be the energy value at the i-th time point in the complete sequence; The minimum energy value in the complete sequence; This represents the maximum energy value in the complete sequence.

[0020] Based on normalized sequences Perform data smoothing to obtain standardized time series. .

[0021] Further optimization involves removing outliers from the original load situation time series using a dynamic thresholding method based on quantiles, resulting in a preliminarily cleaned sequence. :

[0022] Calculate the first quartile of the load situation time series and the third and fourth quartiles ;

[0023] Dynamically calculate the upper and lower limits for outlier detection:

[0024]

[0025]

[0026]

[0027] In the formula, This is the lower threshold for outlier detection; This is the upper threshold for outlier detection; For dynamic adjustment coefficients, Load status time series The arithmetic mean; Load status time series Standard deviation

[0028] All sequences that satisfy or Data points were identified as outliers and removed, resulting in a pre-cleaned sequence. .

[0029] Further preferred, the normalized sequence-based Perform data smoothing to obtain standardized time series. ,include:

[0030]

[0031]

[0032]

[0033] In the formula, Let be the smoothed energy value at the i-th time point; The standard deviation of the Gaussian function; The position index within the sliding window; The optimal smoothing window size; is the first-order autocorrelation coefficient of the sequence.

[0034] Preferably, the standardized time series is subjected to CEEMDAN decomposition to obtain multiple IMF components, including:

[0035]

[0036] In the formula, Represents a standardized time series. The value of the m-th intrinsic mode function at time point t; The value of the residual component at time point t; The number of modes is automatically determined.

[0037] Preferably, the step involves extracting features from each IMF component and calculating the mean feature of each IMF:

[0038]

[0039] In the formula, The mean characteristic of the m-th IMF component;

[0040] Calculate the variance characteristics of each IMF:

[0041]

[0042] In the formula, Let be the variance characteristic of the m-th IMF component;

[0043] Calculate the skewness characteristics of each IMF:

[0044]

[0045] In the formula, For the first Skewness characteristics of each IMF component;

[0046] Calculate the dominant frequency characteristics of each IMF:

[0047]

[0048]

[0049] In the formula, Let be the frequency domain amplitude of the m-th IMF component; This is a frequency index, representing different frequency components; For the first The dominant frequency characteristics of each IMF component;

[0050] Calculate multi-scale energy entropy characteristics:

[0051]

[0052]

[0053]

[0054] In the formula, It is a multi-scale energy entropy characteristic; The energy of the m-th IMF; The sum of energy for all IMF components;

[0055] Perform deep feature learning and compute deep temporal features:

[0056]

[0057] In the formula, This represents a deep temporal feature vector. For parameters Temporal convolutional networks;

[0058] All features form a comprehensive feature vector. .

[0059] Preferably, the set of situational awareness types is set as follows: Where g is the number of situation types, the importance weight of each feature is calculated through an attention mechanism to generate the attention weight of each feature:

[0060]

[0061] In the formula, Let be the attention weight for the d-th feature; This is the transpose of the attention vector, obtained through training; The attention weight matrix is ​​obtained through training. For comprehensive feature vectors The d-th eigenvalue in; This is the attention bias term, obtained through training;

[0062] All based on calculations Construct the feature attention vector A:

[0063]

[0064] In the formula, For feature attention vectors;

[0065] All based on calculations Construct the feature attention vector A:

[0066]

[0067] In the formula, This is the feature attention vector.

[0068] Generate attention-enhanced feature vectors :

[0069]

[0070] In the formula, Feature vectors for attention enhancement.

[0071] Preferably, a multi-level classification architecture is used to perform coarse-grained classification on the attention-enhanced feature vectors to obtain a coarse-grained probability distribution:

[0072]

[0073]

[0074]

[0075] In the formula, This is the coarse-grained situational awareness classification score vector. This is a coarse-grained situational awareness classification weight matrix; This is the coarse-grained situational awareness classification bias vector. For the feature vector to belong to the th The probability of a coarse-grained situational awareness type;

[0076] Based on coarse-grained classification results and attention enhancement features Fine-grained classification yields the probability distribution:

[0077]

[0078]

[0079] In the formula, The score for the feature vector belonging to the g-th fine-grained situational awareness category; This is a fine-grained classification weight matrix; This is a fine-grained classification bias vector;

[0080] The probability distribution is obtained based on coarse-grained probability distribution and fine-grained classification, and the feature vector is determined to belong to the situational awareness type. The probability of:

[0081]

[0082] The feature vector belongs to the situational awareness type. The probability of.

[0083] Secondly, the present invention also provides a power grid load situation perception type determination system based on load situation time series energy values, comprising:

[0084] The data preprocessing module preprocesses the raw collected load situation time series to obtain a standardized time series;

[0085] The CEEMDAN decomposition module performs CEEMDAN decomposition on the standardized time series to obtain multiple IMF components;

[0086] The feature extraction module extracts features from each IMF component, calculates the mean, variance, skewness, dominant frequency, multi-scale energy entropy, and deep temporal series features of each IMF, and forms a comprehensive feature vector from all features.

[0087] The attention-enhanced feature vector generation module calculates the importance weight of each feature through an attention mechanism to generate attention-enhanced feature vectors.

[0088] The classification module adopts a multi-level classification architecture to perform coarse-grained and fine-grained classification on the attention-enhanced feature vectors, and obtain the probability that the feature vectors belong to the situational awareness type.

[0089] The results output module outputs the situational awareness type with the highest probability.

[0090] Preferably, the classification module includes: a coarse-grained classification module and a fine-grained classification module, wherein the coarse-grained classification module obtains a coarse-grained probability distribution; and the fine-grained classification module obtains a fine-grained probability distribution.

[0091] This invention provides a method and system for power grid load situation awareness based on load situation time series energy values. The method constructs a multi-scale feature extraction framework based on CEEMDAN decomposition, combining statistical features such as mean, variance, and skewness with composite features such as multi-scale energy entropy and deep time series features to form a high-dimensional feature vector. By introducing an attention mechanism to dynamically calculate feature importance weights, and in conjunction with a multi-level classification architecture, it achieves progressive identification from coarse-grained to fine-grained, significantly improving the accuracy of situation type discrimination.

[0092] The classification of power grid load status perception methods provided by this invention enables power grid operation and maintenance and dispatch personnel to accurately grasp the load status, changing trends and potential risks from different perspectives, and then formulate targeted control strategies. Attached Figure Description

[0093] The accompanying drawings, which are incorporated in and form part of this specification, illustrate embodiments consistent with the invention and, together with the description, serve to explain the principles of the invention.

[0094] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, for those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0095] Figure 1The flowchart of the power grid load situation perception method based on load situation time series energy value provided by the present invention. Detailed Implementation

[0096] The present invention will be further explained below with reference to specific implementation schemes, but this explanation does not limit the scope of the invention.

[0097] To achieve more accurate operational status perception and decision support for the power system, such as Figure 1 As shown, the power grid load situation perception method based on load situation time series energy values ​​includes:

[0098] S1: Remove outliers from the original load situation time series, fill in missing values ​​through linear interpolation, and perform data normalization and smoothing to obtain a standardized time series;

[0099] Based on the original collected load status time series (in The total length of the sequence. For the first The raw energy values ​​at each time point were used, and a dynamic threshold detection method based on quantiles was employed to remove outliers.

[0100] Calculate the first quartile of the load situation time series (25th percentile) and third quartile (75th percentile).

[0101] Dynamically calculate the upper and lower limits for outlier detection:

[0102]

[0103]

[0104]

[0105] In the formula, This is the lower threshold for outlier detection; This is the upper threshold for outlier detection; For dynamic adjustment coefficients, Load status time series The arithmetic mean; Load status time series Standard deviation

[0106] All sequences that satisfy or Data points were identified as outliers and removed, resulting in a pre-cleaned sequence. .

[0107] Based on the pre-cleaned sequence Linear interpolation was performed on the missing values ​​resulting from the removal process to obtain the complete sequence. :

[0108]

[0109] In the formula, This represents the energy value calculated by interpolation at the missing position j. The nearest non-missing value preceding the missing position j; The nearest non-missing value after the missing position j; This refers to the index position of the missing points in the time series that need to be interpolated.

[0110] Based on complete sequence Perform data normalization to obtain a normalized sequence. :

[0111]

[0112] In the formula, This is the normalized energy value at the i-th time point; Let be the energy value at the i-th time point in the complete sequence; The minimum energy value in the complete sequence; This represents the maximum energy value in the complete sequence.

[0113] Based on normalized sequences Perform data smoothing to obtain standardized time series. .

[0114]

[0115]

[0116]

[0117] In the formula, Let be the smoothed energy value at the i-th time point; The standard deviation of the Gaussian function; The position index within the sliding window; The optimal smoothing window size; is the first-order autocorrelation coefficient of the sequence.

[0118] By combining the aforementioned dynamic thresholding method based on quantiles with data smoothing methods for data preprocessing, data quality is effectively improved.

[0119] S2: Perform CEEMDAN decomposition on the standardized time series to obtain multiple IMF components. Extract features from each IMF component, including mean features, variance features, skewness features, main frequency features, multi-scale energy entropy features, and deep time series features, to form a comprehensive feature vector.

[0120] Standardized time series based on S1 output CEEMDAN decomposition was performed to obtain multiple IMF components. This is used for subsequent feature extraction.

[0121]

[0122] In the formula, The value of the m-th intrinsic mode function at time point t; The value of the residual component at time point t; The number of modes is automatically determined.

[0123] Features are extracted from each IMF component, and the mean feature of each IMF is calculated:

[0124]

[0125] In the formula, Let be the mean characteristic of the m-th IMF component.

[0126] Calculate the variance characteristics of each IMF:

[0127]

[0128] In the formula, Let be the variance characteristic of the m-th IMF component.

[0129] Calculate the skewness characteristics of each IMF:

[0130]

[0131] In the formula, For the first Skewness characteristics of each IMF component.

[0132] Calculate the dominant frequency characteristics of each IMF:

[0133]

[0134]

[0135] In the formula, Let be the frequency domain amplitude of the m-th IMF component; This is a frequency index, representing different frequency components; For the first The dominant frequency characteristics of each IMF component.

[0136] Calculate multi-scale energy entropy characteristics:

[0137]

[0138]

[0139]

[0140] In the formula, It is a multi-scale energy entropy characteristic; The energy of the m-th IMF; This is the sum of the energy of all IMF components.

[0141] Perform deep feature learning and compute deep temporal features:

[0142]

[0143] In the formula, This represents a deep temporal feature vector. For parameters Temporal convolutional networks.

[0144] All the above features are combined into a comprehensive feature vector. .

[0145] By employing CEEMDAN decomposition and combining it with multi-dimensional features (including statistical features, frequency domain features, energy entropy, and deep time series features), a comprehensive capture of multi-scale and deep-level information of load time series was achieved.

[0146] This implementation scheme also introduces an attention mechanism to weight key features and adopts a multi-level classification architecture for coarse-to-fine identification, which improves the accuracy and refinement of situation type judgment.

[0147] S3: Given a set of situational awareness types, calculate the importance weight of each feature through an attention mechanism to generate an attention-enhanced feature vector; adopt a multi-level classification architecture to perform coarse-grained and fine-grained classification to obtain the probability that a feature belongs to a situational awareness type.

[0148] Let the set of situational awareness types be... , where g is the number of situation types.

[0149] The importance weight of each feature is calculated using an attention mechanism, thus obtaining the attention weights for each feature. :

[0150]

[0151] In the formula, Let be the attention weight for the d-th feature; This is the transpose of the attention vector, obtained through training; The attention weight matrix is ​​obtained through training. For comprehensive feature vectors The d-th eigenvalue in; This is the attention bias term, obtained through training.

[0152] All based on calculations Construct the feature attention vector A:

[0153]

[0154] In the formula, This is the feature attention vector.

[0155] Generate attention-enhanced feature vectors :

[0156]

[0157] In the formula, Feature vectors for attention enhancement.

[0158] A multi-level classification architecture is adopted. First, coarse-grained classification is performed to obtain a coarse-grained probability distribution:

[0159]

[0160]

[0161]

[0162] In the formula, This is the coarse-grained situational awareness classification score vector. This is a coarse-grained situational awareness classification weight matrix; This is the coarse-grained situational awareness classification bias vector. For the feature vector to belong to the th The probability of each coarse-grained situational awareness category.

[0163] Based on coarse-grained classification results and attention enhancement features Fine-grained classification yields the probability distribution:

[0164]

[0165]

[0166] In the formula, The score for the feature vector belonging to the g-th fine-grained situational awareness category; This is a fine-grained classification weight matrix; This is a fine-grained classification bias vector; The feature vector belongs to the situational awareness type. The probability of.

[0167] S4: Select the situational awareness type with the highest probability as the final result.

[0168] Select the situational awareness type with the highest probability as the final result:

[0169]

[0170] In the formula, The final type of situational awareness.

[0171] In addition, this implementation plan also provides a power grid load situation awareness type determination system based on load situation time series energy values, including:

[0172] The data preprocessing module preprocesses the raw collected load situation time series to obtain a standardized time series;

[0173] The CEEMDAN decomposition module performs CEEMDAN decomposition on the standardized time series to obtain multiple IMF components;

[0174] The feature extraction module extracts features from each IMF component, calculates the mean, variance, skewness, dominant frequency, multi-scale energy entropy, and deep temporal series features of each IMF, and forms a comprehensive feature vector from all features.

[0175] The attention-enhanced feature vector generation module calculates the importance weight of each feature through an attention mechanism to generate attention-enhanced feature vectors.

[0176] The classification module adopts a multi-level classification architecture to perform coarse-grained and fine-grained classification on the attention-enhanced feature vectors, and obtain the probability that the feature vectors belong to the situational awareness type.

[0177] The results output module outputs the situational awareness type with the highest probability.

[0178] The classification module includes a coarse-grained classification module and a fine-grained classification module, wherein the coarse-grained classification module obtains a coarse-grained probability distribution, and the fine-grained classification module obtains a fine-grained probability distribution.

[0179] In summary, the perception type determination method in this implementation scheme constructs a highly robust feature vector by dynamically adjusting coefficients to eliminate outliers and combining multi-dimensional representations of statistical features, multi-scale energy entropy, and deep temporal features. An attention mechanism is introduced to allocate feature weights, coupled with a coarse-grained classification architecture, to improve the accuracy of load status identification, providing more precise operational status perception and decision support for the power system. This enables more accurate and reliable perception and classification of grid load status, providing strong technical support for the safe and stable operation and intelligent control of the power grid.

[0180] Although preferred embodiments of this application have been described, those skilled in the art, upon learning the basic inventive concept, can make other changes and modifications to these embodiments. Therefore, the appended claims are intended to be interpreted as including the preferred embodiments as well as all changes and modifications falling within the scope of this application.

[0181] Obviously, those skilled in the art can make various modifications and variations to this application without departing from the spirit and scope of this application. Therefore, if such modifications and variations fall within the scope of the claims of this application and their equivalents, this application also intends to include such modifications and variations.

Claims

1. A method for determining the type of power grid load situation perception based on the energy values ​​of load situation time series, characterized in that, include: The original load situation time series was preprocessed to obtain a standardized time series; The standardized time series was decomposed using CEEMDAN to obtain multiple IMF components; Features are extracted from each IMF component, and the mean, variance, skewness, dominant frequency, multi-scale energy entropy, and deep temporal series features of each IMF are calculated. All features are then combined into a comprehensive feature vector. Based on the comprehensive feature vector, a set of situational awareness types is set, and the importance weight of each feature is calculated through an attention mechanism to generate an attention-enhanced feature vector. A multi-level classification architecture is used to perform coarse-grained and fine-grained classification on the attention-enhanced feature vectors to obtain the probability that the feature vectors belong to the situational awareness type: In the formula, This is the coarse-grained situational awareness classification score vector. This is a coarse-grained situational awareness classification weight matrix; This is the coarse-grained situational awareness classification bias vector. For the feature vector to belong to the th The probability of a coarse-grained situational awareness type; Based on coarse-grained classification results and attention enhancement features Fine-grained classification is performed to obtain the probability distribution: In the formula, The score for the feature vector belonging to the g-th fine-grained situational awareness category; This is a fine-grained classification weight matrix; This is a fine-grained classification bias vector; The probability distribution is obtained based on coarse-grained probability distribution and fine-grained classification, and the feature vector is determined to belong to the situational awareness type. The probability of: The feature vector belongs to the situational awareness type. The probability of; The situational awareness type with the highest probability is selected as the final result.

2. The method for determining the power grid load situation perception type based on the original collected load situation time series energy values ​​according to claim 1, characterized in that, Load situation time series ,in The total length of the sequence. For the first The original energy value at each point in time; Outliers in the original load situation time series were removed using a quantile-based dynamic thresholding method, resulting in a preliminarily cleaned series. ; Linear interpolation is used to process the pre-cleaned sequence. After removing the missing values, the complete sequence was obtained. : In the formula, This represents the energy value calculated by interpolation at the missing position j. The nearest non-missing value preceding the missing position j; The nearest non-missing value after the missing position j; The index position of the missing point in the time series that needs to be interpolated; Based on complete sequence Perform data normalization to obtain a normalized sequence. : In the formula, This is the normalized energy value at the i-th time point; Let be the energy value at the i-th time point in the complete sequence; The minimum energy value in the complete sequence; This represents the maximum energy value in the complete sequence. Based on normalized sequences Perform data smoothing to obtain standardized time series. .

3. The method for determining the type of power grid load situation perception based on the energy value of load situation time series according to claim 2, characterized in that, Outliers in the original load situation time series were removed using a quantile-based dynamic thresholding method, resulting in a preliminarily cleaned series. : Calculate the first quartile of the load situation time series and the third and fourth quartiles ; Dynamically calculate the upper and lower limits for outlier detection: In the formula, This is the lower threshold for outlier detection; This is the upper threshold for outlier detection; For dynamic adjustment coefficients, Load status time series The arithmetic mean; Load status time series Standard deviation All sequences that satisfy or Data points were identified as outliers and removed, resulting in a pre-cleaned sequence. .

4. The method for determining the type of power grid load situation perception based on the energy value of load situation time series according to claim 2, characterized in that, The normalized sequence Perform data smoothing to obtain standardized time series. ,include: In the formula, Let be the smoothed energy value at the i-th time point; The standard deviation of the Gaussian function; The position index within the sliding window; The optimal smoothing window size; is the first-order autocorrelation coefficient of the sequence.

5. The method for determining the type of power grid load situation perception based on the energy value of load situation time series according to claim 1, characterized in that, The standardized time series was decomposed using CEEMDAN to obtain multiple IMF components, including: In the formula, Represents a standardized time series. The value of the m-th intrinsic mode function at time point t; The value of the residual component at time point t; The number of modes is automatically determined.

6. The method for determining the type of power grid load situation perception based on the energy value of load situation time series according to claim 1, characterized in that, The process involves extracting features from each IMF component and calculating the mean feature of each IMF: In the formula, The mean characteristic of the m-th IMF component; Calculate the variance characteristics of each IMF: In the formula, Let be the variance characteristic of the m-th IMF component; Calculate the skewness characteristics of each IMF: In the formula, For the first Skewness characteristics of each IMF component; Calculate the dominant frequency characteristics of each IMF: In the formula, Let be the frequency domain amplitude of the m-th IMF component; This is a frequency index, representing different frequency components; For the first The dominant frequency characteristics of each IMF component; Calculate the multi-scale energy entropy characteristics: In the formula, It is a multi-scale energy entropy characteristic; The energy of the m-th IMF; This is the sum of the energy of all IMF components; Perform deep feature learning and compute deep temporal features: In the formula, This represents a deep temporal feature vector. For parameters Temporal convolutional networks; All features form a comprehensive feature vector. .

7. The method for determining the type of power grid load situation perception based on the energy value of load situation time series according to claim 1, characterized in that, Let the set of situational awareness types be... Where g is the number of situation types, the importance weight of each feature is calculated through an attention mechanism to generate the attention weight of each feature: In the formula, Let be the attention weight for the d-th feature; This is the transpose of the attention vector, obtained through training; The attention weight matrix is ​​obtained through training. For the comprehensive feature vector The d-th eigenvalue in; This is the attention bias term, obtained through training; All based on calculations Construct the feature attention vector A: In the formula, For feature attention vectors; All based on calculations Construct the feature attention vector A: In the formula, For feature attention vectors; Generate attention-enhanced feature vectors : In the formula, Feature vectors for attention enhancement.

8. A power grid load situation perception type determination system based on load situation time series energy values, characterized in that, include: The data preprocessing module preprocesses the raw collected load situation time series to obtain a standardized time series; The CEEMDAN decomposition module performs CEEMDAN decomposition on the standardized time series to obtain multiple IMF components; The feature extraction module extracts features from each IMF component, calculates the mean, variance, skewness, dominant frequency, multi-scale energy entropy, and deep temporal series features of each IMF, and forms a comprehensive feature vector from all features. The attention-enhanced feature vector generation module calculates the importance weight of each feature through an attention mechanism to generate attention-enhanced feature vectors. The classification module employs a multi-level classification architecture to perform coarse-grained and fine-grained classification on the attention-enhanced feature vectors, obtaining the probability that the feature vectors belong to the situational awareness type. In the formula, This is the coarse-grained situational awareness classification score vector. This is a coarse-grained situational awareness classification weight matrix; This is the coarse-grained situational awareness classification bias vector. For the feature vector to belong to the th The probability of a coarse-grained situational awareness type; Based on coarse-grained classification results and attention enhancement features Fine-grained classification is performed to obtain the probability distribution: In the formula, The score for the feature vector belonging to the g-th fine-grained situational awareness category; This is a fine-grained classification weight matrix; This is a fine-grained classification bias vector; The probability distribution is obtained based on coarse-grained probability distribution and fine-grained classification, and the feature vector is determined to belong to the situational awareness type. The probability of: The feature vector belongs to the situational awareness type. The probability of; The results output module outputs the situational awareness type with the highest probability.

Images