Performance data processing method and system for thermal power plant operators

The characteristics of performance data of thermal power plant operators are extracted through self-attention mechanism and feedforward neural network, and combined with knowledge distillation modules and expert experience, the problems of insufficient data utilization and insufficient model generalization capabilities in the existing technology are solved, and the more Accurate and comprehensive personnel performance evaluation.

CN118469408BActive Publication Date: 2024-11-05中煤新集利辛发电有限公司 +1
View PDF 9 Cites 0 Cited by

Patent Information

Application Number
CN202410925648.3
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2024-07-11
Publication Date
2024-11-05
Estimated Expiration
2044-07-11

AI Technical Summary

Technical Problem

The existing technology has problems such as insufficient data utilization, limited feature extraction capabilities and poor model generalization capabilities in the performance data processing of thermal power plant operation personnel, resulting in inaccurate and comprehensive performance evaluation.

Method used

The performance data characteristics are extracted by self-attention mechanism and feedforward neural network, combined with the knowledge distillation module to fusion embedding vectors and knowledge vectors, and fused through the multi-head attention mechanism and the gated unit, and introduced expert experience and knowledge for weighted summing to form a comprehensive Performance rating.

Benefits of technology

It improves the accuracy and reliability of performance evaluation, enhances the comprehensive and fine-grained evaluation of employee performance, takes into account the advantages of data-driven and expert experience, and improves the authority and credibility of evaluation results.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN118469408B_ABST
    Figure CN118469408B_ABST
Patent Text Reader

Abstract

The present invention provides a performance data processing method and system for thermal power plant operators, which relates to the field of data processing technology, including obtaining original performance data of thermal power plant operators, converting the pre-processed performance data into multiple data sequences that conform to the input format of a performance evaluation model; inputting the multiple data sequences into a pre-constructed performance evaluation model, extracting performance features of the multiple data sequences through a self-attention mechanism and a feedforward neural network, obtaining an embedding vector characterizing each data sequence, fusing the embedding vector and the knowledge vector through a knowledge distillation module of the performance evaluation model in combination with an attention mechanism to obtain a fused vector; inputting the fused vector into a fully connected layer of the performance evaluation model, and obtaining performance scores of the original performance data in multiple dimensions through nonlinear transformation; and introducing expert experience knowledge to perform weighted summation on the performance scores to obtain a comprehensive performance score.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The invention relates to data processing technology, and in particular to a method and system for processing performance data of operators in a thermal power plant. Background Art

[0002] With the rapid development of the power industry and the intensification of market competition, thermal power plants are facing the challenges of improving operating efficiency, reducing costs and ensuring safe and stable operation. Performance management of thermal power plant operators is one of the key factors in achieving these goals. Traditional performance evaluation methods mainly rely on manual evaluation and experience judgment, which have problems such as strong subjectivity, low efficiency, and difficulty in comprehensively considering various factors, and cannot meet the management needs of modern thermal power plants.

[0003] The existing technology has the following deficiencies in processing the performance data of operators in thermal power plants:

[0004] Insufficient data utilization: Failure to fully utilize the massive data generated during the operation of thermal power plants, such as equipment status, operation records, fault information, etc., which contain important information on employee performance.

[0005] Limited feature extraction capabilities: The feature extraction and representation methods for employee performance data are relatively simple, making it difficult to effectively capture key patterns and dependencies in the data, affecting the accuracy of performance evaluation.

[0006] Poor model generalization ability: Existing performance evaluation models are usually targeted at specific employee roles or work scenarios, lack universality and scalability, and are difficult to adapt to the complex and changeable operating environment of thermal power plants. Summary of the invention

[0007] The embodiments of the present invention provide a method and system for processing performance data of operators in a thermal power plant, which can solve the problems in the prior art.

[0008] According to a first aspect of the embodiments of the present invention,

[0009] Provides a method for processing the performance data of thermal power plant operators, including:

[0010] Obtaining original performance data of thermal power plant operators, preprocessing the original performance data to obtain preprocessed performance data, and converting the preprocessed performance data into multiple data sequences that conform to the input format of the performance evaluation model;

[0011] Input the multiple data sequences into a pre-built performance evaluation model, extract the performance features of the multiple data sequences through a self-attention mechanism and a feedforward neural network, obtain an embedding vector representing each data sequence, pass the embedding vector and the pre-acquired knowledge vector through a knowledge distillation module of the performance evaluation model, and fuse the embedding vector and the knowledge vector in combination with an attention mechanism to obtain a fused vector;

[0012] The fusion vector is input into the fully connected layer of the performance evaluation model, and the performance scores of the original performance data in multiple dimensions are obtained through nonlinear transformation; expert experience knowledge is introduced to perform weighted summation on the performance scores to obtain a comprehensive performance score.

[0013] In an optional embodiment,

[0014] The multiple data sequences are input into a pre-built performance evaluation model, and the performance features of the multiple data sequences are extracted through a self-attention mechanism and a feedforward neural network to obtain an embedding vector representing each data sequence, including:

[0015] Encode the multiple data sequences, perform one-hot encoding on discrete data in the multiple data sequences, normalize continuous data, and divide time series data into time segments of fixed length using a sliding window method, where each time segment corresponds to an input sample, to obtain multiple encoded data sequences;

[0016] Input the encoded multiple data sequences into the self-attention layer, and capture the long-distance dependencies in the data sequences through the sparse self-attention mechanism to obtain a set of context vectors;

[0017] The context vector is passed into a feedforward neural network, and the context vector is transformed and abstracted through multiple fully connected layers and nonlinear activation functions to obtain a set of transformation vectors, each of which represents a high-level feature of a data sequence;

[0018] Multiple transformation vectors output by the feedforward neural network are fused by at least one of concatenation, weighted averaging or attention weighting to obtain an embedding vector representing each data sequence.

[0019] In an optional embodiment,

[0020] Before fusing the embedding vector and the pre-acquired knowledge vector through the knowledge distillation module of the performance evaluation model and combining the attention mechanism to obtain the fused vector, the method further includes obtaining the knowledge vector:

[0021] Constructing an original performance knowledge graph from a pre-acquired performance data source; determining a set of neighbor nodes corresponding to each node in the performance knowledge graph, and determining a neighbor weight of any node and its corresponding neighbor node based on an attention mechanism;

[0022] Aggregating the neighbor nodes according to the neighbor weights to determine the graph vector of any node in the performance knowledge graph;

[0023] Based on the graph vector, residual connection and layer normalization operations are introduced into the performance knowledge graph to obtain the knowledge vector.

[0024] In an optional embodiment,

[0025] The embedding vector and the pre-acquired knowledge vector are passed through the knowledge distillation module of the performance evaluation model, and the embedding vector and the knowledge vector are fused in combination with the attention mechanism to obtain a fused vector including;

[0026] The embedding vector and knowledge vector are input into multiple attention heads respectively. Each attention head maps the embedding vector and knowledge vector to different subspaces through linear transformation to obtain the query matrix, key matrix and value matrix.

[0027] For each attention head, the similarity between the query matrix and the key matrix is ​​calculated by dot product to obtain the attention score; the attention score is normalized to obtain the attention weight of each attention head; the attention weight is multiplied by the value matrix to obtain the initial output of each attention head; the initial outputs of all attention heads are concatenated and the final multi-head attention output is obtained through linear transformation;

[0028] For each embedding vector and the corresponding knowledge vector, a fusion weight of the embedding vector and the knowledge vector is calculated through a gating unit; the embedding vector and the knowledge vector are initially fused according to the fusion weight to obtain an initial fusion value; the initial fusion value is combined with the multi-head attention output to obtain a fusion vector.

[0029] In an optional embodiment,

[0030] Calculating the fusion weight of the embedding vector and the knowledge vector by the gating unit includes:

[0031] g i =σ( W g ⋅[ e i || h i ]+ b g ) ;

[0032] Among them, g i represents the i-th fusion weight, σ represents the sigmoid activation function, W g represents the weight matrix of the gate unit, b g represents the bias vector of the gate control unit, e i 、h i Represent the i-th embedding vector and the i-th knowledge vector respectively;

[0033] The embedding vector and the knowledge vector are initially fused according to the fusion weight to obtain an initial fusion value including:

[0034] ;

[0035] Among them, f i represents the i-th initial fusion value;

[0036] Combining the initial fusion value with the multi-head attention output to obtain a fusion vector includes:

[0037] ;

[0038] Among them, F represents the fusion vector, Gate() represents the gating mechanism, MultiHead() represents the multi-head attention mechanism, E and H represent the set of embedding vectors and the set of knowledge vectors, respectively.

[0039] In an optional embodiment,

[0040] The method also includes training a performance evaluation model:

[0041] In each training batch, using an adversarial sample generation function, a corresponding adversarial sample is generated for each original sample, the original sample and the adversarial sample are merged into augmented training data, and a comprehensive loss function of a performance evaluation model to be trained is determined based on the augmented training data, wherein the comprehensive loss function includes a first loss function and a second loss function, the first loss function includes an adversarial loss, and the second loss function includes a knowledge distillation loss;

[0042] The gradient of the comprehensive loss function to the performance evaluation model parameters is calculated through the back-propagation algorithm. According to the statistical information of the historical gradient norm, the clipping threshold is adaptively calculated, and the gradient exceeding the clipping threshold is clipped. The performance evaluation model is iteratively trained using the clipped gradient until the performance evaluation model converges or reaches the preset number of training rounds.

[0043] In an optional embodiment,

[0044] The first loss function includes:

[0045] ;

[0046] Among them, L adv represents the loss value of the first loss function, represents the expected value of the original sample, x and y represent the input sample of the performance evaluation model and the output sample of the performance evaluation model respectively, D represents the original sample set, δ represents the adversarial perturbation, R represents the maximum amplitude limit of the adversarial perturbation, θ represents the performance evaluation model parameter, P() represents the prediction function, and L() represents the cross entropy loss function;

[0047] The second loss function includes:

[0048] ;

[0049] Among them, L kd Represents the loss value of the second loss function, f s (), f t () represents the prediction function of the student model of the knowledge distillation module and the prediction function of the teacher model of the knowledge distillation module, θ s ,θ t They represent the student model parameters of the knowledge distillation module and the teacher model parameters of the knowledge distillation module respectively;

[0050] The comprehensive loss function includes:

[0051] ;

[0052] Among them, L total represents the loss value of the comprehensive loss function, α and β represent the loss weight value corresponding to the first loss function and the loss weight value corresponding to the second loss function, respectively.

[0053] According to a second aspect of the embodiments of the present invention,

[0054] Provide thermal power plant operator performance data processing system, including:

[0055] The first unit is used to obtain original performance data of the thermal power plant operating personnel, preprocess the original performance data to obtain preprocessed performance data, and convert the preprocessed performance data into multiple data sequences that meet the input format of the performance evaluation model;

[0056] The second unit is used to input the multiple data sequences into a pre-built performance evaluation model, extract the performance features of the multiple data sequences through a self-attention mechanism and a feedforward neural network, obtain an embedding vector representing each data sequence, and fuse the embedding vector and the pre-acquired knowledge vector through a knowledge distillation module of the performance evaluation model in combination with an attention mechanism to obtain a fused vector.

[0057] The third unit is used to input the fusion vector into the fully connected layer of the performance evaluation model, and obtain the performance scores of the original performance data in multiple dimensions through nonlinear transformation; introduce expert experience knowledge to perform weighted summation on the performance scores to obtain a comprehensive performance score.

[0058] According to a third aspect of the embodiments of the present invention,

[0059] An electronic device is provided, comprising:

[0060] processor;

[0061] a memory for storing processor-executable instructions;

[0062] The processor is configured to call the instructions stored in the memory to execute the aforementioned method.

[0063] According to a fourth aspect of the embodiments of the present invention,

[0064] A computer-readable storage medium is provided, on which computer program instructions are stored. When the computer program instructions are executed by a processor, the aforementioned method is implemented.

[0065] Using the self-attention mechanism and feedforward neural network to extract the key features of performance data can effectively capture the long-distance dependencies and high-level semantic information in the data, generate a comprehensive and accurate performance feature representation, and improve the accuracy and reliability of performance evaluation. The knowledge distillation module is introduced to fuse the embedding vector of employee performance data with the pre-acquired expert knowledge vector, making full use of domain knowledge and experience, making up for the shortcomings of data-driven methods, and improving the rationality and interpretability of performance evaluation. Through the fully connected layer of the performance evaluation model, the fusion vector is nonlinearly transformed to obtain the performance score of employees in multiple dimensions, realizing a comprehensive and fine-grained evaluation of employee performance and providing rich reference information for performance management. The introduction of expert experience knowledge to perform weighted summation of performance scores to obtain a comprehensive performance score takes into account the advantages of data-driven and expert experience, and improves the authority and credibility of performance evaluation results. BRIEF DESCRIPTION OF THE DRAWINGS

[0066] Figure 1 It is a flow chart of a method for processing performance data of thermal power plant operators according to an embodiment of the present invention;

[0067] Figure 2 The present invention is a schematic diagram of the structure of a thermal power plant operator performance data processing system according to an embodiment of the present invention. DETAILED DESCRIPTION

[0068] In order to make the purpose, technical solution and advantages of the embodiments of the present invention clearer, the technical solution in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0069] The technical solution of the present invention is described in detail with specific embodiments below. The following specific embodiments can be combined with each other, and the same or similar concepts or processes may not be described in detail in some embodiments.

[0070] Figure 1 FIG. 1 is a flow chart of a method for processing performance data of thermal power plant operators according to an embodiment of the present invention. Figure 1 As shown, the method includes:

[0071] S101. Obtaining original performance data of thermal power plant operators, preprocessing the original performance data to obtain preprocessed performance data, and converting the preprocessed performance data into multiple data sequences that conform to the input format of the performance evaluation model;

[0072] In an optional embodiment,

[0073] The multiple data sequences are input into a pre-built performance evaluation model, and the performance features of the multiple data sequences are extracted through a self-attention mechanism and a feedforward neural network to obtain an embedding vector representing each data sequence, including:

[0074] Encode the multiple data sequences, perform one-hot encoding on discrete data in the multiple data sequences, normalize continuous data, and divide time series data into time segments of fixed length using a sliding window method, where each time segment corresponds to an input sample, to obtain multiple encoded data sequences;

[0075] Input the encoded multiple data sequences into the self-attention layer, and capture the long-distance dependencies in the data sequences through the sparse self-attention mechanism to obtain a set of context vectors;

[0076] The context vector is passed into a feedforward neural network, and the context vector is transformed and abstracted through multiple fully connected layers and nonlinear activation functions to obtain a set of transformation vectors, each of which represents a high-level feature of a data sequence;

[0077] Multiple transformation vectors output by the feedforward neural network are fused by at least one of concatenation, weighted averaging or attention weighting to obtain an embedding vector representing each data sequence.

[0078] For example,

[0079] One-hot encodes discrete data in multiple data sequences. Identify discrete features in the data sequence, such as employee position, educational background, etc. Create a binary vector for each discrete feature value, with a length equal to the total number of feature values. For each data sample, set the binary vector element corresponding to the discrete feature value position to 1 and the rest to 0.

[0080] Normalize continuous data in multiple data sequences. Calculate the minimum and maximum values ​​of continuous features. Normalize each continuous feature value to map it to the interval [0, 1].

[0081] For time series data, a sliding window method is used to divide it into time segments of fixed length. Determine the size and step size of the sliding window, such as a window size of 30 days and a step size of 7 days. Starting from the starting position of the time series data, take the sliding window size as a time segment, and extract the data in the segment as an input sample. Move the sliding window backward by one step, extract the next time segment as a new input sample, and repeat the process until the end of the time series data. After completing the above processing, multiple encoded data sequences are obtained, each of which consists of several fixed-length input samples.

[0082] Input multiple encoded data sequences into the self-attention layer, and extract the context information of the data sequence through the sparse self-attention mechanism. Calculate the sparse attention weight matrix between the input samples. Randomly select a part of the input samples as the query vector (query) and the key vector (key), and the rest of the samples as the value vector (value). Calculate the similarity score between the query vector and the key vector. Commonly used similarity functions include dot product, cosine similarity, etc. Apply the softmax function to the similarity score to obtain the normalized sparse attention weight matrix.

[0083] The context vector is calculated using the sparse attention weight matrix. The sparse attention weight matrix is ​​multiplied with the value vector to obtain the weighted summed context vector. A linear transformation and residual connection are applied to the context vector to obtain the final self-attention layer output.

[0084] The context vector output by the self-attention layer is passed into the feedforward neural network to transform and abstract the features. The feedforward neural network is usually constructed by multiple fully connected layers and nonlinear activation functions. Set the number of layers of the feedforward neural network and the number of neurons in each layer. Select a suitable nonlinear activation function, such as ReLU, LeakyReLU, ELU, etc. Input the context vector to the first layer of the feedforward neural network. Apply a linear transformation to the context vector, that is, multiply it with the weight matrix and add a bias term. Apply a nonlinear activation function to the result of the linear transformation to obtain the output of the first layer. Use the output of the first layer as the input of the next layer until the last layer of the feedforward neural network. The output of the last layer of the feedforward neural network is the transformed feature vector, each of which represents the high-level features of a data sequence. Fusion of multiple transformation vectors output by the feedforward neural network to obtain the embedding vector representing each data sequence.

[0085] Concatenation and fusion: Multiple transformation vectors are directly concatenated in the feature dimension to obtain a long vector as the embedding vector. A weight is assigned to each transformation vector, which can be a pre-set fixed value or dynamically calculated through the attention mechanism. Multiple transformation vectors are multiplied by the corresponding weights and then summed to obtain the weighted average embedding vector.

[0086] The attention mechanism is used to calculate the weight of each transformation vector, usually based on the similarity score between the transformation vector and a global query vector. Multiple transformation vectors are multiplied by the corresponding attention weights and then summed to obtain the attention-weighted embedding vector. The obtained embedding vector concentrates on the high-level feature information of each data sequence and can be used for subsequent performance evaluation tasks.

[0087] This embodiment achieves effective encoding of different types of data by one-hot encoding discrete data, normalizing continuous data, and dividing time series data into time segments using a sliding window method, unifies the data format, improves the quality and availability of data, and lays a good foundation for subsequent feature extraction. The sparse self-attention mechanism is used to process the encoded data sequence, and the sparse attention weights between samples are calculated to effectively capture the long-distance dependencies in the data sequence, overcome the limitation that traditional recurrent neural networks are difficult to process long sequence data, extract more comprehensive and accurate context information, and enhance the model's understanding and representation of complex performance data. The context vector output by the self-attention layer is passed into the feedforward neural network, and the features are transformed and abstracted through multiple fully connected layers and nonlinear activation functions, realizing high-level representation learning of features, extracting high-level semantic features of the data sequence, enhancing the model's feature expression ability and generalization ability, and improving the accuracy of performance evaluation.

[0088] In an optional embodiment,

[0089] Before fusing the embedding vector and the pre-acquired knowledge vector through the knowledge distillation module of the performance evaluation model and combining the attention mechanism to obtain the fused vector, the method further includes obtaining the knowledge vector:

[0090] Constructing an original performance knowledge graph from a pre-acquired performance data source; determining a set of neighbor nodes corresponding to each node in the performance knowledge graph, and determining a neighbor weight of any node and its corresponding neighbor node based on an attention mechanism;

[0091] Aggregating the neighbor nodes according to the neighbor weights to determine the graph vector of any node in the performance knowledge graph;

[0092] Based on the graph vector, residual connection and layer normalization operations are introduced into the performance knowledge graph to obtain the knowledge vector.

[0093] Exemplarily, an original performance knowledge graph is constructed from a pre-acquired performance data source. Specifically, the node types of the performance knowledge graph are determined, such as employee nodes, department nodes, performance indicator nodes, etc. Node entities and their attribute information are extracted from the performance data source, such as employee personal information, department information, performance indicator values, etc. Edge connections are established between nodes based on the relationships between node entities, such as the department to which an employee belongs, the association between performance indicators and employees, etc. The extracted node entities, attribute information, and edge connections are organized into an original performance knowledge graph.

[0094] Determine the set of neighbor nodes corresponding to each node in the performance knowledge graph. Traverse each node in the performance knowledge graph. For each node, find the nodes directly connected to it to form the set of neighbor nodes of the node. Store each node and its corresponding set of neighbor nodes for subsequent calculations.

[0095] Determine the neighbor weights of any node and its corresponding neighbor nodes based on the attention mechanism. For any node in the performance knowledge graph, calculate the attention weights between it and its neighbor nodes. Concatenate the feature vector of the current node with the feature vector of the neighbor node as the input of the attention mechanism. Through the calculation of the attention mechanism, the attention weights between the current node and each neighbor node are obtained. The attention weights are normalized so that their sum is 1.

[0096] According to the neighbor weights, the neighbor nodes are aggregated to determine the graph vector of any node in the performance knowledge graph. For any node in the performance knowledge graph, the neighbor nodes are weighted summed according to their neighbor weights. The feature vector of the current node is concatenated with the feature vector of the neighbor nodes after weighted summation. The concatenated vector is transformed through a fully connected layer to obtain the graph vector representation of the current node.

[0097] Based on the graph vector, residual connection and layer normalization operations are introduced in the performance knowledge graph to obtain the knowledge vector. The graph vector is layer normalized to make its mean 0 and variance 1. The layer-normalized graph vector is residually connected with the original node feature vector, that is, the two are added. The vector after residual connection is layer normalized again to obtain the final knowledge vector. The knowledge vector and the embedding vector are fused through the knowledge distillation module of the performance evaluation model.

[0098] The knowledge vector is input into the knowledge distillation module together with the previously obtained embedding vector. In the knowledge distillation module, the attention mechanism is used to calculate the fusion weight between the knowledge vector and the embedding vector. According to the fusion weight, the knowledge vector and the embedding vector are weighted summed to obtain the fused vector representation. The fused vector is transformed through a fully connected layer to obtain the final fused vector.

[0099] This embodiment constructs the original performance knowledge graph from the pre-acquired performance data source, and structurally represents key elements such as employees, departments, performance indicators, and the relationships between them, forming a comprehensive and systematic performance knowledge base, which provides a rich and reliable data foundation for the subsequent acquisition of knowledge vectors. The attention mechanism is used to determine the neighbor weights between any node and its neighbor nodes in the performance knowledge graph, which can adaptively learn the importance between nodes, highlight the influence of key nodes, suppress the interference of noise nodes, and improve the accuracy and robustness of knowledge representation. The neighbor nodes are aggregated by neighbor weights to obtain a graph vector representation of the node in the performance knowledge graph, which effectively integrates the feature information of the node itself and the contextual information of the neighbor nodes, obtains a more comprehensive and semantically rich node representation, and enhances the expressive power of the knowledge vector. The acquired knowledge vector is integrated with the data-driven embedding vector through the attention mechanism, which can make full use of information from two different sources, complement and promote each other, and obtain a more accurate and comprehensive performance feature representation, which improves the performance and generalization ability of the performance evaluation model.

[0100] S102. Input the multiple data sequences into a pre-built performance evaluation model, extract the performance features of the multiple data sequences through a self-attention mechanism and a feedforward neural network, obtain an embedding vector representing each data sequence, pass the embedding vector and the pre-acquired knowledge vector through a knowledge distillation module of the performance evaluation model, and fuse the embedding vector and the knowledge vector in combination with an attention mechanism to obtain a fused vector;

[0101] In an optional embodiment,

[0102] The embedding vector and the pre-acquired knowledge vector are passed through the knowledge distillation module of the performance evaluation model, and the embedding vector and the knowledge vector are fused in combination with the attention mechanism to obtain a fused vector including;

[0103] The embedding vector and knowledge vector are input into multiple attention heads respectively. Each attention head maps the embedding vector and knowledge vector to different subspaces through linear transformation to obtain the query matrix, key matrix and value matrix.

[0104] For each attention head, the similarity between the query matrix and the key matrix is ​​calculated by dot product to obtain the attention score; the attention score is normalized to obtain the attention weight of each attention head; the attention weight is multiplied by the value matrix to obtain the initial output of each attention head; the initial outputs of all attention heads are concatenated and the final multi-head attention output is obtained through linear transformation;

[0105] For each embedding vector and the corresponding knowledge vector, a fusion weight of the embedding vector and the knowledge vector is calculated through a gating unit; the embedding vector and the knowledge vector are initially fused according to the fusion weight to obtain an initial fusion value; the initial fusion value is combined with the multi-head attention output to obtain a fusion vector.

[0106] Exemplarily, the embedding vector and the knowledge vector are input into multiple attention heads for processing. The number of attention heads can be set according to task requirements and computing resources. For each attention head, the following operations are performed:

[0107] The embedding vector is mapped to the query subspace through linear transformation to obtain the query matrix; the knowledge vector is mapped to the key subspace through linear transformation to obtain the key matrix; the knowledge vector is mapped to the value subspace through linear transformation to obtain the value matrix.

[0108] For each attention head, calculate the attention score and get the attention weight. For each attention head, do the following:

[0109] The similarity between the query matrix and the key matrix is ​​calculated by dot product to get the attention score. Apply the softmax function to normalize the attention score to get the attention weight of the attention head. Multiply the attention weight with the value matrix to get the initial output of each attention head. For each attention head, do the following:

[0110] Multiply the attention weights by the value matrix to get the initial output of the attention head. Concatenate the initial outputs of all attention heads and use a linear transformation to get the final multi-head attention output. Concatenate the initial outputs of all attention heads in the feature dimension to get the concatenated matrix. Map the concatenated matrix to the output space through a linear transformation to get the final multi-head attention output.

[0111] The fusion weights of the embedding vector and the knowledge vector are calculated through the gating unit, and the initial fusion is performed. For each embedding vector and the corresponding knowledge vector, the initial fusion value is combined with the multi-head attention output to obtain the final fusion vector.

[0112] For each initial fusion value, perform the following operations: concatenate the initial fusion value with the multi-head attention output in the feature dimension to obtain the concatenated vector. Map the concatenated vector to the output space through linear transformation to obtain the final fusion vector.

[0113] In an optional embodiment,

[0114] Calculating the fusion weight of the embedding vector and the knowledge vector by the gating unit includes:

[0115] g i =σ( W g ⋅[ e i || h i ]+ b g ) ;

[0116] Among them, g i represents the i-th fusion weight, σ represents the sigmoid activation function, W g represents the weight matrix of the gate unit, b g represents the bias vector of the gate control unit, e i 、h i Represent the i-th embedding vector and the i-th knowledge vector respectively;

[0117] The embedding vector and the knowledge vector are initially fused according to the fusion weight to obtain an initial fusion value including:

[0118] ;

[0119] Among them, f i represents the i-th initial fusion value;

[0120] Combining the initial fusion value with the multi-head attention output to obtain a fusion vector includes:

[0121] ;

[0122] Among them, F represents the fusion vector, Gate() represents the gating mechanism, MultiHead() represents the multi-head attention mechanism, E and H represent the set of embedding vectors and the set of knowledge vectors, respectively.

[0123] This embodiment introduces a multi-head attention mechanism and calculates the correlation between the embedding vector and the knowledge vector from different subspace perspectives. It can capture more comprehensive and fine-grained interaction information between the two information sources, fully explore their potential semantic associations, achieve diversity and complementarity of feature representation, and improve the representation and generalization capabilities of the fusion vector.

[0124] In each attention head, the similarity between the query matrix and the key matrix is ​​calculated by dot product to obtain the attention score, which is normalized to obtain the attention weight, which can adaptively learn the importance of the embedding vector and the knowledge vector in different subspaces, highlight the influence of key information, suppress the interference of noise information, and improve the accuracy and robustness of the fusion process. The attention weight is multiplied by the value matrix to obtain the initial output of each attention head, and then the outputs of all attention heads are spliced ​​together and the final multi-head attention output is obtained through linear transformation, which realizes information aggregation and high-level feature extraction across subspaces, and enhances the semantic expression and abstraction capabilities of the fusion vector.

[0125] For each embedding vector and the corresponding knowledge vector, the gating unit calculates their fusion weights, which can adaptively adjust the importance of the two information sources and achieve dynamic trade-offs and optimal combinations. When the information of the embedding vector is more critical, the fusion weight will give it a higher weight; when the information of the knowledge vector is more important, the fusion weight will give it a higher weight. This adaptive fusion mechanism can flexibly adjust the fusion strategy according to different evaluation objects and scenarios, improving the adaptability and effectiveness of the fusion results.

[0126] S103. Input the fusion vector into the fully connected layer of the performance evaluation model, and obtain the performance scores of the original performance data in multiple dimensions through nonlinear transformation; introduce expert experience knowledge to perform weighted summation on the performance scores to obtain a comprehensive performance score.

[0127] Exemplarily, the fusion vector is input into the fully connected layer of the performance evaluation model, and the performance scores of the original performance data in multiple dimensions are obtained through nonlinear transformation, including:

[0128] At the top level of the performance evaluation model, a fully connected layer can be used to perform nonlinear transformations on the fused vectors to generate performance scores for the original performance data in multiple dimensions. The fully connected layer achieves a high degree of abstraction of the feature space and improves the representation capability by performing weighted combination and nonlinear activation of the input features.

[0129] Specifically, let the fusion vector be , where d is the dimension of the fusion vector. First, the fusion vector v is input into the fully connected layer, and the hidden representation is obtained through linear transformation. :

[0130] ;

[0131] in, is the weight matrix of the fully connected layer, is the bias term, and m is the dimension of the hidden representation.

[0132] By introducing nonlinear features, nonlinear activation functions can be applied to the hidden representation , and get the activated hidden representation :

[0133] ;

[0134] In this application, the ReLU activation function is selected, that is, , because it can effectively alleviate the gradient vanishing problem and accelerate the convergence of the model.

[0135] Finally, the activated hidden representation a can be mapped to multiple evaluation dimensions of the original performance data through another linear transformation to obtain the final performance score vector :

[0136] ;

[0137] in, is the weight matrix of the output layer, is the bias term, and k is the number of evaluation dimensions.

[0138] Through the nonlinear transformation process of the fully connected layer, the fusion vector can be mapped to multiple evaluation dimensions of the original performance data, and the performance score of each dimension can be obtained. This nonlinear transformation can effectively capture the high-level semantic information contained in the fusion vector and convert it into the score of the corresponding dimension, thereby realizing a comprehensive evaluation of the original performance data.

[0139] In order to further improve the representation ability and generalization performance of the model, multiple hidden layers can be introduced before the fully connected layer to build a deeper neural network structure. At the same time, in order to prevent the occurrence of overfitting problems, dropout regularization can be added between the fully connected layers to randomly shut down some neurons to enhance the robustness and generalization ability of the model.

[0140] During the model training process, the gradient of the loss function to the model parameters is calculated through the back-propagation algorithm, and the model parameters are updated using an optimization algorithm (such as the Adam algorithm) to minimize the difference between the predicted performance score and the actual score. After multiple rounds of iterative optimization, the model can learn the nonlinear mapping relationship between the fusion vector and the multi-dimensional performance score, thereby achieving accurate evaluation of the original performance data.

[0141] In an optional embodiment,

[0142] The method also includes training a performance evaluation model:

[0143] In each training batch, using an adversarial sample generation function, a corresponding adversarial sample is generated for each original sample, the original sample and the adversarial sample are merged into augmented training data, and a comprehensive loss function of a performance evaluation model to be trained is determined based on the augmented training data, wherein the comprehensive loss function includes a first loss function and a second loss function, the first loss function includes an adversarial loss, and the second loss function includes a knowledge distillation loss;

[0144] The gradient of the comprehensive loss function to the performance evaluation model parameters is calculated through the back-propagation algorithm. According to the statistical information of the historical gradient norm, the clipping threshold is adaptively calculated, and the gradient exceeding the clipping threshold is clipped. The performance evaluation model is iteratively trained using the clipped gradient until the performance evaluation model converges or reaches the preset number of training rounds.

[0145] In an optional embodiment,

[0146] The first loss function includes:

[0147] ;

[0148] Among them, L adv represents the loss value of the first loss function, represents the expected value of the original sample, x and y represent the input sample of the performance evaluation model and the output sample of the performance evaluation model respectively, D represents the original sample set, δ represents the adversarial perturbation, R represents the maximum amplitude limit of the adversarial perturbation, θ represents the performance evaluation model parameter, P() represents the prediction function, and L() represents the cross entropy loss function;

[0149] The second loss function includes:

[0150] ;

[0151] Among them, L kd Represents the loss value of the second loss function, f s (), f t () represents the prediction function of the student model of the knowledge distillation module and the prediction function of the teacher model of the knowledge distillation module, θ s ,θ t They represent the student model parameters of the knowledge distillation module and the teacher model parameters of the knowledge distillation module respectively;

[0152] The comprehensive loss function includes:

[0153] ;

[0154] Among them, L total represents the loss value of the comprehensive loss function, α and β represent the loss weight value corresponding to the first loss function and the loss weight value corresponding to the second loss function, respectively.

[0155] Exemplarily, for each training batch, an adversarial sample generation function is used to generate adversarial samples.

[0156] For each original sample in the batch, perform the following operations: Generate the corresponding adversarial sample using the adversarial sample generation function. Combine the original sample and the adversarial sample into an augmented training sample pair.

[0157] The first loss function includes:

[0158] ;

[0159] Among them, L adv represents the loss value of the first loss function, represents the expected value of the original sample, x and y represent the input sample of the performance evaluation model and the output sample of the performance evaluation model respectively, D represents the original sample set, δ represents the adversarial perturbation, R represents the maximum amplitude limit of the adversarial perturbation, θ represents the performance evaluation model parameter, P() represents the prediction function, and L() represents the cross entropy loss function;

[0160] The second loss function includes:

[0161] ;

[0162] Among them, L kd Represents the loss value of the second loss function, f s (), f t () represents the prediction function of the student model of the knowledge distillation module and the prediction function of the teacher model of the knowledge distillation module, θ s,θ t They represent the student model parameters of the knowledge distillation module and the teacher model parameters of the knowledge distillation module respectively;

[0163] The comprehensive loss function includes:

[0164] ;

[0165] Among them, L total represents the loss value of the comprehensive loss function, α and β represent the loss weight value corresponding to the first loss function and the loss weight value corresponding to the second loss function, respectively.

[0166] Calculate the gradient of the comprehensive loss function with respect to the performance evaluation model parameters through the back-propagation algorithm. Back-propagate the comprehensive loss function to calculate its gradient with respect to the model parameters. Record the gradient norm of the current batch and add it to the statistical information of the historical gradient norm. Adaptively calculate the clipping threshold based on the statistical information of the historical gradient norm, and clip the gradients that exceed the clipping threshold. Adaptively calculate the clipping threshold of the current batch based on the statistical information of the historical gradient norm (such as the mean and standard deviation). Clip the gradients of the current batch. Using the clipped gradients, iteratively update the parameters of the performance evaluation model through an optimization algorithm (such as Adam, SGD, etc.). Using the clipped gradients, calculate the parameter update amount through the optimization algorithm, and apply the parameter update amount to the current model parameters.

[0167] Exemplarily, the clipping threshold is adaptively calculated based on statistical information of historical gradient norms.

[0168] Calculate the mean and standard deviation of the historical gradient norm: Assume that the set of historical gradient norms is {g_1, g_2, ..., g_t}, where t is the number of the current batch. According to the mean and standard deviation, adaptively calculate the clipping threshold of the current batch: Set the clipping factor to control the size of the clipping threshold. Calculate the clipping threshold , where μ and φ represent the mean and standard deviation of the historical gradient norm, and C represents the clipping factor.

[0169] Clip the gradients that exceed the clipping threshold and clip the gradients of the current batch:

[0170] if , then the gradient is clipped to: .

[0171] Otherwise, keep the gradient unchanged. Represents the gradient of the comprehensive loss function with respect to the model parameters.

[0172] Using the clipped gradients, the parameters of the performance evaluation model are iteratively updated through an optimization algorithm. Using the clipped gradients, the parameter updates are calculated through an optimization algorithm (such as Adam, SGD, etc.). The parameter updates are applied to the current model parameters.

[0173] This embodiment effectively enhances the diversity and robustness of the training data by utilizing the adversarial sample generation function in each training batch to generate a corresponding adversarial sample for each original sample, and merging the original sample and the adversarial sample into expanded training data. The introduction of adversarial samples can simulate potential noise disturbances and adversarial attacks, improve the generalization ability and stability of the model in the face of these challenges, and reduce the risk of overfitting. The comprehensive loss function of the performance evaluation model to be trained is determined based on the expanded training data, wherein the comprehensive loss function includes a first loss function and a second loss function. The first loss function includes an adversarial loss, which is used to measure the performance of the model on adversarial samples, prompting the model to learn more robust feature representations, and improving the model's adversarial defense capabilities. The second loss function includes a knowledge distillation loss, which is used to measure the degree of information retention of the model when fusing the embedding vector and the knowledge vector, ensuring that key information is not lost during the fusion process, and improving the quality and interpretability of the fusion vector.

[0174] By calculating the gradient of the comprehensive loss function to the performance evaluation model parameters through the back-propagation algorithm, the optimization goal of the loss function can be effectively passed to the model parameters, guiding the model to adjust and update in the direction of reducing the comprehensive loss. Gradient information contains the sensitivity of the model parameters to the loss function and is the key driving force for optimizing the model. According to the statistical information of the historical gradient norm, the clipping threshold is adaptively calculated to clip the gradient that exceeds the clipping threshold. Gradient clipping technology can effectively control the norm of the gradient and prevent the occurrence of the gradient explosion problem. By adaptively adjusting the clipping threshold, the clipping strength of the gradient can be dynamically controlled according to the training status at different stages, while ensuring the optimization stability, the effective information of the gradient can be retained as much as possible, accelerating the convergence speed of the model.

[0175] Figure 2 FIG. 1 is a schematic diagram of the structure of a thermal power plant operator performance data processing system according to an embodiment of the present invention. Figure 2 As shown, the system comprises:

[0176] The first unit is used to obtain original performance data of the thermal power plant operating personnel, preprocess the original performance data to obtain preprocessed performance data, and convert the preprocessed performance data into multiple data sequences that meet the input format of the performance evaluation model;

[0177] The second unit is used to input the multiple data sequences into a pre-built performance evaluation model, extract the performance features of the multiple data sequences through a self-attention mechanism and a feedforward neural network, obtain an embedding vector representing each data sequence, and fuse the embedding vector and the pre-acquired knowledge vector through a knowledge distillation module of the performance evaluation model in combination with an attention mechanism to obtain a fused vector.

[0178] The third unit is used to input the fusion vector into the fully connected layer of the performance evaluation model, and obtain the performance scores of the original performance data in multiple dimensions through nonlinear transformation; introduce expert experience knowledge to perform weighted summation on the performance scores to obtain a comprehensive performance score.

[0179] According to a third aspect of the embodiments of the present invention,

[0180] An electronic device is provided, comprising:

[0181] processor;

[0182] a memory for storing processor-executable instructions;

[0183] The processor is configured to call the instructions stored in the memory to execute the aforementioned method.

[0184] According to a fourth aspect of the embodiments of the present invention,

[0185] A computer-readable storage medium is provided, on which computer program instructions are stored. When the computer program instructions are executed by a processor, the aforementioned method is implemented.

[0186] The present invention may be a method, an apparatus, a system and / or a computer program product. The computer program product may include a computer-readable storage medium carrying computer-readable program instructions for executing various aspects of the present invention.

[0187] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or replace some or all of the technical features therein with equivalents. However, these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for processing performance data of thermal power plant operators, characterized in that: include: Obtaining original performance data of thermal power plant operators, preprocessing the original performance data to obtain preprocessed performance data, and converting the preprocessed performance data into multiple data sequences that conform to the input format of the performance evaluation model; Input the multiple data sequences into a pre-built performance evaluation model, extract the performance features of the multiple data sequences through a self-attention mechanism and a feedforward neural network, obtain an embedding vector representing each data sequence, pass the embedding vector and the pre-acquired knowledge vector through a knowledge distillation module of the performance evaluation model, and fuse the embedding vector and the knowledge vector in combination with an attention mechanism to obtain a fused vector; The fusion vector is input into the fully connected layer of the performance evaluation model, and the performance scores of the original performance data in multiple dimensions are obtained through nonlinear transformation; expert experience knowledge is introduced to perform weighted summation on the performance scores to obtain a comprehensive performance score; The multiple data sequences are input into a pre-built performance evaluation model, and the performance features of the multiple data sequences are extracted through a self-attention mechanism and a feedforward neural network to obtain an embedding vector representing each data sequence, including: Encode the multiple data sequences, perform one-hot encoding on discrete data in the multiple data sequences, normalize continuous data, and divide time series data into time segments of fixed length using a sliding window method, where each time segment corresponds to an input sample, to obtain multiple encoded data sequences; Input the encoded multiple data sequences into the self-attention layer, and capture the long-distance dependencies in the data sequences through the sparse self-attention mechanism to obtain a set of context vectors; The context vector is passed into a feedforward neural network, and the context vector is transformed and abstracted through multiple fully connected layers and nonlinear activation functions to obtain a set of transformation vectors, each of which represents a high-level feature of a data sequence; The multiple transformation vectors output by the feedforward neural network are fused by at least one of concatenation, weighted averaging or attention weighting to obtain an embedding vector representing each data sequence; Before fusing the embedding vector and the pre-acquired knowledge vector through the knowledge distillation module of the performance evaluation model and combining the attention mechanism to obtain the fused vector, the method further includes obtaining the knowledge vector: Constructing an original performance knowledge graph from a pre-acquired performance data source; determining a set of neighbor nodes corresponding to each node in the performance knowledge graph, and determining a neighbor weight of any node and its corresponding neighbor node based on an attention mechanism; Aggregating the neighbor nodes according to the neighbor weights to determine the graph vector of any node in the performance knowledge graph; Based on the graph vector, a residual connection and a layer normalization operation are introduced into the performance knowledge graph to obtain the knowledge vector; The embedding vector and the pre-acquired knowledge vector are passed through the knowledge distillation module of the performance evaluation model, and the embedding vector and the knowledge vector are fused in combination with the attention mechanism to obtain a fused vector including; The embedding vector and the knowledge vector are input into multiple attention heads respectively. Each attention head maps the embedding vector and the knowledge vector to different subspaces through linear transformation to obtain the query matrix, key matrix and value matrix. For each attention head, do the following: The embedding vector is mapped to the query subspace through linear transformation to obtain the query matrix; the knowledge vector is mapped to the key subspace through linear transformation to obtain the key matrix; the knowledge vector is mapped to the value subspace through linear transformation to obtain the value matrix; For each attention head, the similarity between the query matrix and the key matrix is ​​calculated by dot product to obtain the attention score; the attention score is normalized to obtain the attention weight of each attention head; the attention weight is multiplied by the value matrix to obtain the initial output of each attention head; the initial outputs of all attention heads are concatenated and the final multi-head attention output is obtained through linear transformation; For each embedding vector and the corresponding knowledge vector, a fusion weight of the embedding vector and the knowledge vector is calculated by a gating unit; the embedding vector and the knowledge vector are initially fused according to the fusion weight to obtain an initial fusion value; the initial fusion value is combined with the multi-head attention output to obtain a fusion vector; Calculating the fusion weight of the embedding vector and the knowledge vector by the gating unit includes: ; in, g i Indicates i fusion weights, σ ​​represents the sigmoid activation function, W g represents the weight matrix of the gating unit, b g represents the bias vector of the gating unit, e i , h i Respectively represent i The embedding vector and i knowledge vectors; The embedding vector and the knowledge vector are initially fused according to the fusion weight to obtain an initial fusion value including: ; in, f i Indicates i Initial fusion value; Combining the initial fusion value with the multi-head attention output to obtain a fusion vector includes: ; in, F represents the fusion vector, Gate() represents the gating mechanism, MultiHead() represents the multi-head attention mechanism, E , H They represent the set of embedding vectors and the set of knowledge vectors respectively.

2. The method according to claim 1, characterized in that The method also includes training a performance evaluation model: In each training batch, using an adversarial sample generation function, a corresponding adversarial sample is generated for each original sample, the original sample and the adversarial sample are merged into augmented training data, and a comprehensive loss function of a performance evaluation model to be trained is determined based on the augmented training data, wherein the comprehensive loss function includes a first loss function and a second loss function, the first loss function includes an adversarial loss, and the second loss function includes a knowledge distillation loss; The gradient of the comprehensive loss function to the performance evaluation model parameters is calculated through the back-propagation algorithm. According to the statistical information of the historical gradient norm, the clipping threshold is adaptively calculated, and the gradient exceeding the clipping threshold is clipped. The performance evaluation model is iteratively trained using the clipped gradient until the performance evaluation model converges or reaches the preset number of training rounds.

3. The method according to claim 2, characterized in that The first loss function includes: ; in, L adv represents the loss value of the first loss function, represents the expected value of the original sample, x , y They represent the input samples of the performance evaluation model and the output samples of the performance evaluation model respectively. D represents the original sample set, δ represents the adversarial perturbation, R represents the maximum amplitude limit of the adversarial disturbance, θ represents the performance evaluation model parameter, P() represents the prediction function, L() represents the cross entropy loss function; The second loss function includes: ; in, L kd represents the loss value of the second loss function, f s ()、f t () represents the prediction function of the student model of the knowledge distillation module and the prediction function of the teacher model of the knowledge distillation module, θ s , θ t They represent the student model parameters of the knowledge distillation module and the teacher model parameters of the knowledge distillation module respectively; The comprehensive loss function includes: ; in, L total represents the loss value of the comprehensive loss function, α and β represent the loss weight value corresponding to the first loss function and the loss weight value corresponding to the second loss function, respectively.

4. A thermal power plant operator performance data processing system, used to implement the method described in any one of claims 1 to 3, characterized in that: include: The first unit is used to obtain original performance data of the thermal power plant operating personnel, preprocess the original performance data to obtain preprocessed performance data, and convert the preprocessed performance data into multiple data sequences that meet the input format of the performance evaluation model; The second unit is used to input the multiple data sequences into a pre-built performance evaluation model, extract the performance features of the multiple data sequences through a self-attention mechanism and a feedforward neural network, obtain an embedding vector representing each data sequence, and fuse the embedding vector and the pre-acquired knowledge vector through a knowledge distillation module of the performance evaluation model in combination with an attention mechanism to obtain a fused vector. The third unit is used to input the fusion vector into the fully connected layer of the performance evaluation model, and obtain the performance scores of the original performance data in multiple dimensions through nonlinear transformation; introduce expert experience knowledge to perform weighted summation on the performance scores to obtain a comprehensive performance score.

5. An electronic device, characterized in that: include: processor; a memory for storing processor-executable instructions; The processor is configured to call the instructions stored in the memory to execute the method described in any one of claims 1 to 3.

6. A computer-readable storage medium having computer program instructions stored thereon, characterized in that: When the computer program instructions are executed by a processor, the method according to any one of claims 1 to 3 is implemented.

Citation Information

Patent Citations

  • Aspect-level emotion prediction method and device based on semantic syntax

    CN115860006A

  • Performance evaluation method and system based on graph convolutional neural network

    CN116109195A

  • Multimodal medical image conversion method and system based on knowledge distillation and adversarial attack

    CN116596910A

  • Performance evaluation method based on attention mechanism neural network

    CN117273516A

  • Online evaluation score evaluation method based on artificial intelligence

    CN117852974A