Power terminal equipment credibility evaluation method and system based on multi-source data fusion

Abstract

The invention provides a power terminal equipment credibility evaluation method and system based on multi-source data fusion, and the method comprises the steps: obtaining a state data sequence, and enabling each piece of equipment state data in the state data sequence to correspond to one piece of power terminal equipment; inputting the state data sequence into a preset evaluation model, so that the evaluation model performs feature embedding on each piece of equipment state data in the state data sequence to obtain embedded representation data of each piece of equipment state data, generating a corresponding attention score matrix according to each piece of embedded representation data based on a self-attention learning mechanism; and simultaneously performing data mapping on each attention score matrix in a parallel operation mode through an NAT Decoder decoder to generate a credibility evaluation result of each power terminal device, thereby improving the efficiency and accuracy of credibility evaluation of each power terminal device.

Description

Technical Field

[0001] The present application relates to the technical field of monitoring and evaluation of power terminal equipment, and in particular to a method and system for evaluating the credibility of power terminal equipment based on multi-source data fusion. Background Art

[0002] With the rapid development of modern society, the scale and complexity of power systems are increasing. As key components of power systems, smart power terminals, such as smart meters, smart switches, and smart transformers, play a vital role in the production, transmission, distribution, and use of electricity. These devices can realize real-time monitoring, remote control, and data interaction of power parameters, effectively improving the reliability, safety, and operating efficiency of power systems. However, there are still many challenges: (1) Power terminals face network security threats, the operating environment is complex and changeable, and terminal devices may be affected by external environmental factors or internal faults. Effective security protection measures are required for the collected data to prevent data from being tampered with, lost, or leaked. (2) The upgrading of various equipment in the power system makes the system highly dynamic and uncertain. The accuracy and reliability of the data need to be verified and maintained. In response to the above challenges, how to use limited equipment-related status information to conduct efficient and accurate credibility assessment of smart power terminals deserves further research. Credibility assessment of power terminal equipment refers to a comprehensive assessment process that determines whether it can reliably perform tasks and resist various risks in the power system by quantitatively analyzing multiple indicators such as the equipment's operating status, behavioral characteristics, and environmental adaptability. Its core goal is to ensure the data authenticity, functional stability and system interaction security of terminal devices in the Power Internet of Things (PIoT).

[0003] In recent years, time series prediction algorithms have emerged in numerous fields. Time series data has unique properties: it is collected in chronological order, and the current data is intrinsically linked to previous data. Data changes exhibit trends and periodicity, but they can also experience irregular changes due to accidental factors. This characteristic makes the application and research of time series prediction algorithms crucial in various fields. While existing time series prediction methods can already assess the security and reliability of mobile devices, the amount of data in power systems is vast, complex, and diverse, requiring the model's decision-making process and results to be highly transparent and interpretable. Furthermore, new devices are constantly being added to new power systems, and device types and operating environments are becoming increasingly diverse, requiring models to be adaptable and generalizable. A single prediction algorithm is unable to meet the data and model calculation requirements of power systems. Summary of the Invention

[0004] In response to the above technical problems, the present application provides a method and system for credibility assessment of power terminal equipment based on multi-source data fusion, which improves the efficiency and accuracy of credibility assessment of various power terminal equipment.

[0005] In a first aspect, an embodiment of the present application provides a method for evaluating the credibility of power terminal equipment based on multi-source data fusion, comprising: Acquire a status data sequence, where each piece of device status data in the status data sequence corresponds to a power terminal device, and the device status data includes real-time operation data, terminal characteristic data, and time characteristic data of the corresponding power terminal device; Inputting the state data sequence into a preset evaluation model so that the evaluation model performs feature embedding on each device state data item in the state data sequence to obtain embedded representation data for each device state data item, and generating a corresponding attention score matrix based on each embedded representation data item based on a self-attention learning mechanism, wherein the embedded representation data includes position embedding information, time embedding information, and feature scalars; The NAT Decoder decodes the attention score matrices in parallel to generate a credibility assessment result for each power terminal device. Among them, the evaluation model is obtained by training the initial evaluation model in a supervised learning manner based on the historical status data of each power terminal device, and the initial evaluation model is constructed based on the Transformer-CNN hybrid architecture.

[0006] The embodiment of the present application provides a method for evaluating the credibility of power terminal equipment based on multi-source data fusion, which forms a state data sequence with a multi-dimensional feature combination by integrating real-time operation data, terminal features and time features. Then, an evaluation model based on a Transformer-CNN hybrid architecture is used to perform feature embedding and self-attention learning on the state data sequence to generate several attention score matrices, wherein the Transformer architecture can use the self-attention mechanism to capture the temporal dependency and global correlation of the power equipment state, and the CNN architecture is used to extract local features and enhance sensitivity to sudden events. The Transformer-CNN hybrid architecture combines the advantages of both, which can model long-term dependencies and capture local details, improve the adaptability of the model to complex power data, and improve the accuracy of credibility evaluation. Finally, the NAT Decoder decoder is used to parallel map the attention score matrix and obtain the credibility evaluation results of each power terminal device at the same time, avoiding the delay problem of traditional serial decoding, meeting the real-time evaluation requirements of massive equipment, and improving the efficiency of credibility evaluation of various power terminal devices.

[0007] Furthermore, the acquisition of the status data sequence includes: The preset monitoring system collects real-time current data, real-time voltage data, real-time power consumption and environmental status information of each power terminal device to build real-time operating data of each power terminal device; Obtaining corresponding data upload frequency, data upload success rate, data upload delay, business traffic ratio, digital certificate information, and historical credibility evaluation results from each of the power terminal devices to construct terminal feature data of each power terminal device; Based on a preset time alignment algorithm, aligning each of the real-time operation data with a preset timestamp according to the collection time of each of the real-time operation data, and then constructing time feature data corresponding to the real-time operation data according to the timestamp; Performing data cleaning, data segmentation, missing value processing, feature construction, and data normalization on each of the real-time operation data, terminal feature data, and time feature data to obtain a preprocessed data set; The correlation values between the various features in the preprocessed data set are calculated by the Pearson correlation coefficient method, and a number of related features are determined from each feature according to each of the correlation values, and then corresponding data are extracted from the preprocessed data set according to each of the related features to construct the state data sequence.

[0008] The embodiment of the present application provides a method for obtaining a state data sequence, which obtains the real-time operation data, terminal feature data, and time feature data of each power terminal device in different ways, realizes the effective fusion of multi-dimensional data, and further solves the noise problem in the power data through operations such as missing value filling and outlier removal, thereby improving the accuracy of subsequent credibility assessment. In addition, this embodiment also performs Pearson feature screening on the preprocessed data set, retains key features through correlation analysis, reduces redundant dimensions, reduces computational overhead while retaining core information, and improves the efficiency of subsequent credibility assessment of various power terminal devices.

[0009] In one possible implementation, the evaluation model performs feature embedding on each piece of device status data in the status data sequence to obtain embedded representation data for each piece of device status data, including: Capturing the sequential characteristics of each piece of data in the state data sequence based on sine-cosine function position coding to obtain position embedded information of each piece of device state data in the state data sequence; Extracting each time data corresponding to each time feature data in the state data sequence, and mapping each time data to a vector space of a preset dimension by position encoding, thereby constructing time embedding information for each device state data in the state data sequence; Extracting each feature data corresponding to each real-time operation data and each terminal feature data in the state data sequence, and constructing a feature scalar for each device state data in the state data sequence; The position embedding information, time embedding information and characteristic scalar of each piece of device status data are combined to obtain embedding representation data corresponding to each piece of device status data.

[0010] An embodiment of the present application provides a feature embedding method. In the position embedding process, the absolute position relationship of the data sequence is captured through sine and cosine coding, avoiding the sequential dependency limitations of traditional RNNs and enhancing spatial positioning capabilities. In the time embedding process, the time feature is mapped into a vector through position coding to model the time dependency of the device state and reflect the change trend of the device over time. Finally, the position embedding information, time embedding information and feature scalars of each device state data are combined to form a high-dimensional representation of embedded representation data. By fusing information of different granularities, the classification accuracy of the subsequent model is improved, thereby improving the accuracy of the credibility assessment of various power terminal devices.

[0011] In one possible implementation, the self-attention learning mechanism generates a corresponding attention score matrix according to each of the embedded representation data, including: Multiplying the preset query weight matrix, feature weight matrix and information weight matrix with each of the embedded representation data respectively to obtain a query vector matrix, a feature vector matrix and an information vector matrix of each of the embedded representation data; Inputting each of the query vector matrix, the feature vector matrix, and the information vector matrix into a preset SoftMax function so that the SoftMax function generates an attention score matrix corresponding to each of the embedded representation data; The query weight matrix, feature weight matrix and information weight matrix are obtained through model training.

[0012] An embodiment of the present application provides a method for generating a corresponding attention score matrix based on embedded representation data. The method dynamically calculates the correlation strength between features by calculating the query vector matrix, feature vector matrix and information vector matrix (Query, Key, Value matrix) of the embedded representation data, captures the current key equipment status of the power terminal equipment through the self-attention mechanism, and then calculates the attention score of each feature based on the three vector matrices through the SoftMax function to generate an attention score matrix, quantify the contribution of different features, and improve the model interpretability and evaluation accuracy.

[0013] Furthermore, when the NAT Decoder decoder performs data mapping on any attention score matrix to generate the credibility assessment result of the corresponding power terminal device, the NAT Decoder decoder simultaneously performs data mapping on each of the attention score matrices in a parallel operation manner to generate the credibility assessment result of each power terminal device, including: Performing a weighted sum operation on the attention score matrix and the information vector matrix corresponding to the power terminal device to obtain a comprehensive feature vector; Inputting the comprehensive feature vector into a preset classifier or regressor, performing data mapping according to a preset decision boundary, and generating a credibility assessment result of the corresponding power terminal equipment; The classifier, regressor and decision boundary are obtained through model training.

[0014] This embodiment of the present application provides a method for generating corresponding credibility assessment results based on an attention score matrix. Before performing data mapping, a weighted summation operation is performed on the attention score matrix and the information vector matrix to obtain a comprehensive feature vector. This operation avoids over-reliance on a single feature to generate credibility assessment results and improves the stability of the assessment results through multi-dimensional feature complementarity. During the data mapping process, a decision boundary that is dynamically adjusted in advance through supervised learning is introduced to assist in data mapping. This allows the model to adapt to the differences in credibility thresholds of different devices, thereby improving the accuracy of credibility assessments for various power terminal devices.

[0015] In one possible implementation, the power terminal equipment credibility assessment method further includes, after obtaining each attention score matrix, performing a sparse operation on each of the attention score matrices to generate corresponding sparse matrices; The NAT Decoder is used to perform data mapping on each of the sparse matrices to generate a credibility evaluation result for each power terminal device.

[0016] Furthermore, performing a sparse operation on each of the attention score matrices to generate corresponding sparse matrices includes: Using a preset global sparsification function to perform a sparsification operation on each of the attention score matrices to generate corresponding sparse matrices; In the sparsification operation, for any of the attention score matrices, each element in the attention score matrix is traversed and the value of each element is compared with the adaptive threshold. If the value of the current element is greater than the adaptive threshold, the value of the current element is retained, otherwise the current element is marked as noise. After completing the traversal of the attention score matrix, the corresponding sparse matrix is constructed based on the non-noise elements in the attention score matrix.

[0017] The Transformer model learns the global dependency information of the power terminal equipment data sequence through self-attention, but it is also affected by context-independent information, which reduces the accuracy of the prediction. To solve this problem, the embodiment of the present application introduces a sparse attention mechanism in the credibility assessment process. After performing a sparse operation on each attention score matrix and then performing data mapping, the key information in the power terminal equipment data sequence is retained, the impact of noise information is reduced, and the efficiency and accuracy of the credibility assessment of various power terminal equipment are improved.

[0018] In one possible implementation, the evaluation model is obtained by training an initial evaluation model in a supervised learning manner based on historical status data of each power terminal device, including: Collecting a number of historical status data from each power terminal device, including historical operation data, historical terminal characteristic data, and historical time characteristic data; According to the actual operating conditions and safety standards of each power terminal device, the historical status data are annotated, a credibility evaluation value corresponding to each historical status data is determined, and a training data set is constructed; Constructing an initial evaluation model based on a Transformer-CNN hybrid architecture and initializing model parameters of the initial evaluation model, including a query weight matrix, a feature weight matrix, an information weight matrix, position encoding parameters, and weight parameters and bias parameters of each network layer; According to the training data set and the preset loss function, the initial evaluation model is iteratively trained for multiple rounds using a gradient descent algorithm, and the model parameters of the initial evaluation model are updated during the training process to finally obtain the evaluation model.

[0019] In a second aspect, an embodiment of the present application provides a power terminal equipment credibility assessment system based on multi-source data fusion, including an acquisition module, an assessment module, and a decoding module; The acquisition module is used to acquire a status data sequence, wherein each device status data in the status data sequence corresponds to an electric power terminal device, and the device status data includes real-time operation data, terminal characteristic data, and time characteristic data of the corresponding electric power terminal device; The evaluation module is configured to input the state data sequence into a preset evaluation model, so that the evaluation model performs feature embedding on each device state data in the state data sequence, obtains embedded representation data for each device state data, and generates a corresponding attention score matrix based on each embedded representation data based on a self-attention learning mechanism, wherein the embedded representation data includes position embedding information, time embedding information, and feature scalars; The decoding module is used to simultaneously perform data mapping on each of the attention score matrices in a parallel operation manner through a NAT Decoder decoder to generate a credibility assessment result for each power terminal device; Among them, the evaluation model is obtained by training the initial evaluation model in a supervised learning manner based on the historical status data of each power terminal device, and the initial evaluation model is constructed based on the Transformer-CNN hybrid architecture.

[0020] In one possible implementation, the power terminal equipment credibility assessment system further includes a sparseness module, which is configured to perform a sparseness operation on each attention score matrix after obtaining each attention score matrix to generate each corresponding sparse matrix; The decoding module performs data mapping on each of the sparse matrices through a NAT Decoder to generate a credibility assessment result for each power terminal device. BRIEF DESCRIPTION OF THE DRAWINGS

[0021] Figure 1 A flowchart of a method for evaluating the credibility of power terminal equipment based on multi-source data fusion provided in an embodiment of the present application; Figure 2 A schematic diagram of the structure of a NAT Decoder in a method for credibility assessment of power terminal equipment based on multi-source data fusion provided in an embodiment of the present application; Figure 3 A schematic diagram of a specific implementation process of a method for evaluating the credibility of power terminal equipment based on multi-source data fusion provided in an embodiment of the present application; Figure 4 A schematic diagram of a process for sparse processing in a method for credibility assessment of power terminal equipment based on multi-source data fusion provided in an embodiment of the present application; Figure 5 A schematic structural diagram of a power terminal equipment credibility assessment system based on multi-source data fusion provided in an embodiment of the present application. DETAILED DESCRIPTION

[0022] The following will be combined with the drawings in the embodiments of this application to clearly and completely describe the technical solutions in the embodiments of this application. Obviously, the embodiments described are only part of the embodiments of this application, not all of the embodiments. Based on the embodiments in this application, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of this application.

[0023] It should be noted that the step numbers herein are for convenience of explanation of the specific embodiments and do not serve to define the order in which the steps are to be performed. In the description of this application, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of technical features indicated. Therefore, a feature designated "first" or "second" may explicitly or implicitly include one or more of such features.

[0024] Example 1: like Figure 1 As shown, the first embodiment provides a method for evaluating the credibility of power terminal equipment based on multi-source data fusion, including steps S1-S3: Step S1: Acquire a status data sequence, where each piece of device status data in the status data sequence corresponds to a power terminal device, and the device status data includes real-time operation data, terminal characteristic data, and time characteristic data of the corresponding power terminal device; Step S2: inputting the state data sequence into a preset evaluation model, so that the evaluation model performs feature embedding on each device state data in the state data sequence to obtain embedded representation data for each device state data, and generating a corresponding attention score matrix based on each embedded representation data based on a self-attention learning mechanism, wherein the embedded representation data includes position embedding information, time embedding information, and feature scalars; Step S3: Using a NAT Decoder, data mapping is performed on each of the attention score matrices simultaneously in a parallel operation manner to generate a credibility assessment result for each power terminal device. Among them, the evaluation model is obtained by training the initial evaluation model in a supervised learning manner based on the historical status data of each power terminal device, and the initial evaluation model is constructed based on the Transformer-CNN hybrid architecture.

[0025] The embodiment of the present application provides a method for evaluating the credibility of power terminal equipment based on multi-source data fusion, which forms a state data sequence with a multi-dimensional feature combination by integrating real-time operation data, terminal features and time features. Then, an evaluation model based on a Transformer-CNN hybrid architecture is used to perform feature embedding and self-attention learning on the state data sequence to generate several attention score matrices, wherein the Transformer architecture can use the self-attention mechanism to capture the temporal dependency and global correlation of the power equipment state, and the CNN architecture is used to extract local features and enhance sensitivity to sudden events. The Transformer-CNN hybrid architecture combines the advantages of both, which can model long-term dependencies and capture local details, improve the adaptability of the model to complex power data, and improve the accuracy of credibility evaluation. Finally, the NAT Decoder decoder is used to parallel map the attention score matrix and obtain the credibility evaluation results of each power terminal device at the same time, avoiding the delay problem of traditional serial decoding, meeting the real-time evaluation requirements of massive equipment, and improving the efficiency of credibility evaluation of various power terminal devices.

[0026] Furthermore, in step S1, the acquiring of the state data sequence includes: The preset monitoring system collects real-time current data, real-time voltage data, real-time power consumption and environmental status information of each power terminal device to build real-time operating data of each power terminal device; Obtaining corresponding data upload frequency, data upload success rate, data upload delay, business traffic ratio, digital certificate information, and historical credibility evaluation results from each of the power terminal devices to construct terminal feature data of each power terminal device; Based on a preset time alignment algorithm, aligning each of the real-time operation data with a preset timestamp according to the collection time of each of the real-time operation data, and then constructing time feature data corresponding to the real-time operation data according to the timestamp; Performing data cleaning, data segmentation, missing value processing, feature construction, and data normalization on each of the real-time operation data, terminal feature data, and time feature data to obtain a preprocessed data set; The correlation values between the various features in the preprocessed data set are calculated by the Pearson correlation coefficient method, and a number of related features are determined from each feature according to each of the correlation values, and then corresponding data are extracted from the preprocessed data set according to each of the related features to construct the state data sequence.

[0027] The embodiment of the present application provides a method for obtaining a state data sequence, which obtains the real-time operation data, terminal feature data, and time feature data of each power terminal device in different ways, realizes the effective fusion of multi-dimensional data, and further solves the noise problem in the power data through operations such as missing value filling and outlier removal, thereby improving the accuracy of subsequent credibility assessment. In addition, this embodiment also performs Pearson feature screening on the preprocessed data set, retains key features through correlation analysis, reduces redundant dimensions, reduces computational overhead while retaining core information, and improves the efficiency of subsequent credibility assessment of various power terminal devices.

[0028] In a preferred embodiment, real-time operational data includes power parameter data such as current, voltage, and power usage. This data reflects the power consumption and transmission status of the device during operation. For example, a smart meter collects real-time user power usage data, and a smart transformer monitors its own current and voltage values. Furthermore, it may include environmental status information, such as temperature, humidity, and other environmental factors, as environmental conditions may affect device operation. For example, high temperatures may affect device heat dissipation, thereby affecting its performance and reliability.

[0029] Terminal feature data primarily includes operational characteristics of power terminal equipment, such as data upload frequency, success rate, latency, business traffic share, digital certificate information, and historical credibility assessment results. The frequency, success rate, and latency of data uploads reflect the quality and stability of data during transmission and processing. Frequent interruptions or excessive latency in data uploads can impact device credibility assessments. Business traffic share reflects the device's share of the overall system traffic, reflecting its level of business activity and importance. Digital certificate information is used to verify device identity and data security, ensuring the data source is reliable and has not been tampered with. Historical credibility assessment results, which are past evaluations of device credibility, can provide a reference for current assessments and help the model better learn about changing trends in device credibility.

[0030] Temporal feature data mainly includes time information such as years, months, and holidays. Since some data, such as environmental status information and power consumption, are significantly affected by time, collecting this time data can help the model consider the role of time factors when analyzing device status, thereby improving the accuracy of credibility evaluation.

[0031] The collected power terminal equipment data is preprocessed, and data cleaning, data segmentation, missing value processing, feature construction, feature selection, data normalization, etc. are performed according to the characteristics of the power terminal data set. Among them, the Pearson correlation coefficient method is used to calculate the correlation between each feature and the target value, and then feature selection is performed based on the calculation results. This can reduce the difficulty and time of model training. The Pearson correlation coefficient calculation formula is as follows: Wherein, X and Y represent two different variables. In this embodiment, A data sequence that can represent a device characteristic (such as current, voltage, etc.), The data sequence represents the preset target value (comprehensive indicator related to device reliability). The linear correlation between the device characteristics and the target value is measured by calculating the correlation coefficient between the two. n represents the data sequence and The number of data points in the equation is called the sample size. It reflects the size of the data used to calculate the correlation coefficient. The number of samples will affect the accuracy and reliability of the correlation coefficient calculation. and The data series are and The data points. In practical applications, for different device characteristics and target value data sequences, and will correspond to specific observation values.

[0032] In one possible implementation, in step S2, the evaluation model performs feature embedding on each piece of device status data in the status data sequence to obtain embedded representation data of each piece of device status data, including: Capturing the sequential characteristics of each piece of data in the state data sequence based on sine-cosine function position coding to obtain position embedded information of each piece of device state data in the state data sequence; Extracting each time data corresponding to each time feature data in the state data sequence, and mapping each time data to a vector space of a preset dimension by position encoding, thereby constructing time embedding information for each device state data in the state data sequence; Extracting each feature data corresponding to each real-time operation data and each terminal feature data in the state data sequence, and constructing a feature scalar for each device state data in the state data sequence; The position embedding information, time embedding information and characteristic scalar of each piece of device status data are combined to obtain embedding representation data corresponding to each piece of device status data.

[0033] An embodiment of the present application provides a feature embedding method. In the position embedding process, the absolute position relationship of the data sequence is captured through sine and cosine coding, avoiding the sequential dependency limitations of traditional RNNs and enhancing spatial positioning capabilities. In the time embedding process, the time feature is mapped into a vector through position coding to model the time dependency of the device state and reflect the change trend of the device over time. Finally, the position embedding information, time embedding information and feature scalars of each device state data are combined to form a high-dimensional representation of embedded representation data. By fusing information of different granularities, the classification accuracy of the subsequent model is improved, thereby improving the accuracy of the credibility assessment of various power terminal devices.

[0034] Position embedding information is used to identify the relative position of data within a sequence. When collecting data from power terminal equipment, each data point is in a specific order. For example, for data collected over time, such as current and voltage, these data are sequenced according to the order in which they were collected. The order of each data point in the sequence is its position index. This index reflects the order of the data and provides the model with information about the temporal or logical order of the data, enabling the model to analyze based on the data's sequential characteristics and explore temporal dependencies between data points.

[0035] Time embedding information is obtained by converting time-related data during device data collection, primarily including time information such as years, months, and holidays. Because some device data (such as environmental status information and electricity consumption) is significantly affected by time, this time information is crucial for understanding how device status changes. During model processing, this time information is embedded, forming the model input along with position embedding and feature scalars. This conversion may involve encoding the time information, such as mapping it to a vector space of a specific dimension using a method similar to position encoding, or constructing a time feature matrix based on the time information. This allows the model to learn the time-related characteristics of the data, enabling more accurate assessment of device credibility.

[0036] Feature scalars are obtained by converting real-time operational data and terminal feature data. During the data preprocessing phase, these raw data undergo cleaning, missing value processing, feature construction and selection, and data normalization to convert them into feature scalars suitable for model input. For example, the Pearson correlation coefficient method is used to calculate the correlation between each feature and the target value. Key features are selected, and then the minimum-maximum normalization method is used to scale the data to the range of 0-1. The final feature scalar is obtained, providing the model with key information for determining device reliability.

[0037] In a preferred embodiment, the implementation formula of the sine and cosine function position encoding is as follows: Among them, i is the position index of a single power terminal equipment data in the sequence, which is mainly used to distinguish the calculation of different dimensions in the position coding. ) uses the sine function to calculate the positional encoding value, in odd dimensions ( ) is calculated using the cosine function; The dimension of semantic embedding of device data determines the complexity of encoding and representing data by the CNN-Transformer model. It can provide richer feature representation capabilities, but it will also increase the amount of computation and the complexity of model training; pos represents the absolute position of the data in the sequence.

[0038] In the task of evaluating the credibility of power terminal equipment, in addition to position embedding information, time embedding information and feature scalars also need to be considered. Therefore, the final embedding representation data consists of position embedding information, feature scalars, and time embedding. The formula is as follows: in To balance the size factor between feature variables and position embedding and time embedding, is the feature scalar, PE is the position embedding information, and TE is the time embedding information.

[0039] In one possible implementation, in step S2, the self-attention learning mechanism generates a corresponding attention score matrix according to each of the embedded representation data, including: Multiplying the preset query weight matrix, feature weight matrix and information weight matrix with each of the embedded representation data respectively to obtain a query vector matrix, a feature vector matrix and an information vector matrix of each of the embedded representation data; Inputting each of the query vector matrix, the feature vector matrix, and the information vector matrix into a preset SoftMax function so that the SoftMax function generates an attention score matrix corresponding to each of the embedded representation data; The query weight matrix, feature weight matrix and information weight matrix are obtained through model training.

[0040] An embodiment of the present application provides a method for generating a corresponding attention score matrix based on embedded representation data. The method dynamically calculates the correlation strength between features by calculating the query vector matrix, feature vector matrix and information vector matrix (Query, Key, Value matrix) of the embedded representation data, captures the current key equipment status of the power terminal equipment through the self-attention mechanism, and then calculates the attention score of each feature based on the three vector matrices through the SoftMax function to generate an attention score matrix, quantify the contribution of different features, and improve the model interpretability and evaluation accuracy.

[0041] In a preferred embodiment, it is assumed that the power terminal equipment data sequence is , whose embedding is expressed as ,in It is the embedding representation of a piece of power terminal equipment data, which is synthesized by position embedding, time embedding and feature scalar embedding. Q, K and V are used to represent the query, key and value vector matrices of the power terminal equipment data sequence respectively. The calculation formula for the Query, Key, and Value vector matrices is as follows: in, is the query weight matrix, which maps the input feature vector to the query space to generate the query vector matrix In the data processing scenario of power terminal equipment, The original features (such as current, voltage, etc.) collected by the device can be linearly transformed to obtain the query vector used to query other data point information. Multiply them together to get their own query vectors, which represent the direction and intensity of each data point's "query" for other data points. For example, for the current data collected at a certain moment, after After the transformation, the obtained Query vector can be used to measure the correlation between the current data at this moment and the data of various devices at other moments.

[0042] It is the feature (Key) weight matrix, which is used to map the input feature vector to the Key space and generate the feature vector matrix . A unique feature representation is generated for each data point as the "identity" of the data point. In the self-attention mechanism, the query vector is similar to the key vector to determine the degree of association between different data points.

[0043] It is the information (Value) weight matrix, which is used to map the input feature vector to the Value space and generate the information vector matrix . Contains the actual information content of each element in the data sequence. When the model determines the association between data points through the calculation of Query and Key vectors, it will use these associations to Get the corresponding information.

[0044] In obtaining 、 、 back, The similarity between the key calculation items and all power terminal equipment data is calculated, and then the calculation results are scaled to avoid the gradient vanishing problem caused by excessive similarity values. Finally, the final output is obtained through the SoftMax layer. The calculation formula of the Self Attention score matrix SC is as follows: in, Indicates the dimension of the Key vector. When scaling the result, divide by The purpose of this scaling operation is to avoid the vanishing gradient problem caused by excessively large similarity values. In deep learning, vanishing gradients can make model training difficult and prevent effective parameter updates. This scaling operation ensures the stability and effectiveness of the self-attention mechanism during the computational process, enabling the model to better learn relationships between data. The function is used to convert the input value into a probability distribution. In this formula, Function pair The results are processed so that the sum of the output results is 1, thereby obtaining the relative attention weights between each data point and other data points. These weights reflect the importance of different data points in the comprehensive evaluation.

[0045] Furthermore, in step S3, when the NAT Decoder performs data mapping on any attention score matrix to generate a credibility evaluation result of the corresponding power terminal device, the NAT Decoder performs data mapping on each of the attention score matrices simultaneously in a parallel operation manner to generate a credibility evaluation result of each power terminal device, including: Performing a weighted sum operation on the attention score matrix and the information vector matrix corresponding to the power terminal device to obtain a comprehensive feature vector; Inputting the comprehensive feature vector into a preset classifier or regressor, performing data mapping according to a preset decision boundary, and generating a credibility assessment result of the corresponding power terminal equipment; The classifier, regressor and decision boundary are obtained through model training.

[0046] This embodiment of the present application provides a method for generating corresponding credibility assessment results based on an attention score matrix. Before performing data mapping, a weighted summation operation is performed on the attention score matrix and the information vector matrix to obtain a comprehensive feature vector. This operation avoids over-reliance on a single feature to generate credibility assessment results and improves the stability of the assessment results through multi-dimensional feature complementarity. During the data mapping process, a decision boundary that is dynamically adjusted in advance through supervised learning is introduced to assist in data mapping. This allows the model to adapt to the differences in credibility thresholds of different devices, thereby improving the accuracy of credibility assessments for various power terminal devices.

[0047] In a preferred embodiment, NAT Decoder (Non-Autoregressive Transformer Decoder) breaks through the traditional Transformer step-by-step decoding model. In the traditional model, the decoder needs to use the output of the previous step as input to generate the next prediction result in sequence, which is inefficient. However, NAT Decoder takes advantage of parallel computing and, when processing power terminal equipment data sequences, it integrates the feature representation processed by the Transformer model through the self-attention mechanism and the sparse attention mechanism. Figure 2 As shown, the NAT Decoder has an internal parallel processing structure that can simultaneously calculate and convert information from multiple locations. It directly maps input feature representations to multiple device credibility evaluation results according to specific rules and algorithms. This mapping process may involve multiple linear transformations, activation function operations, and weighted feature combinations. For example, input features are linearly transformed using a set of weight matrices, and then the data distribution is adjusted using an activation function. Finally, the processed data is converted into corresponding credibility evaluation results based on different evaluation dimensions and thresholds. This allows for the simultaneous output of multiple evaluation results, significantly improving evaluation efficiency.

[0048] The reliability assessment results primarily refer to the equipment reliability rating. These can be discrete levels, such as high reliability, medium reliability, and low reliability, or specific numerical scores, such as a reliability score from 0 to 100, with higher scores indicating higher equipment reliability. These evaluation results comprehensively reflect the equipment's integrity, availability, and confidentiality, helping power system operators assess equipment reliability and make appropriate decisions, such as maintenance or replacement.

[0049] In one possible implementation, the power terminal equipment credibility assessment method further includes, after obtaining each attention score matrix, performing a sparse operation on each of the attention score matrices to generate corresponding sparse matrices; The NAT Decoder decoder performs data mapping on each of the sparse matrices to generate a credibility evaluation result for each power terminal device.

[0050] Furthermore, performing a sparse operation on each of the attention score matrices to generate corresponding sparse matrices includes: Using a preset global sparsification function to perform a sparsification operation on each of the attention score matrices to generate corresponding sparse matrices; In the sparsification operation, for any of the attention score matrices, each element in the attention score matrix is traversed and the value of each element is compared with the adaptive threshold. If the value of the current element is greater than the adaptive threshold, the value of the current element is retained, otherwise the current element is marked as noise. After completing the traversal of the attention score matrix, the corresponding sparse matrix is constructed based on the non-noise elements in the attention score matrix.

[0051] The Transformer model learns the global dependency information of the power terminal equipment data sequence through self-attention, but it is also affected by context-independent information, which reduces the accuracy of the prediction. To solve this problem, the embodiment of the present application introduces a sparse attention mechanism in the credibility assessment process. After performing a sparse operation on each attention score matrix and then performing data mapping, the key information in the power terminal equipment data sequence is retained, the impact of noise information is reduced, and the efficiency and accuracy of the credibility assessment of various power terminal equipment are improved.

[0052] In a preferred embodiment, Figure 3 As shown in Figure 2, after the terminal device status data is preprocessed, feature embedded, and the attention score matrix SC is generated based on the self-attention mechanism, the attention score matrix is uniformly processed in the global sparsification module. The global sparsification module introduces the global sparsification function and adaptive parameters , the attention score matrix SC is sparsely processed to retain the key information in the power terminal equipment data sequence and reduce the influence of noise information. The expression of the sparse processing is as follows: in is an adaptive parameter. When the attention score is greater than When the attention score is less than When , it means that this part of the information may be noise. The specific process of sparse processing is as follows Figure 4 As shown, in Figure 4 In [1], Q, K, and V represent the query vector, key vector, and value vector, respectively. F represents a linear transformation operation. The scale layer is used to calculate the attention score matrix SC, and the globally sparse layer is used to perform sparse processing on the attention score matrix. After sparse processing, the sparse attention score matrix is obtained, which is input into the SoftMax function for normalization and then calculated with the Value vector matrix V. The specific calculation formula is as follows: Through this calculation, the final attention representation is obtained, which retains important contextual information and eliminates the influence of noise. It provides high-quality feature information for the subsequent use of NAT Decoder for device credibility evaluation, enabling the model to more accurately evaluate the credibility of power terminal equipment.

[0053] In one possible implementation, the evaluation model is obtained by training an initial evaluation model in a supervised learning manner based on historical status data of each power terminal device, including: Collecting a number of historical status data from each power terminal device, including historical operation data, historical terminal characteristic data, and historical time characteristic data; According to the actual operating conditions and safety standards of each power terminal device, the historical status data are annotated, a credibility evaluation value corresponding to each historical status data is determined, and a training data set is constructed; Constructing an initial evaluation model based on a Transformer-CNN hybrid architecture and initializing model parameters of the initial evaluation model, including a query weight matrix, a feature weight matrix, an information weight matrix, position encoding parameters, and weight parameters and bias parameters of each network layer; According to the training data set and the preset loss function, the initial evaluation model is iteratively trained for multiple rounds using a gradient descent algorithm, and the model parameters of the initial evaluation model are updated during the training process to finally obtain the evaluation model.

[0054] In a preferred embodiment, the specific training process of the model includes the following steps: 1. Data collection and annotation Multi-source data is collected from various types of power terminal equipment, covering device collection characteristics (such as current, voltage, power consumption, and environmental parameters), data processing characteristics (data upload frequency, success rate, and latency), device characteristics (traffic traffic share, digital certificate information, and historical evaluation ratings), and time-related data (year, month, and holidays). The collected data is annotated based on the actual operating conditions and safety standards of the equipment, and the true value of the device credibility corresponding to each data sample is determined to construct a training dataset.

[0055] 2. Data preprocessing Data cleaning: Remove noise, outliers, and duplicate data from the data. For example, delete obviously erroneous current and voltage data caused by sensor failure during the acquisition process, as well as duplicated data points.

[0056] Missing value processing: Missing values in the data are handled using methods such as mean value filling and the K-nearest neighbor method. For example, if the power consumption data for a certain period is missing, it can be filled based on the average power consumption of adjacent time periods or the power consumption data of similar devices.

[0057] Feature construction and selection: The Pearson correlation coefficient is used to calculate the correlation between each feature and the target value (device credibility) to screen out key features. For example, the correlation coefficient between features such as current and voltage and device credibility is calculated. Features with high correlation are retained, while redundant features with low correlation are removed to reduce model training complexity.

[0058] Data normalization: Use the min-max normalization method to scale data to the range of 0-1, eliminating the dimensionality effects between different features and improving the convergence speed of model training. For example, electricity consumption data can be converted according to the min-max normalization formula to keep it within a uniform numerical range.

[0059] 3. Model initialization Construct a model structure based on CNN-Transformer and initialize the model parameters, including the weight matrix corresponding to the Query, Key, and Value vector matrices ( 、 、 ), positional encoding parameters, and weights and biases of other network layers.

[0060] 4. Model training The preprocessed data is input into the CNN-Transformer model. After feature embedding, attention score matrix calculation, sparsification operation and NAT Decoder decoding, the model outputs the corresponding credibility assessment results.

[0061] 5. Model training optimization Define the loss function: Select an appropriate loss function, such as the mean square error (MSE) loss function, to measure the difference between the device credibility predicted by the model and the actual labeled value. The formula is: ,in is the sample size, is the real device credibility value, The device credibility value predicted by the model.

[0062] Select an optimization algorithm: Use a gradient descent algorithm (such as the Adam optimizer) to adjust the parameters in the model (such as 、 、 The weight matrix and other network layer parameters are calculated so that the loss function value gradually decreases. In each iteration, the optimizer updates the parameters based on the gradient information. The formula is: ,in is the current parameter value, is the learning rate, is the gradient of the loss function with respect to the parameters.

[0063] Training and Validation: Multiple rounds of training are performed. In each round, the model updates its parameters based on the training data. Simultaneously, a validation set (a portion of the training data that is not used for training) is used to evaluate model performance. The loss function and other evaluation metrics (such as accuracy and recall) are monitored to prevent overfitting. If the model's performance on the validation set stops improving, training can be stopped early to save the optimal model parameters.

[0064] 6. Model Evaluation Use a test set (data that does not overlap with the training or validation sets) to comprehensively evaluate the trained model. Calculate metrics such as the model's accuracy, recall, and F1 score on the test set to assess the accuracy and reliability of the model's device credibility assessment. Based on the evaluation results, analyze the model's strengths and weaknesses, and adjust and optimize the model structure, parameter settings, or data processing methods to further improve model performance.

[0065] In summary, compared with the prior art, the embodiments of the present application have the following beneficial technical effects: This embodiment of the present application discloses a method for assessing the credibility of power terminal equipment based on multi-source data fusion. First, the source data obtained from the equipment is preprocessed, including data cleaning, missing value processing, feature construction, feature selection, and data normalization. The relevant features of the preprocessed data are then input into a Transformer-CNN hybrid model to model the device status data sequence. The features are then processed using self-attention and sparse attention mechanisms to retain key information. A NAT decoder is then used to improve the speed of device credibility assessment. This embodiment comprehensively assesses multiple aspects of device status information, introducing a Transformer to extract global information, a CNN to capture local temporal features, and a sparsification function to retain important contextual information and eliminate noise. By training the Transformer-CNN hybrid model, the model learns the characteristic relationships between various device status information and device credibility assessments, allowing the model to be used for credibility assessment. This effectively addresses the significant uncertainty and dynamic nature of equipment in power systems. Furthermore, the model is highly scalable and adaptable, capable of addressing emerging demands in power systems and enabling continuous optimization of credibility assessments for a large number of terminal devices.

[0066] Example 2: like Figure 5 As shown, the second embodiment provides a power terminal equipment credibility assessment system based on multi-source data fusion, including an acquisition module 10, an assessment module 20 and a decoding module 30; The acquisition module 10 is used to acquire a status data sequence, wherein each device status data in the status data sequence corresponds to a power terminal device, and the device status data includes real-time operation data, terminal feature data, and time feature data of the corresponding power terminal device; The evaluation module 20 is configured to input the state data sequence into a preset evaluation model, so that the evaluation model performs feature embedding on each device state data in the state data sequence, obtains embedded representation data for each device state data, and generates a corresponding attention score matrix based on each embedded representation data based on a self-attention learning mechanism, wherein the embedded representation data includes position embedding information, time embedding information, and feature scalars; The decoding module 30 is used to simultaneously perform data mapping on each of the attention score matrices in a parallel operation manner through a NAT Decoder to generate a credibility assessment result for each power terminal device; Among them, the evaluation model is obtained by training the initial evaluation model in a supervised learning manner based on the historical status data of each power terminal device, and the initial evaluation model is constructed based on the Transformer-CNN hybrid architecture.

[0067] Furthermore, the acquisition module 10 acquires the state data sequence, including: The preset monitoring system collects real-time current data, real-time voltage data, real-time power consumption and environmental status information of each power terminal device to build real-time operating data of each power terminal device; Obtaining corresponding data upload frequency, data upload success rate, data upload delay, business traffic ratio, digital certificate information, and historical credibility evaluation results from each of the power terminal devices to construct terminal feature data of each power terminal device; Based on a preset time alignment algorithm, aligning each of the real-time operation data with a preset timestamp according to the collection time of each of the real-time operation data, and then constructing time feature data corresponding to the real-time operation data according to the timestamp; Performing data cleaning, data segmentation, missing value processing, feature construction, and data normalization on each of the real-time operation data, terminal feature data, and time feature data to obtain a preprocessed data set; The correlation values between the various features in the preprocessed data set are calculated by the Pearson correlation coefficient method, and a number of related features are determined from each feature according to each of the correlation values, and then corresponding data are extracted from the preprocessed data set according to each of the related features to construct the state data sequence.

[0068] In one possible implementation, the evaluation model performs feature embedding on each piece of device status data in the status data sequence to obtain embedded representation data for each piece of device status data, including: Capturing the sequential characteristics of each piece of data in the state data sequence based on sine-cosine function position coding to obtain position embedded information of each piece of device state data in the state data sequence; Extracting each time data corresponding to each time feature data in the state data sequence, and mapping each time data to a vector space of a preset dimension by position encoding, thereby constructing time embedding information for each device state data in the state data sequence; Extracting each feature data corresponding to each real-time operation data and each terminal feature data in the state data sequence, and constructing a feature scalar for each device state data in the state data sequence; The position embedding information, time embedding information and characteristic scalar of each piece of device status data are combined to obtain embedding representation data corresponding to each piece of device status data.

[0069] In one possible implementation, the self-attention learning mechanism generates a corresponding attention score matrix according to each of the embedded representation data, including: Multiplying the preset query weight matrix, feature weight matrix and information weight matrix with each of the embedded representation data respectively to obtain a query vector matrix, a feature vector matrix and an information vector matrix of each of the embedded representation data; Inputting each of the query vector matrix, the feature vector matrix, and the information vector matrix into a preset SoftMax function so that the SoftMax function generates an attention score matrix corresponding to each of the embedded representation data; The query weight matrix, feature weight matrix and information weight matrix are obtained through model training.

[0070] Furthermore, when the NAT Decoder is used to perform data mapping on any attention score matrix to generate a credibility assessment result of the corresponding power terminal device, the decoding module 30 simultaneously performs data mapping on each of the attention score matrices in a parallel operation manner through the NAT Decoder to generate a credibility assessment result of each power terminal device, including: Performing a weighted sum operation on the attention score matrix and the information vector matrix corresponding to the power terminal device to obtain a comprehensive feature vector; Inputting the comprehensive feature vector into a preset classifier or regressor, performing data mapping according to a preset decision boundary, and generating a credibility assessment result of the corresponding power terminal equipment; The classifier, regressor and decision boundary are obtained through model training.

[0071] In one possible implementation, the power terminal equipment credibility assessment system further includes a sparseness module, which is configured to perform a sparseness operation on each attention score matrix after obtaining each attention score matrix to generate each corresponding sparse matrix; The decoding module performs data mapping on each of the sparse matrices through a NAT Decoder to generate a credibility assessment result for each power terminal device.

[0072] Furthermore, the sparsification module performs a sparsification operation on each of the attention score matrices to generate corresponding sparse matrices, including: Using a preset global sparsification function to perform a sparsification operation on each of the attention score matrices to generate corresponding sparse matrices; In the sparsification operation, for any of the attention score matrices, each element in the attention score matrix is traversed and the value of each element is compared with the adaptive threshold. If the value of the current element is greater than the adaptive threshold, the value of the current element is retained, otherwise the current element is marked as noise. After completing the traversal of the attention score matrix, the corresponding sparse matrix is constructed based on the non-noise elements in the attention score matrix.

[0073] In one possible implementation, the power terminal device credibility assessment system further includes a model training module, which is configured to obtain the assessment model by training an initial assessment model in a supervised learning manner based on historical status data of each power terminal device, and includes an acquisition unit, a data set construction unit, a model initialization unit, and a training unit; The acquisition unit is used to collect a number of historical status data from each power terminal device, including historical operation data, historical terminal feature data and historical time feature data; The data set construction unit is used to perform data annotation on the plurality of historical status data according to the actual operating conditions and safety standards of each power terminal device, determine the credibility evaluation value corresponding to each historical status data, and construct a training data set; The model initialization unit is used to construct an initial evaluation model based on the Transformer-CNN hybrid architecture and initialize the model parameters of the initial evaluation model, wherein the model parameters include a query weight matrix, a feature weight matrix, an information weight matrix, position encoding parameters, and weight parameters and bias parameters of each network layer; The training unit is used to perform multiple rounds of iterative training on the initial evaluation model using a gradient descent algorithm based on the training data set and a preset loss function, and to update the model parameters of the initial evaluation model during the training process to finally obtain the evaluation model.

[0074] The embodiment of the present application provides a credibility assessment system for power terminal equipment based on multi-source data fusion, which forms a state data sequence of multi-dimensional feature combination by integrating real-time operation data, terminal features and time features. Then, an evaluation model based on the Transformer-CNN hybrid architecture is used to perform feature embedding and self-attention learning on the state data sequence to generate several attention score matrices, wherein the Transformer architecture can use the self-attention mechanism to capture the temporal dependency and global correlation of the power equipment state, and the CNN architecture is used to extract local features and enhance the sensitivity to sudden events. The Transformer-CNN hybrid architecture combines the advantages of both, which can model long-term dependencies and capture local details, improve the adaptability of the model to complex power data, and improve the accuracy of credibility assessment. Finally, the attention score matrix is mapped in parallel through the NAT Decoder decoder, and the credibility assessment results of each power terminal device are obtained at the same time, avoiding the delay problem of traditional serial decoding, meeting the real-time assessment requirements of massive equipment, and improving the efficiency of credibility assessment of various power terminal devices.

[0075] The more detailed working principle and process flow of this embodiment can be referred to, but not limited to, the relevant records of the first embodiment.

[0076] The specific embodiments described above further illustrate the objectives, technical solutions, and beneficial effects of this application. It should be understood that the above descriptions are merely specific embodiments of this application and are not intended to limit the scope of protection of this application. In particular, it should be noted that any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of this application by those skilled in the art should be included within the scope of protection of this application.

Claims

1. A method for evaluating the credibility of power terminal equipment based on multi-source data fusion, characterized in that: include: Acquire a status data sequence, where each piece of device status data in the status data sequence corresponds to a power terminal device, and the device status data includes real-time operation data, terminal characteristic data, and time characteristic data of the corresponding power terminal device; Inputting the state data sequence into a preset evaluation model so that the evaluation model performs feature embedding on each device state data item in the state data sequence to obtain embedded representation data for each device state data item, and generating a corresponding attention score matrix based on each embedded representation data item based on a self-attention learning mechanism, wherein the embedded representation data includes position embedding information, time embedding information, and feature scalars; The NAT Decoder decodes the attention score matrices in parallel to generate a credibility assessment result for each power terminal device. Among them, the evaluation model is obtained by training the initial evaluation model in a supervised learning manner based on the historical status data of each power terminal device, and the initial evaluation model is constructed based on the Transformer-CNN hybrid architecture.

2. The method for evaluating the reliability of power terminal equipment based on multi-source data fusion according to claim 1, characterized in that: The acquisition state data sequence includes: The preset monitoring system collects real-time current data, real-time voltage data, real-time power consumption and environmental status information of each power terminal device to build real-time operating data of each power terminal device; Obtaining corresponding data upload frequency, data upload success rate, data upload delay, business traffic ratio, digital certificate information, and historical credibility evaluation results from each of the power terminal devices to construct terminal feature data of each power terminal device; Based on a preset time alignment algorithm, aligning each of the real-time operation data with a preset timestamp according to the collection time of each of the real-time operation data, and then constructing time feature data corresponding to the real-time operation data according to the timestamp; Performing data cleaning, data segmentation, missing value processing, feature construction, and data normalization on each of the real-time operation data, terminal feature data, and time feature data to obtain a preprocessed data set; The correlation values between the various features in the preprocessed data set are calculated by the Pearson correlation coefficient method, and a number of related features are determined from each feature according to each of the correlation values, and then corresponding data are extracted from the preprocessed data set according to each of the related features to construct the state data sequence.

3. The method for evaluating the credibility of power terminal equipment based on multi-source data fusion according to claim 1, characterized in that: The evaluation model performs feature embedding on each piece of device status data in the status data sequence to obtain embedded representation data of each piece of device status data, including: Capturing the sequential characteristics of each piece of data in the state data sequence based on sine-cosine function position coding to obtain position embedded information of each piece of device state data in the state data sequence; Extracting each time data corresponding to each time feature data in the state data sequence, and mapping each time data to a vector space of a preset dimension by position encoding, thereby constructing time embedding information for each device state data in the state data sequence; Extracting each feature data corresponding to each real-time operation data and each terminal feature data in the state data sequence, and constructing a feature scalar for each device state data in the state data sequence; The position embedding information, time embedding information and characteristic scalar of each piece of device status data are combined to obtain embedding representation data corresponding to each piece of device status data.

4. The method for evaluating the credibility of power terminal equipment based on multi-source data fusion according to claim 1, characterized in that: The self-attention learning mechanism generates a corresponding attention score matrix according to each of the embedded representation data, including: Multiplying the preset query weight matrix, feature weight matrix and information weight matrix with each of the embedded representation data respectively to obtain a query vector matrix, a feature vector matrix and an information vector matrix of each of the embedded representation data; Inputting each of the query vector matrix, the feature vector matrix, and the information vector matrix into a preset SoftMax function so that the SoftMax function generates an attention score matrix corresponding to each of the embedded representation data; The query weight matrix, feature weight matrix and information weight matrix are obtained through model training.

5. The method for evaluating the credibility of power terminal equipment based on multi-source data fusion according to claim 4, characterized in that: When data mapping is performed on any attention score matrix by the NAT Decoder to generate a credibility evaluation result of the corresponding power terminal device, the NAT Decoder simultaneously performs data mapping on each of the attention score matrices in a parallel operation manner to generate a credibility evaluation result of each power terminal device, including: Performing a weighted sum operation on the attention score matrix and the information vector matrix corresponding to the power terminal device to obtain a comprehensive feature vector; Inputting the comprehensive feature vector into a preset classifier or regressor, performing data mapping according to a preset decision boundary, and generating a credibility assessment result of the corresponding power terminal equipment; The classifier, regressor and decision boundary are obtained through model training.

6. The method for evaluating the credibility of power terminal equipment based on multi-source data fusion according to claim 1, characterized in that: The power terminal equipment credibility assessment method further includes, after obtaining each attention score matrix, performing a sparse operation on each of the attention score matrices to generate corresponding sparse matrices; The NAT Decoder decoder performs data mapping on each of the sparse matrices to generate a credibility evaluation result for each power terminal device.

7. The method for evaluating the credibility of power terminal equipment based on multi-source data fusion according to claim 6, characterized in that: The performing a sparse operation on each of the attention score matrices to generate corresponding sparse matrices includes: Using a preset global sparsification function to perform a sparsification operation on each of the attention score matrices to generate corresponding sparse matrices; In the sparsification operation, for any of the attention score matrices, each element in the attention score matrix is traversed and the value of each element is compared with the adaptive threshold. If the value of the current element is greater than the adaptive threshold, the value of the current element is retained, otherwise the current element is marked as noise. After completing the traversal of the attention score matrix, the corresponding sparse matrix is constructed based on the non-noise elements in the attention score matrix.

8. A method for evaluating the credibility of power terminal equipment based on multi-source data fusion according to any one of claims 1 to 7, characterized in that: The evaluation model is obtained by training the initial evaluation model in a supervised learning manner based on the historical status data of each power terminal device, including: Collecting a number of historical status data from each power terminal device, including historical operation data, historical terminal characteristic data, and historical time characteristic data; According to the actual operating conditions and safety standards of each power terminal device, the historical status data are annotated, a credibility evaluation value corresponding to each historical status data is determined, and a training data set is constructed; Constructing an initial evaluation model based on a Transformer-CNN hybrid architecture and initializing model parameters of the initial evaluation model, including a query weight matrix, a feature weight matrix, an information weight matrix, position encoding parameters, and weight parameters and bias parameters of each network layer; According to the training data set and the preset loss function, the initial evaluation model is iteratively trained for multiple rounds using a gradient descent algorithm, and the model parameters of the initial evaluation model are updated during the training process to finally obtain the evaluation model.

9. A power terminal equipment credibility assessment system based on multi-source data fusion, characterized in that: It includes acquisition module, evaluation module and decoding module; The acquisition module is used to acquire a status data sequence, wherein each device status data in the status data sequence corresponds to an electric power terminal device, and the device status data includes real-time operation data, terminal characteristic data, and time characteristic data of the corresponding electric power terminal device; The evaluation module is configured to input the state data sequence into a preset evaluation model, so that the evaluation model performs feature embedding on each device state data in the state data sequence, obtains embedded representation data for each device state data, and generates a corresponding attention score matrix based on each embedded representation data based on a self-attention learning mechanism, wherein the embedded representation data includes position embedding information, time embedding information, and feature scalars; The decoding module is used to simultaneously perform data mapping on each of the attention score matrices in a parallel operation manner through a NAT Decoder decoder to generate a credibility assessment result for each power terminal device; Among them, the evaluation model is obtained by training the initial evaluation model in a supervised learning manner based on the historical status data of each power terminal device, and the initial evaluation model is constructed based on the Transformer-CNN hybrid architecture.

10. The power terminal equipment credibility assessment system based on multi-source data fusion according to claim 9, characterized in that: The power terminal equipment credibility assessment system further includes a sparseness module, which is configured to perform a sparseness operation on each attention score matrix after obtaining each attention score matrix to generate each corresponding sparse matrix; The decoding module performs data mapping on each of the sparse matrices through a NAT Decoder to generate a credibility assessment result for each power terminal device.

Images