An artificial intelligence-based device state prediction and risk assessment method

Abstract

The application provides a device state prediction method and device based on artificial intelligence, comprising: obtaining the operation log of a new energy device, using a principal component analysis model to extract and label the feature of the denoised device operation signal data; based on a generative adversarial network, the labeled characteristic value data is expanded, and a non-parametric verification method is used to verify the expanded data to obtain training data; a long short-term memory network (LSTM) based on an attention mechanism is used to train a new energy device operation state intelligent prediction model to obtain a trained operation state intelligent prediction model; finally, the model output value is determined to realize accurate prediction of the state of the electrical equipment in a substation and risk assessment. The principal component method is used for dimension reduction to reduce the number of evaluation indexes, and the long short-term memory network (LSTM) based on the attention mechanism is used to improve the accuracy and efficiency of the device health degree evaluation; the generalization ability and prediction performance of the model are enhanced; and the stability and reliability of the model are improved.

Description

Technical Field

[0001] This application relates to the field of power maintenance technology, and in particular to an artificial intelligence-based method for equipment condition prediction and risk assessment. Background Technology

[0002] With the rapid development of the power industry, the operating environment of power facilities is becoming increasingly complex, and equipment failures and defects pose a severe challenge to the stability and safety of these facilities. In particular, critical facilities such as high-voltage transmission lines, substations, and power generation equipment, once they fail, can not only affect the stable operation of the power system but also lead to widespread power outages, and even cause casualties and property damage. Therefore, how to promptly detect potential failures and defects during equipment operation has become an urgent problem for the power industry to solve.

[0003] Currently, traditional power facility inspection methods often follow fixed schedules or rely on operator experience, depending heavily on manual inspections and periodic checks. This approach is not only labor-intensive but also limited by the frequency and efficiency of manual inspections, making it difficult to detect minor defects in a timely manner. Furthermore, this method fails to accurately reflect the actual usage and wear levels of equipment, potentially leading to premature or delayed maintenance, thus increasing unnecessary maintenance costs or the risk of unexpected failures. In addition, traditional methods lack continuous monitoring and trend prediction of equipment operating status, making it difficult to meet the increasingly complex needs of power facility management.

[0004] With the development of industrial internet and big data technologies, data-driven equipment health assessments are becoming increasingly possible. By collecting and analyzing the vast amounts of data generated during equipment operation, its health status can be assessed more accurately, enabling predictive maintenance, optimizing resource allocation, reducing unexpected downtime, and improving overall production efficiency.

[0005] However, existing methods for predicting the health status of power facilities suffer from several technical problems. Firstly, insufficient feature extraction from operational data results in incomplete and inefficient data features, and some important information is easily lost, affecting the accuracy and reliability of power facility health status prediction. Secondly, inaccurate predictions of power facility health status lead to improper inspection and management, increasing safety hazards.

[0006] Therefore, it is particularly important to invent an artificial intelligence-based method for predicting equipment status and assessing risks to accurately predict equipment status. Summary of the Invention

[0007] In view of this, it is necessary to provide an artificial intelligence-based method for predicting equipment status and assessing risks, which can at least overcome one of the above-mentioned shortcomings.

[0008] This invention provides an artificial intelligence-based method and apparatus for predicting equipment status, comprising: acquiring the operation logs of new energy equipment; using a principal component analysis model to extract and label features from the denoised equipment operation signal data; expanding the labeled feature data based on a generative adversarial network; using a non-parametric verification method to verify the expanded data to obtain training data; training an intelligent prediction model for the operating status of new energy equipment using a Long Short-Term Memory (LSTM) network based on an attention mechanism to obtain a trained intelligent prediction model for operating status; and finally determining the model output value to achieve accurate prediction of the status of electrical equipment in substations and to conduct risk assessment. The use of principal component analysis to reduce the number of evaluation indicators and the use of an LSTM model based on an attention mechanism improve the accuracy and efficiency of equipment health assessment; enhance the model's generalization ability and predictive performance; and improve the model's stability and reliability.

[0009] In a first aspect, embodiments of this application provide a method for predicting device status and assessing risks based on artificial intelligence, the method comprising:

[0010] The operation logs of the new energy equipment are obtained, the operation logs are preprocessed, and the denoised equipment operation signal data are extracted and labeled using a principal component analysis model to obtain labeled feature value data that characterizes the equipment's operating status and health status.

[0011] The labeled feature data is augmented using generative adversarial networks. The augmented data is then validated using a nonparametric verification method, and both the augmented and original data are selected as training data.

[0012] A smart prediction model for the operating status of new energy equipment is established using a Long Short-Term Memory (LSTM) network based on an attention mechanism. The training data is used as the input vector to the model. Through model learning and training, the output value of the model is determined, and the trained smart prediction model for the operating status is obtained.

[0013] The features extracted in real time are used as input vectors into the model to determine the model output value, thereby enabling accurate prediction of the status of electrical equipment in substations and risk assessment.

[0014] In one embodiment, the preprocessing of the runtime log includes:

[0015] Remove non-character tags and stop words from the runtime log, keeping only key characters, and split compound tags into individual words according to the camelCase naming convention.

[0016] In one embodiment, the step of using principal component analysis to extract and label features from the denoised device operating signal data includes:

[0017] Obtain the preprocessed runtime logs and map the runtime log dataset to a high-dimensional space;

[0018] In the high-dimensional space, the high-dimensional data is projected onto a low-dimensional space with the same dimensions as the original by using PCA transformation to reduce the dimensionality of the runtime log dataset, and finally the reconstructed dataset is obtained.

[0019] The degree of anomaly for each sample is determined by calculating the reconstruction error. The reconstruction error E s The formula for calculating (x) is:

[0020]

[0021] Where d is the preset dimension. Let C represent the centroid of the i-th cluster after clustering, where i and j represent the number of clusters and the dimension count, respectively. i Let i represent the i-th family, and k represent the number of clusters.

[0022] In one embodiment, the augmentation of labeled feature data based on generative adversarial networks and the verification of the augmented data using a nonparametric validation method include:

[0023] Generative adversarial networks are used to augment and enhance log data;

[0024] The KS checksum method is used to filter and expand the data. The calculation method of the KS checksum method is D. iss as follows:

[0025]

[0026] Among them, F n (X) represents the empirical distribution function of sample X, F m (Y) represents the empirical distribution function of sample Y, and n and m represent the number of samples X and Y, respectively;

[0027] When D iss If the data is within the preset threshold range, the qualified augmented data is retained; otherwise, it is discarded.

[0028] In one embodiment, the generative adversarial network is constructed in the following manner:

[0029] The generator deceives the discriminator by producing defects that resemble actual sample data;

[0030] The discriminator extracts features of abnormal data by introducing a self-attention mechanism into a convolutional layer, and then classifies the abnormal data features.

[0031] The discriminator loss function is obtained by judging the supervisory loss function of the discriminator through cross-entropy.

[0032] The network model is trained using a generative adversarial approach between the generator and the discriminator, and the loss function is optimized; thus, the generative adversarial network model is improved.

[0033] In one embodiment, the step of establishing an intelligent prediction model for the operating status of new energy equipment using a Long Short-Term Memory (LSTM) network based on an attention mechanism includes:

[0034] S1, Input Layer: Standardizes the labeled feature data according to the preset dimensional data to obtain input data with multiple dimensions;

[0035] S2, Word Vector Layer: The input data is transformed into a sequence of word vectors using a vector machine model;

[0036] S3, the first LSTM layer: transforms the word vector sequence into high-dimensional features, calculates the candidate memory information of the word vector sequence at the current time through the input gate, obtains the hidden state through the calculation results of the forget gate and the input gate, and outputs the feature information of the word vector sequence.

[0037] S4, Sentence Vector Layer: The word vector sequence feature information is weighted by global information to obtain the sentence vector sequence;

[0038] S5, the second LSTM layer: transforms the sentence vector sequence into high-dimensional features, calculates the candidate memory information of the sentence vector sequence at the current time through the input gate, obtains the hidden state through the calculation results of the forget gate and the input gate, and outputs the feature information of the sentence vector sequence.

[0039] S6, Attention Layer: Perform self-attention calculation on the sentence vector feature sequence to obtain self-attention features;

[0040] S7, Feature Fusion Layer: Self-attention features are concatenated and fused to obtain enhanced features;

[0041] S8, Output Layer: The enhanced features are input into the fully connected layer. The fully connected layer integrates the enhanced feature information and maps the multi-dimensional feature information into a two-dimensional feature input to obtain the prediction result of the substation electrical equipment status at the next moment.

[0042] In one embodiment, the standardization process of the labeled feature value data based on preset dimensional data includes:

[0043] The standardization process for labeled feature value data based on preset dimensional data includes:

[0044]

[0045] Where X represents the original data. The mean of the original data, σ is the standard deviation of the original data.

[0046] In one embodiment, in step S6, the loss function L com for:

[0047] L com =λ1l1+λ2l2,

[0048]

[0049] in y represents the model output value, t and T represent the number of real samples, and λ1 and λ2 represent the weight coefficients.

[0050] In one embodiment, the fusion operation in step S7 is as follows:

[0051] The channel is then upgraded after dimensionality reduction using a fully connected layer;

[0052] Channel multiplication is performed with shallow features, and then fusion with deep features to obtain channel alignment, resulting in the aligned fused feature F. m ;

[0053] Features of fusion F m The calculation formula is:

[0054]

[0055] Among them, A v () indicates the calculation of the pooling average, M a () indicates the calculation of the global pooling maximum value, where w1 and w2 represent the weights of adjacent connection layers, and F h F1 represents the deep features, F1 represents the initial features, and δ represents the activation function.

[0056] Secondly, embodiments of this application provide an artificial intelligence-based device for predicting and assessing equipment status, applied to the artificial intelligence-based device status prediction and risk assessment method described in the first aspect, the device comprising:

[0057] The data acquisition module is used to acquire the operation logs of new energy equipment, preprocess the operation logs, and use a principal component analysis model to extract and label the features of the denoised equipment operation signal data to obtain labeled feature value data that characterizes the equipment's operating status and health status.

[0058] The data processing module is used to augment labeled feature data based on generative adversarial networks, verify the augmented data using nonparametric verification methods, and select the augmented data and the original data as training data.

[0059] The model building module is used to establish an intelligent prediction model for the operating status of new energy equipment using a Long Short-Term Memory (LSTM) network based on an attention mechanism. The training data is input into the model as the model input vector. Through model learning and training, the model output value is determined, and the trained intelligent prediction model for the operating status is obtained.

[0060] The status prediction module is used to input the features extracted in real time into the model as input vectors, determine the model output value, realize accurate prediction of the status of electrical equipment in substations, and conduct risk assessment.

[0061] Thirdly, embodiments of this application provide an electronic device, including:

[0062] processor;

[0063] Memory used to store processor-executable instructions;

[0064] The processor is configured to implement the AI-based device state prediction and risk assessment method as described in the first aspect when executing the instructions.

[0065] Fourthly, embodiments of this application provide a computer-readable storage medium including instructions that instruct a device to perform the artificial intelligence-based device state prediction and risk assessment method as described in the first aspect.

[0066] This invention provides an artificial intelligence-based method and apparatus for predicting equipment status, comprising: acquiring the operation logs of new energy equipment; using a principal component analysis model to extract and label features from the denoised equipment operation signal data; expanding the labeled feature data based on a generative adversarial network; using a non-parametric verification method to verify the expanded data to obtain training data; training an intelligent prediction model for the operating status of new energy equipment using a Long Short-Term Memory (LSTM) network based on an attention mechanism to obtain a trained intelligent prediction model for operating status; and finally determining the model output value to achieve accurate prediction of the status of electrical equipment in substations and to conduct risk assessment. The use of principal component analysis to reduce the number of evaluation indicators and the use of an LSTM model based on an attention mechanism improve the accuracy and efficiency of equipment health assessment; enhance the model's generalization ability and predictive performance; and improve the model's stability and reliability. Attached Figure Description

[0067] Figure 1 This is a schematic diagram of a device status prediction and risk assessment method based on artificial intelligence provided in an embodiment of this application.

[0068] Figure 2 A schematic diagram of the structure of a Long Short-Term Memory (LSTM) network based on an attention mechanism provided in another embodiment of this application.

[0069] Figure 3 This is a schematic diagram of the execution flow of an attention-based Long Short-Term Memory (LSTM) network model provided in another embodiment of this application.

[0070] Figure 4 A schematic diagram of an AI-based device status prediction and risk assessment device module provided in an embodiment of this application.

[0071] Figure 5 A schematic diagram of an electronic device provided in an embodiment of this application. Detailed Implementation

[0072] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of this application, and not all of them.

[0073] It should be noted that, in the embodiments of this application, "at least one" refers to one or more, and "more than one" refers to two or more. Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this application belongs. The terminology used in the specification of this application is for the purpose of describing particular embodiments only and is not intended to be limiting of this application.

[0074] It should be noted that in the embodiments of this application, the terms "first," "second," etc., are used only for descriptive purposes and should not be construed as indicating or implying relative importance, nor as indicating or implying order. Features specified as "first" or "second" may explicitly or implicitly include one or more of the stated features. In the description of the embodiments of this application, words such as "exemplary" or "for example" are used to indicate examples, illustrations, or explanations. Any embodiment or design scheme described as "exemplary" or "for example" in the embodiments of this application should not be construed as being more preferred or advantageous than other embodiments or design schemes. Specifically, the use of words such as "exemplary" or "for example" is intended to present the relevant concepts in a concrete manner.

[0075] Based on the embodiments described in this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0076] With the rapid development of the power industry, the operating environment of power facilities is becoming increasingly complex, and equipment failures and defects pose a severe challenge to the stability and safety of these facilities. In particular, critical facilities such as high-voltage transmission lines, substations, and power generation equipment, once they fail, can not only affect the stable operation of the power system but also lead to widespread power outages, and even cause casualties and property damage. Therefore, it is crucial to develop an artificial intelligence-based method for equipment condition prediction and risk assessment to accurately predict equipment conditions.

[0077] Currently, traditional methods for inspecting power facilities largely rely on manual inspections and periodic checks. This approach is not only labor-intensive but also limited by the frequency and efficiency of manual inspections, making it difficult to detect minor defects in a timely manner. Furthermore, traditional methods lack continuous monitoring and trend prediction of equipment operating status, making it difficult to meet the increasingly complex needs of power facility management.

[0078] In view of this, this application provides an artificial intelligence-based method for predicting equipment status and assessing risks, which can accurately predict equipment status and perform risk assessment.

[0079] Figure 1 This is a schematic diagram of a device status prediction and risk assessment method based on artificial intelligence provided in an embodiment of this application. Figure 1 The AI-based equipment status prediction and risk assessment method shown includes at least the following steps:

[0080] S110, acquire the operation log of the new energy equipment, preprocess the operation log, use principal component analysis model to extract and label the features of the denoised equipment operation signal data, and obtain labeled feature value data that characterizes the equipment operation status and health status.

[0081] In this embodiment of the application, step S110 preprocesses the running log, including: removing non-character tags and stop words from the running log, retaining only key characters, and splitting composite tags into individual words according to the principle of camelCase naming.

[0082] Specifically, the purpose of log analysis data preprocessing can also include filtering non-compliant data, cleaning up meaningless data, and performing format conversion and standardization to filter and separate basic data for different themes according to subsequent needs. Data preprocessing can be achieved through MapReduce programs. For example, clickstream models are used to generate pageviews and visit information tables to facilitate subsequent statistical analysis.

[0083] Understandably, CamelCase naming convention is a set of naming rules (conventions) used in computer programming. As its name suggests, CamelCase refers to the use of mixed uppercase and lowercase letters to form variable and function names. Programmers often adopt a consistent and readable naming convention to facilitate communication among peers.

[0084] Specifically, the main steps of traditional principal component analysis are: selecting initial variables based on the research question; determining whether principal component analysis is appropriate (KMO and Bartlett tests); calculating the correlation coefficient matrix, eigenvalues ​​and eigenvectors, and calculating the variance contribution rate and cumulative variance contribution rate; selecting principal components; interpreting the principal components; and further analysis after principal component analysis.

[0085] These are the basic steps of principal component analysis. It's important to note that different software and tools may have slightly different implementation details, but the concepts are similar.

[0086] It is understood that the use of principal component analysis model to extract and label features from denoised equipment operating signal data in this embodiment includes:

[0087] Obtain the preprocessed runtime logs and map the runtime log dataset to a high-dimensional space;

[0088] In the high-dimensional space, the high-dimensional data is projected onto a low-dimensional space with the same dimensions as the original by using PCA transformation to reduce the dimensionality of the runtime log dataset, and finally the reconstructed dataset is obtained.

[0089] The degree of anomaly for each sample is determined by calculating the reconstruction error. The reconstruction error E s The formula for calculating (x) is:

[0090]

[0091] Where d is the preset dimension. Let C represent the centroid of the i-th cluster after clustering, where i and j represent the number of clusters and the dimension count, respectively. i Let i represent the i-th family, and k represent the number of clusters.

[0092] S112, based on generative adversarial networks, augments the labeled feature data, uses a non-parametric verification method to verify the augmented data, and selects the augmented data and the original data as training data.

[0093] The main purpose of principal component analysis (PCA) is to combine many variables in the original data and reduce them to a small group of new variables that can express all the information in the original data, thereby effectively describing the overall internal structure of the original data.

[0094] The following are the effective results of principal component analysis:

[0095] (1) Data dimensionality reduction: Principal component analysis transforms the original data into a set of new variables, namely principal components, through linear transformation. These new variables can explain most of the variation in the original data, thus achieving data dimensionality reduction.

[0096] (2) Improve data representation: Obtain higher representation accuracy from the original variables, especially when the amount of data is large.

[0097] (3) Identify the latent structure in the data: Principal component analysis can identify the latent structure in the data and reduce the number of variables while retaining as much information as possible.

[0098] (4) Simplify the data analysis process: Principal component analysis simplifies the data analysis process by extracting the principal components in the data.

[0099] Therefore, the effect of principal component analysis is to help us better understand the intrinsic structure of data by reducing dimensionality, improving data representation, identifying the underlying structure in the data, and simplifying the data analysis process.

[0100] In this embodiment of the application, the augmentation of labeled feature data based on generative adversarial networks and the verification of the augmented data using a nonparametric validation method include:

[0101] Generative adversarial networks are used to augment and enhance log data;

[0102] The KS checksum method is used to filter and expand the data. The calculation method of the KS checksum method is D. iss as follows:

[0103]

[0104] Among them, F n (X) represents the empirical distribution function of sample X, F m (Y) represents the empirical distribution function of sample Y, and n and m represent the number of samples X and Y, respectively;

[0105] When D issWithin a preset threshold range, qualified augmented data is retained; otherwise, it is discarded. It's understandable that Generative Adversarial Networks (GANs) are deep learning models. The core idea of ​​GANs is to generate data through adversarial training, making the generated samples as close as possible to real samples.

[0106] Understandably, the generator's task is to generate realistic data samples. It accepts random noise as input and attempts to generate a sample that looks real (such as an image or text). The discriminator's task is to distinguish whether the input sample comes from a real dataset or is a fake sample generated by the generator. It outputs a probability value indicating whether the input sample is real or generated. The training process of a GAN is a game-theoretic process, with the generator and discriminator competing against each other.

[0107] Understandably, the generator attempts to generate fake samples that can deceive the discriminator, making it difficult for the discriminator to distinguish between generated and real samples. The discriminator, on the other hand, tries to improve its discriminative ability to more accurately identify real and fake samples. Adversarial training: The training process of a GAN is a zero-sum game, with the generator and discriminator competing against each other. The generator tries to maximize the discriminator's misclassification rate for generated samples, while the discriminator tries to minimize this misclassification rate.

[0108] Specifically, the generative adversarial network is constructed in the following way:

[0109] The generator deceives the discriminator by producing defects that resemble actual sample data;

[0110] The discriminator extracts abnormal data features by introducing a self-attention mechanism into the convolutional layer, and classifies the abnormal data features.

[0111] The discriminator loss function is obtained by judging the supervisory loss function of the discriminator through cross-entropy;

[0112] The network model is trained and the loss function is optimized by using a generative adversarial approach between the generator and the discriminator.

[0113] Improve the generative adversarial network model.

[0114] Specifically, the training process of a Generative Adversarial Network (GAN) is as follows: Initialization: The generator and discriminator are initialized to random states; Generation: The generator generates a fake sample from random noise; Discrimination: The discriminator classifies the generated sample and the real sample, calculates its prediction results, and provides feedback based on its accuracy in distinguishing between the generated sample and the real sample; Optimization: The generator and discriminator update their parameters through backpropagation. The generator adjusts its generation strategy based on the feedback from the discriminator, while the discriminator improves its accuracy by distinguishing between real and generated samples; Iteration: This process is repeated until the generator can generate very realistic samples, and the discriminator can hardly distinguish between the generated samples and the real samples.

[0115] S114. An intelligent prediction model for the operating status of new energy equipment is established using a Long Short-Term Memory (LSTM) network based on an attention mechanism. The training data is used as the model input vector and input into the model. Through model learning and training, the model output value is determined, and the trained intelligent prediction model for the operating status is obtained.

[0116] Understandably, it's well-known that LSTMs don't handle long sequences and highlighting important information well, leading to poor performance in some situations. Attention mechanisms, which mimic the characteristics of human visual attention, can effectively address this problem. LSTM (Long Short-Term Memory) networks and attention mechanisms are two important techniques in deep learning; they can be combined to improve model performance, especially when processing sequential data.

[0117] Specifically, the attention mechanism determines which parts of the input sequence to focus on through weight distribution. This allows the model to dynamically adjust its focus when generating the output, in order to better capture key information in the input sequence. In this way, by combining the long-term dependency capture capability of LSTM with the dynamic focus adjustment capability of the attention mechanism, our model can more effectively handle various complex sequence processing tasks and be applied to more fields.

[0118] The core of LSTM is a cell state and three gating mechanisms (input gate, forget gate, and output gate). These gates dynamically control the flow of information by filtering it, allowing the model to selectively remember or forget certain information. Cell State: The cell state is the memory unit of the LSTM, used to carry long-term contextual information. It avoids rapid gradient vanishing through linear propagation (a small number of weighted operations and gating updates). Gating Mechanism: Each gate consists of a sigmoid activation function and some dot product operations, with an output range of [0, 1], representing the degree to which a certain portion of information passes through. Forget Gate: Controls which information needs to be deleted from the cell state. The formula is: f tThe output of the forget gate determines whether to retain previous information in the cell state. The input gate controls which new information needs to be added to the cell state. This involves two steps: determining which information needs to be updated. Generate candidate information: Final update of cell state: Output gate: Determines the hidden state (i.e., output) at the current moment. The hidden state is:

[0119] Where "*" represents convolution, used Hadama product, b f b i b c b o Let W represent the bias terms of the forget gate, input gate, candidate information, and output gate, respectively, and let W be the weight matrices of the forget gate, input gate, and output gate, respectively. xf W xi W xc and W xo These are the weight matrices for the forget gate, input gate, candidate information, and output gate, respectively. σ is the sigmoid activation function, and tanh is the hyperbolic tangent activation function. To maintain consistent output and input sizes, zero-padding is performed before the convolution operation.

[0120] The core of LSTM lies in dynamically controlling the flow of information through a gating mechanism, enabling memory and forgetting in conjunction with cell state. This is achieved through the following steps: the forget gate determines which information needs to be forgotten; the input gate determines which new information needs to be added; the cell state is updated to store the new memory; and the output gate controls the output at the current time step.

[0121] Specifically, such as Figure 2 The diagram shown is a schematic diagram of the Long Short-Term Memory (LSTM) network model structure based on the attention mechanism provided in another embodiment of this application. Figure 3 The diagram shown illustrates the execution flow of a Long Short-Term Memory (LSTM) network model based on an attention mechanism, according to another embodiment of this application. The model is divided into eight layers: input layer, word vector layer, first LSTM layer, sentence vector layer, second LSTM layer, attention layer, fusion feature layer, and output layer.

[0122] The structure and function of the intelligent prediction model for the operating status of new energy equipment, which utilizes an attention-based Long Short-Term Memory (LSTM) network, include:

[0123] S1, Input Layer: Standardizes the labeled feature data according to the preset dimensional data to obtain input data with multiple dimensions;

[0124] S2, Word Vector Layer: The input data is transformed into a sequence of word vectors using a vector machine model;

[0125] S3, the first LSTM layer: transforms the word vector sequence into high-dimensional features, calculates the candidate memory information of the word vector sequence at the current time through the input gate, obtains the hidden state through the calculation results of the forget gate and the input gate, and outputs the feature information of the word vector sequence.

[0126] S4, Sentence Vector Layer: The word vector sequence feature information is weighted by global information to obtain the sentence vector sequence;

[0127] S5, the second LSTM layer: transforms the sentence vector sequence into high-dimensional features, calculates the candidate memory information of the sentence vector sequence at the current time through the input gate, obtains the hidden state through the calculation results of the forget gate and the input gate, and outputs the feature information of the sentence vector sequence.

[0128] S6, Attention Layer: Perform self-attention calculation on the sentence vector feature sequence to obtain self-attention features;

[0129] S7, Feature Fusion Layer: Self-attention features are concatenated and fused to obtain enhanced features;

[0130] S8, Output Layer: The enhanced features are input into the fully connected layer. The fully connected layer integrates the enhanced feature information and maps the multi-dimensional feature information into a two-dimensional feature input to obtain the prediction result of the substation electrical equipment status at the next moment.

[0131] Specifically, the standardization process for labeled feature value data based on preset dimensional data includes:

[0132] The standardization process for labeled feature value data based on preset dimensional data includes:

[0133]

[0134] Where X represents the original data. The mean of the original data, σ is the standard deviation of the original data.

[0135] Specifically, in step S6, the loss function L com for:

[0136] L com =λ1l1+λ2l2,

[0137]

[0138] in y represents the model output value, t and T represent the number of real samples, and λ1 and λ2 represent the weight coefficients.

[0139] Specifically, in step S7, the steps of the fusion operation are as follows:

[0140] The channel is then upgraded after dimensionality reduction using a fully connected layer;

[0141] Channel multiplication is performed with shallow features, and then fusion with deep features to obtain channel alignment, resulting in the aligned and fused feature F. m ;

[0142] Features of fusion F m The calculation formula is:

[0143]

[0144] Among them, A v () indicates the calculation of the pooling average, M a () indicates the calculation of the global pooling maximum value, where w1 and w2 represent the weights of adjacent connection layers, and F h F1 represents the deep features, F1 represents the initial features, and δ represents the activation function.

[0145] Fully connected layers have too many parameters and are computationally too expensive. However, after convolution and pooling, the size of a CNN is greatly reduced, and only one or two fully connected layers are used at the end, making the computational cost within an acceptable range.

[0146] S116 is used to input the features extracted in real time into the model as the model input vector, determine the model output value, realize accurate prediction of the status of electrical equipment in substations, and conduct risk assessment.

[0147] Specifically, it can realize real-time monitoring of equipment operating status, release alarm and early warning information, and grasp and judge the equipment operating status in real time; realize the collection and fusion of equipment status-related status data, and conduct in-depth mining and analysis of multi-dimensional big data; realize the summary and analysis of equipment operation problems, provide support for the formulation of equipment operation and maintenance strategies and anti-accident measures, provide decision support for production command, and effectively guide the development of equipment maintenance, technical transformation and other work.

[0148] In this embodiment, the device system aims to guide equipment condition-based maintenance and full lifecycle management. Based on comprehensive status information of substation equipment and centered on an application model derived from big data analysis, it expands its functions in multiple dimensions around defect / fault diagnosis and prediction, providing significant guidance for the operation and maintenance of substation equipment. The system involves targeted in-depth mining and application of equipment status data, primarily including: big data → typical defect database; big data → fault prediction; big data → family-related defects; big data → equipment lifespan prediction, etc.

[0149] It is understood that the AI-based equipment status prediction and risk assessment method provided in this application can achieve fully automated status prediction and risk assessment, which not only improves the efficiency of power facility maintenance but also reduces equipment failures and downtime through early warning. Furthermore, by integrating various advanced AI technologies and data processing methods, the system can process and analyze various complex information about power facilities, greatly improving the accuracy and reliability of fault prediction and reducing safety accidents and economic losses caused by equipment failures.

[0150] Figure 4 This is a schematic diagram of an AI-based device status prediction and risk assessment module provided in another embodiment of this application. It is applied to applications such as… Figure 1 The method for predicting equipment status and assessing risks based on artificial intelligence is shown. Figure 4 The AI-based equipment status prediction and risk assessment device shown includes: a data acquisition module 1, a data processing module 2, a model building module 3, and a status prediction module 4. Its functions are similar to... Figure 1 The parts shown in the text are the same or similar, and will not be repeated here.

[0151] In this embodiment of the application, the data acquisition module 1 is used to acquire the operation log of the new energy equipment, preprocess the operation log, use the principal component analysis model to extract and label the features of the denoised equipment operation signal data, and obtain labeled feature value data that characterizes the equipment operation status and health status.

[0152] In this embodiment of the application, the data processing module 2 is used to augment the labeled feature data based on the generative adversarial network, verify the augmented data using a non-parametric verification method, and select the augmented data and the original data as training data.

[0153] In this embodiment, model building module 3 is used to establish an intelligent prediction model for the operating status of new energy equipment using a Long Short-Term Memory (LSTM) network based on an attention mechanism. Training data is input into the model as the model input vector. Through model learning and training, the model output value is determined, resulting in a trained intelligent prediction model for the operating status. Please refer to the following for specific acquisition methods. Figure 1-4 The details and their corresponding descriptions are not repeated here.

[0154] In this embodiment of the application, the state prediction module 4 is used to input the features extracted in real time into the model as the model input vector, determine the model output value, and realize accurate prediction of the state of electrical equipment in the substation.

[0155] Figure 5 This is an electronic device provided in one embodiment of this application. For example... Figure 5As shown, the electronic device includes at least the following components: processor 101 and memory 100, communication interface 103, and bus 102.

[0156] In this embodiment, the memory 100 is used to store executable instructions of the processor 101, which, when configured to execute instructions, implements... Figure 3 The method shown is based on artificial intelligence for predicting equipment status and assessing risks.

[0157] In embodiments of this application, a computer-readable storage medium includes instructions that instruct a device to perform the method as described in the first aspect. For example, the instructions instruct the device to perform... Figure 4 The process steps illustrate an AI-based method for predicting equipment status and assessing risks.

[0158] In one embodiment of this application, the program operating in the electronic device may be a program that controls a central processing unit (CPU) or similar device to achieve the functions of the above-described embodiments of the present invention (a program that enables the computer to function). The information processed by these devices is then temporarily stored in random access memory (RAM) during processing, and subsequently stored in various ROMs such as read-only memory (Flash ROM) and hard disk drives (HDDs), and read, corrected, and written by the CPU as needed.

[0159] It should be noted that a portion of the electronic device described in the above embodiments can also be implemented using a computer. In this case, the program for implementing the control function can be recorded on a computer-readable recording medium, and the program recorded on the recording medium can be read into the computer and executed.

[0160] It should be noted that the term "computer" as used here refers to a computer built into an electronic device, employing hardware including an operating system and peripheral devices. Furthermore, "computer-readable recording media" refers to removable media such as floppy disks, magneto-optical disks, ROMs, and CD-ROMs, as well as storage devices such as hard drives built into a computer.

[0161] Furthermore, a "computer-readable recording medium" can include: a medium that dynamically stores a program for a short period of time, such as a communication line used when transmitting a program via a network such as the Internet or a communication line such as a telephone line; or a medium that stores a program for a fixed period of time, such as volatile memory inside a computer that serves as a server or client in this case. In addition, the aforementioned program can be a program used to implement the above-mentioned functions, or it can be a program that can implement the above-mentioned functions by combining with programs already recorded in the computer.

[0162] Furthermore, the electronic device in the above embodiments can also be implemented as an assembly (device group) composed of multiple devices. Each device constituting the device group can possess some or all of the functions or functional blocks of the electronic device in the above embodiments. As a device group, it is sufficient to have all the functions or functional blocks of the electronic device.

[0163] It is understood that the AI-based equipment status prediction and risk assessment method, apparatus, electronic device, and storage medium provided in this application can effectively improve the operation and maintenance level of power facilities and solve the problem that traditional manual inspections cannot identify facility defects in real time and comprehensively. Through the application of this system, potential fault hazards in facilities can be quickly discovered, and corresponding repair suggestions and optimization schemes can be provided through intelligent analysis, thereby minimizing equipment downtime and maintenance costs. At the same time, the system's adaptive capability allows it to continuously learn and optimize, improving performance based on more real-time collected data, further enhancing its adaptability to different types of power facilities and its ability to predict future defects.

[0164] Those skilled in the art should recognize that the above embodiments are only used to illustrate this application and are not intended to limit this application. Any appropriate changes and variations made to the above embodiments within the essential spirit and scope of this application fall within the scope of protection claimed in this application.

Claims

1. A method for predicting equipment status and assessing risks based on artificial intelligence, characterized in that, The method includes: The operation logs of the new energy equipment are obtained, the operation logs are preprocessed, and the denoised equipment operation signal data are extracted and labeled using a principal component analysis model to obtain labeled feature value data that characterizes the equipment's operating status and health status. The labeled feature data is augmented based on generative adversarial networks, and the augmented data is verified using a nonparametric verification method. The augmented data and the original data are then selected as training data. A smart prediction model for the operating status of new energy equipment is established using a Long Short-Term Memory (LSTM) network based on an attention mechanism. The training data is used as the model input vector. Through model learning and training, the model output value is determined, and the trained smart prediction model for the operating status is obtained. The attention-based Long Short-Term Memory (LSTM) network model is divided into eight layers: input layer, word vector layer, first LSTM layer, sentence vector layer, second LSTM layer, attention layer, fusion feature layer, and output layer. A smart prediction model for the operating status of new energy equipment is established using an attention-based Long Short-Term Memory (LSTM) network, outlining its structure and functions, including: S1, Input Layer: Standardizes the labeled feature data according to the preset dimensional data to obtain input data with multiple dimensions; S2, Word Vector Layer: The input data is transformed into a sequence of word vectors using a vector machine model; S3, the first LSTM layer: transforms the word vector sequence into high-dimensional features, calculates the candidate memory information of the word vector sequence at the current time through the input gate, obtains the hidden state through the calculation results of the forget gate and the input gate, and outputs the feature information of the word vector sequence. S4, Sentence Vector Layer: The word vector sequence feature information is weighted by global information to obtain the sentence vector sequence; S5, the second LSTM layer: transforms the sentence vector sequence into high-dimensional features, calculates the candidate memory information of the sentence vector sequence at the current time through the input gate, obtains the hidden state through the calculation results of the forget gate and the input gate, and outputs the feature information of the sentence vector sequence. S6, Attention Layer: Perform self-attention calculation on the sentence vector feature sequence to obtain self-attention features; S7, Feature Fusion Layer: Self-attention features are concatenated and fused to obtain enhanced features; S8, Output layer: The enhanced features are input into the fully connected layer. The fully connected layer integrates the enhanced feature information and maps the multi-dimensional feature information into a two-dimensional feature input to obtain the prediction result of the substation electrical equipment status at the next moment. The standardization process for labeled feature value data based on preset dimensional data includes: ， in This is the original data. The average of the raw data, The standard deviation of the original data; In step S6, the loss function for: ， ， ， in Output values ​​for the model. The values ​​are the measured values, t=1,2,...,T are the sample indices, and T represents the number of true samples. , Indicates the weighting coefficient; In step S7, the steps of the fusion operation are as follows: The channel is then upgraded after dimensionality reduction using a fully connected layer; Channel multiplication is performed with shallow features, and then fusion is performed with deep features to obtain channel alignment, resulting in aligned and fused features. ; Features of fusion The calculation formula is: ， in, This indicates the calculation of the pooling average. This indicates the calculation of the global pooling maximum value. , These represent the weights of adjacent connection layers, Indicates deep features, Indicates initial features, Indicates the activation function; Fully connected layers have too many parameters and too much computation, but after convolution and pooling, the size of CNN is greatly reduced. Finally, using one or two fully connected layers keeps the computation within an acceptable range. The features extracted in real time are used as input vectors into the model to determine the model output value, thereby enabling accurate prediction of the status of electrical equipment in substations and risk assessment.

2. The method for predicting equipment status and assessing risks based on artificial intelligence according to claim 1, characterized in that, The preprocessing of the runtime log includes: Remove non-character tags and stop words from the runtime log, keeping only key characters, and split compound tags into individual words according to the camelCase naming convention.

3. The method for predicting equipment status and assessing risks based on artificial intelligence according to claim 1, characterized in that, The process of using principal component analysis to extract and label features from the denoised equipment operating signal data includes: Obtain the preprocessed runtime logs and map the runtime log dataset to a high-dimensional space; In the high-dimensional space, the high-dimensional data is projected onto a low-dimensional space with the same dimensions as the original by using PCA transformation to reduce the dimensionality of the runtime log dataset, and finally the reconstructed dataset is obtained. The degree of anomaly for each sample is determined by calculating the reconstruction error.

4. The method for predicting equipment status and assessing risks based on artificial intelligence according to claim 2, characterized in that, The process of augmenting labeled feature data using generative adversarial networks and then verifying the augmented data using nonparametric validation methods includes: Generative adversarial networks are used to augment and enhance log data; The KS verification method is used to filter the augmented data. If the data is within the preset threshold range after verification, the qualified augmented data is retained; otherwise, it is discarded.

5. The method for predicting equipment status and assessing risks based on artificial intelligence according to claim 4, characterized in that, The generative adversarial network is constructed in the following way: The generator deceives the discriminator by producing defects that resemble actual sample data; The discriminator extracts features of abnormal data by introducing a self-attention mechanism into a convolutional layer, and classifies the abnormal data features. The discriminator loss function is obtained by judging the supervisory loss function of the discriminator through cross-entropy; The network model is trained and the loss function is optimized by using a generative adversarial approach between the generator and the discriminator. Improve the generative adversarial network model.

6. An artificial intelligence-based equipment status prediction and risk assessment device, applied to the artificial intelligence-based equipment status prediction and risk assessment method as described in any one of claims 1 to 5, characterized in that, The device includes: The data acquisition module is used to acquire the operation logs of new energy equipment, preprocess the operation logs, use principal component analysis model to extract and label the features of the denoised equipment operation signal data, and obtain labeled feature value data that characterizes the equipment's operating status and health status. The data processing module is used to augment labeled feature data based on generative adversarial networks, verify the augmented data using nonparametric verification methods, and select the augmented data and the original data as training data. The model building module is used to build an intelligent prediction model of the operating status of new energy equipment using a Long Short-Term Memory (LSTM) network based on an attention mechanism. The training data is input into the model as the model input vector. Through model learning and training, the model output value is determined, and the trained intelligent prediction model of the operating status is obtained. The status prediction module is used to input the features extracted in real time into the model as input vectors, determine the model output value, realize accurate prediction of the status of electrical equipment in substations, and conduct risk assessment.

Images