Asset performance auxiliary evaluation method based on NLP

Abstract

The present invention relates to the technical field of power grid asset evaluation, and in particular to an asset performance auxiliary evaluation method based on NLP. By normalizing structured data, cleaning and uniformly formatting unstructured data, a context-aware word embedding model is used to convert unstructured text data into a vector form, and feature weights are dynamically adjusted in combination with the operating status and maintenance cycle of the equipment, so that the changes and risks of the equipment in different states can be reflected in real time. The emotional information in the text is analyzed through a multi-level emotional perception model, the deep correlation between the text and the equipment status is captured, and a comprehensive performance evaluation feature library is constructed; finally, based on the feature library, a deep neural network is used to perform a comprehensive evaluation of asset performance, so as to achieve more accurate risk evaluation. Through multi-source data fusion and context perception, the limitations of traditional evaluation methods are broken through, and the accuracy and comprehensiveness of power grid asset risk evaluation are significantly improved.

Description

Technical Field

[0001] The present invention relates to the technical field of power grid asset evaluation, and in particular to an asset performance auxiliary evaluation method based on NLP. Background Art

[0002] As an important part of modern social infrastructure, the power grid is responsible for the transmission and distribution of electricity, ensuring the normal operation of production and daily life in all walks of life. With the growth of electricity demand and technological progress, the challenges faced by the power grid in the operation process are becoming more and more complex, especially in terms of equipment failure, maintenance and operational efficiency. In order to ensure the stable operation of the power grid, various assets in the power grid must be efficiently managed and evaluated.

[0003] Grid assets usually include hardware equipment such as substations, transmission lines, distribution equipment, and monitoring systems. The operating status of these devices directly affects the reliability and efficiency of the grid. Therefore, asset performance evaluation occupies an important position in the grid industry. Traditional asset performance evaluation methods mainly rely on structured data, such as equipment operating data, maintenance records, and fault history. These data are analyzed manually and with simple statistical analysis tools to assess the health of equipment, warn of potential risks, and guide maintenance decisions.

[0004] However, as the scale and complexity of power grids increase, traditional asset performance evaluation methods face several limitations:

[0005] Insufficient structured data: Although structured data is very important in equipment performance analysis, it usually cannot fully reflect the operating status of the equipment, especially when the equipment has atypical failures or potential risks, structured data often finds it difficult to capture its hidden associated information.

[0006] The importance of unstructured data is underestimated: The unstructured data generated in the power grid (such as operator records, maintenance reports, equipment failure logs, user feedback, etc.) has rich contextual information and emotional content, but these data are often not effectively utilized. Traditional methods fail to fully tap the potential value behind these unstructured data.

[0007] Limitations of the assessment model: Existing asset performance assessment models mostly use a single data source or analysis method and lack a comprehensive understanding of the equipment's operating status. Especially when faced with complex equipment interactions and operating environments, it is difficult to accurately predict risks and conduct comprehensive assessments. Summary of the invention

[0008] The present invention provides an asset performance auxiliary evaluation method based on NLP, which is intended to solve the technical problem that the current use of structured data for evaluation leads to model evaluation limitations, making it difficult to accurately predict risks and conduct comprehensive evaluations.

[0009] The NLP-based asset performance auxiliary evaluation method includes the following steps:

[0010] Step 1: Obtain structured data and unstructured data in the power grid field, normalize the structured data, clean the unstructured data, and unify the format after removing irrelevant information;

[0011] Step 2: Perform text cleaning based on the processed unstructured data, and convert the text data into vector form using a context-aware word embedding model, wherein the context-aware word embedding model dynamically adjusts feature weights based on the equipment operating status and maintenance cycle;

[0012] Step 3: Capture the association between text and device status through a multi-level emotion perception model, and build a comprehensive performance evaluation feature library by fusing the output of the multi-level emotion perception model with the processed structured data;

[0013] Step 4: Use a deep neural network to perform a comprehensive evaluation of asset performance using the comprehensive feature vectors of the comprehensive performance evaluation feature library, thereby achieving asset risk assessment;

[0014] Step 5: The system provides risk assessment reports to operators based on the assessment structure to help decision makers specify subsequent asset management strategies.

[0015] The NLP-based asset performance auxiliary evaluation method provided by the present invention aims to solve the limitations faced by traditional evaluation methods based on structured data, especially the technical bottlenecks in accurately predicting risks and comprehensively evaluating them. Traditional methods mainly rely on structured data of power grid equipment (such as operating status, fault history, etc.), but these data cannot fully capture the potential risks and complex equipment behaviors of the equipment, resulting in the evaluation results often being superficial and difficult to provide accurate risk warnings and performance evaluations. The present invention solves this problem by introducing unstructured data (such as equipment failure logs, operator records, etc.) and using NLP technology for in-depth analysis. First, the structured data is normalized and the unstructured data is cleaned and uniformly formatted to ensure data quality and consistency. Then, a context-aware word embedding model is used to convert unstructured text data into vector form, and the feature weights are dynamically adjusted in combination with the operating status and maintenance cycle of the equipment, so that the changes and risks of the equipment in different states can be reflected in real time. Then, the emotional information in the text is analyzed through a multi-level emotional perception model to capture the deep association between the text and the equipment status, and this information is effectively integrated with the structured data to construct a comprehensive performance evaluation feature library. Finally, based on the feature library, a deep neural network is used to conduct a comprehensive assessment of asset performance, thereby achieving a more accurate risk assessment and providing decision makers with a scientific asset management strategy through a risk assessment report. This method breaks through the limitations of traditional assessment methods through multi-source data fusion and situational awareness, and significantly improves the accuracy and comprehensiveness of power grid asset risk assessment.

[0016] Preferably, step 2 comprises the following steps:

[0017] Input Layer:

[0018] Text input: Input the operation and maintenance text data of power grid equipment, and obtain the basic word embedding of the text through the pre-trained BERT model. Suppose the input text is T = [w1, w2, ..., w n ], each word w is transformed into i Convert to vector e(w i );

[0019] Equipment operation and maintenance status and maintenance cycle information: Assume that the equipment status is S = [s1, s2, ..., s m ], where each s i is a device status parameter, and the maintenance cycle information is M = [m1, m2, ..., m k ], indicating the maintenance cycle data of the equipment;

[0020] Context-aware embedding generation: embedding e(w i), introduce a weighted mechanism for equipment status and maintenance cycle, and adjust the word embedding according to the equipment status S and maintenance cycle M:

[0021]

[0022] Where: represents the word embedding after weighting by equipment status and maintenance cycle; W s and W m The learned weight representations are used to adjust the impact of equipment status and maintenance cycle on word embedding;

[0023] Transformer encoding layer: The context-aware embedding is input into the Transformer encoding layer of the BERT model, and the contextual relationship of each layer is captured through the self-attention mechanism to generate accurate contextual semantic representation:

[0024]

[0025] Where: H i Represents the final embedding vector of each word, which is output after being processed by the Transformer encoding layer.

[0026] Preferably, step 3 comprises the following steps:

[0027] Sentiment labeling: In each text segment, the sentiment label of the text is extracted through the sentiment analysis model. Let the output of sentiment analysis be E i ;

[0028] Sentiment-enhanced representation: based on the sentiment label E of each text i , for the semantic representation H i For enhancement, an emotional enhancement network is used to adjust the weights of semantic representations according to emotional labels:

[0029] H i,enhanced =H i +γ·f(E i );

[0030] Where: H i is the semantic representation obtained in step 2; γ represents a learnable enhancement coefficient; f(E i ) represents the emotion label E i Calculate the enhancement vector;

[0031] Device state perception layer: associate the device state information with the sentiment information of the text, embed the state information into a high-dimensional space through a multi-layer perceptron, and obtain the device state embedding representation S emb ;

[0032] Emotion and device state association layer: enhance the emotional representation of the text H i,enhanced and the device state embedding representation S emb Fusion is performed to establish the association between text sentiment and device status:

[0033] H final =α·H i,enhanced +β·S emb ;

[0034] Where: α and β are the learned weights; H final represents the final fusion feature;

[0035] Fusion with structured data:

[0036] H fusion =α′·H final +β′·D;

[0037] Where: α′ and β′ are the learned weights; D represents structured data;

[0038] Comprehensive performance evaluation feature library: By integrating the emotion perception model with the device status information, the comprehensive feature vector H is obtained. fusion .

[0039] Preferably, the calculation steps of the enhancement vector are as follows:

[0040] Emotional label embedding: embed the emotional label E i Convert to embedding vector e i ;

[0041] Constructing a self-attention layer: Fusing the sentiment tag embedding with the semantic representation of the text:

[0042] f(E i )=Self-Attention(H i , e i );

[0043] Where: H i Represents the original representation of the text; e i The embedding vector representing the sentiment tag; the Self-Attention mechanism is used to adjust the semantic representation of the text according to the embedding of the sentiment tag;

[0044] Output emotion enhancement vector: The final output f(E i ) is the weighted fusion vector of sentiment label and text representation.

[0045] Preferably, the structure of the deep neural network is as follows:

[0046] Input layer: receives the comprehensive feature vector H fusion, the dimension of the vector is d fusion ;

[0047]

[0048] The first fully connected layer: input feature vector X input , output a vector consisting of h1 neurons; use the ReLU activation function:

[0049] H1=ReLU(W1× input +b1);

[0050] Where: represents the weight matrix; represents the bias vector;

[0051] The second fully connected layer: input feature vector H1, output vector consisting of h2 neurons; use ReLU activation function:

[0052] H2=ReLU(W2H1+b2);

[0053] Where: represents the weight matrix; represents the bias vector;

[0054] The second fully connected layer: input feature vector H2, output vector consisting of h3 neurons; use ReLU activation function:

[0055] H3=ReLU(W3H2+b3);

[0056] Where: represents the weight matrix; represents the bias vector;

[0057] Output layer: Predict the risk score of an asset using a regression task:

[0058] y pred =W out H3+b out ;

[0059] Where: represents the weight matrix, Expresses paranoia.

[0060] Preferably, the loss function of the deep neural network is as follows:

[0061]

[0062] Where: N represents the number of samples; λ represents the L2 regularization coefficient; α i Represents the confidence of the i-th sample; represents the true risk score of the i-th sample; represents the risk score of the i-th sample predicted by the model; W k represents the weight matrix of the kth layer.

[0063] The beneficial effects of the present invention are embodied in:

[0064] The NLP-based asset performance auxiliary evaluation method provided by the present invention aims to solve the limitations faced by traditional evaluation methods based on structured data, especially the technical bottlenecks in accurately predicting risks and comprehensively evaluating them. Traditional methods mainly rely on structured data of power grid equipment (such as operating status, fault history, etc.), but these data cannot fully capture the potential risks and complex equipment behaviors of the equipment, resulting in the evaluation results often being superficial and difficult to provide accurate risk warnings and performance evaluations. The present invention solves this problem by introducing unstructured data (such as equipment failure logs, operator records, etc.) and using NLP technology for in-depth analysis. First, the structured data is normalized and the unstructured data is cleaned and uniformly formatted to ensure data quality and consistency. Then, a context-aware word embedding model is used to convert unstructured text data into vector form, and the feature weights are dynamically adjusted in combination with the operating status and maintenance cycle of the equipment, so that the changes and risks of the equipment in different states can be reflected in real time. Then, the emotional information in the text is analyzed through a multi-level emotional perception model to capture the deep association between the text and the equipment status, and this information is effectively integrated with the structured data to construct a comprehensive performance evaluation feature library. Finally, based on the feature library, a deep neural network is used to conduct a comprehensive assessment of asset performance, thereby achieving a more accurate risk assessment and providing decision makers with a scientific asset management strategy through a risk assessment report. This method breaks through the limitations of traditional assessment methods through multi-source data fusion and situational awareness, and significantly improves the accuracy and comprehensiveness of power grid asset risk assessment. BRIEF DESCRIPTION OF THE DRAWINGS

[0065] Figure 1 The specific flow chart provided by the present invention is as follows. DETAILED DESCRIPTION

[0066] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0067] Combine the following Figure 1 The specific implementation modes of the present invention are described in detail:

[0068] The NLP-based asset performance auxiliary evaluation method includes the following steps:

[0069] Step 1: Obtain structured data and unstructured data in the power grid field, normalize the structured data, clean the unstructured data, and unify the format after removing irrelevant information;

[0070] In the power grid field, asset performance evaluation relies on two main types of data: structured data and unstructured data.

[0071] Structured data: This type of data usually refers to data that can be represented in a table or database format, with a clear format and fields. In the field of power grids, structured data includes but is not limited to:

[0072] The operating status of the equipment (such as current, voltage, temperature, load, etc.)

[0073] Maintenance records (such as maintenance time, maintenance cycle, fault history, etc.)

[0074] Equipment life cycle data (such as equipment installation time, expected service life, etc.)

[0075] Fault data (such as fault occurrence time, type, handling method, repair time, etc.)

[0076] Unstructured data: This type of data is usually in free text form and contains rich contextual information. In the field of power grids, unstructured data mainly includes:

[0077] Fault log: detailed description of equipment failure, including observations of on-site staff, repair process, etc. Operator report: daily inspection records, description of abnormal conditions during equipment operation

[0078] User feedback: Customer feedback information about problems or failures in power grid operation

[0079] Maintenance report: detailed text description of the cause of equipment failure and the repair process

[0080] Sensor data: For example, status information recorded by a device through visual or sound sensors, which is often in the form of text or images;

[0081] Normalize structured data:

[0082] Although structured data has clear fields and formats, the value ranges and unit differences of its different features may affect the accuracy of subsequent analysis. Normalization is a method of standardizing structured data, which aims to scale each feature value to a uniform range to avoid certain data dominating the model's learning process due to scale differences.

[0083] Data cleaning for unstructured data:

[0084] Unstructured data, such as text data like fault logs and maintenance records, usually contains noise, redundant information, and content with inconsistent formats. Therefore, data cleaning is required. The goal of data cleaning is to remove these irrelevant or meaningless parts to make the data cleaner and usable for subsequent analysis.

[0085] Specific steps of data cleaning:

[0086] Removing stop words: Stop words refer to common words that are useless for analysis in text processing, such as "de", "shi", "zai", etc. In unstructured data in the power grid field, removing stop words can effectively reduce unnecessary calculations and noise.

[0087] Removing irrelevant information: Filter out information unrelated to the performance of power grid equipment through rules or predefined keywords, such as certain irrelevant operation remarks, personal information, etc. Regular expressions or predefined keyword lists can be used to identify and remove irrelevant content.

[0088] Correcting spelling mistakes: There may be spelling mistakes or different spelling forms in texts such as equipment names and fault descriptions, which need to be uniformly corrected and standardized. For example, unify "substation" and "transformer substation" into a standard term.

[0089] Standardizing time formats: There may be different time formats in logs (such as "2023-12-09" and "12 / 09 / 2023"), which need to be unified into a standard time format for convenient subsequent analysis.

[0090] Word segmentation and stemming: Chinese texts usually need to be segmented, and English texts can be stemmed (for example, unify "maintenance", "maintaining" into "maintain").

[0091] Unifying the format after removing irrelevant information:

[0092] After data cleaning, unstructured data needs to be further unified in format to ensure data consistency and usability.

[0093] Solutions for format unification:

[0094] Unifying text encoding: All text data needs to be unified in encoding format (such as UTF-8) to avoid problems such as garbled characters or inability to parse caused by inconsistent encoding.

[0095] Data structuring processing: Convert the cleaned text data into a structured form (such as tables, data frames). Each record includes fields such as device identification, timestamp, fault type, text description, etc., to facilitate subsequent processing and analysis.

[0096] Feature extraction: Extract features related to power grid asset performance evaluation from text data, such as keywords, sentiment, fault categories, etc. These features can be processed using natural language processing techniques (such as TF-IDF, Word2Ve c , BERT, etc.)

[0097] Step 2: Perform text cleaning based on the processed unstructured data, and convert the text data into vector form using a context-aware word embedding model, wherein the context-aware word embedding model dynamically adjusts feature weights based on the equipment operating status and maintenance cycle;

[0098] The step 2 comprises the following steps:

[0099] Input Layer:

[0100] Text input: Input the operation and maintenance text data of power grid equipment, and obtain the basic word embedding of the text through the pre-trained BERT model. Suppose the input text is T = [w1, w2, ..., w n ], each word w is transformed into i Convert to vector e(w i );

[0101] Equipment operation and maintenance status and maintenance cycle information: Assume that the equipment status is S = [s1, s2, ..., s m ], where each s i is a device status parameter, and the maintenance cycle information is M = [m1, m2, ..., m k ], indicating the maintenance cycle data of the equipment;

[0102] Context-aware embedding generation: embedding e(w i ), introduce a weighted mechanism for equipment status and maintenance cycle, and adjust the word embedding according to the equipment status S and maintenance cycle M:

[0103]

[0104] Where: represents the word embedding after weighting by equipment status and maintenance cycle; W s and W m The learned weight representations are used to adjust the impact of equipment status and maintenance cycle on word embedding;

[0105] Transformer encoding layer: The context-aware embedding is input into the Transformer encoding layer of the BERT model, and the contextual relationship of each layer is captured through the self-attention mechanism to generate accurate contextual semantic representation:

[0106]

[0107] Where: H i Represents the final embedding vector of each word, which is output after being processed by the Transformer encoding layer.

[0108] Step 3: Capture the association between text and device status through a multi-level emotion perception model, and build a comprehensive performance evaluation feature library by fusing the output of the multi-level emotion perception model with the processed structured data;

[0109] The step 3 comprises the following steps:

[0110] Sentiment labeling: In each text segment, the sentiment label of the text is extracted through the sentiment analysis model. Let the output of sentiment analysis be E i ;

[0111] Sentiment-enhanced representation: based on the sentiment label E of each text i , for the semantic representation H i For enhancement, an emotional enhancement network is used to adjust the weights of semantic representations according to emotional labels:

[0112] H i，enhanced =H i +γ·f(E i );

[0113] Where: H i is the semantic representation obtained in step 2; γ represents a learnable enhancement coefficient; f(E i ) represents the emotion label E i Calculate the enhancement vector;

[0114] Device state perception layer: associate the device state information with the sentiment information of the text, embed the state information into a high-dimensional space through a multi-layer perceptron, and obtain the device state embedding representation S emb ;

[0115] Emotion and device state association layer: enhance the emotional representation of the text H i，enhanced and the device state embedding representation S emb Fusion is performed to establish the association between text sentiment and device status:

[0116] H final =α·H i，enhanced +β·Semb ;

[0117] Where: α and β are the learned weights; H final represents the final fusion feature;

[0118] Fusion with structured data:

[0119] H fusion =α′·H final +β′·D;

[0120] Where: α′ and β′ are the learned weights; D represents structured data;

[0121] Comprehensive performance evaluation feature library: By integrating the emotion perception model with the device status information, the comprehensive feature vector H is obtained. fusion .

[0122] The calculation steps of the enhancement vector are as follows:

[0123] Emotional label embedding: embed the emotional label E i Convert to embedding vector e i ;

[0124] Constructing a self-attention layer: Fusing the sentiment tag embedding with the semantic representation of the text:

[0125] f(E i )=Self-Attention(H i , e i );

[0126] Where: H i Represents the original representation of the text; e i The embedding vector representing the sentiment tag; the Self-Attention mechanism is used to adjust the semantic representation of the text according to the embedding of the sentiment tag;

[0127] Output emotion enhancement vector: The final output f(E i ) is the weighted fusion vector of sentiment label and text representation.

[0128] Step 4: Use a deep neural network to perform a comprehensive evaluation of asset performance using the comprehensive feature vectors of the comprehensive performance evaluation feature library, thereby achieving asset risk assessment;

[0129] The structure of the deep neural network is as follows:

[0130] Input layer: receives the comprehensive feature vector H fusion , the dimension of the vector is d fusion ;

[0131]

[0132] The first fully connected layer: input feature vector X input , output a vector consisting of h1 neurons; use the ReLU activation function:

[0133] H1=ReLU(W1× input +b1);

[0134] Where: represents the weight matrix; represents the bias vector;

[0135] The second fully connected layer: input feature vector H1, output vector consisting of h2 neurons; use ReLU activation function:

[0136] H2=ReLU(W2H1+b2);

[0137] Where: represents the weight matrix; represents the bias vector;

[0138] The second fully connected layer: input feature vector H2, output vector consisting of h3 neurons; use ReLU activation function:

[0139] H3=ReLU(W3H2+b3);

[0140] Where: represents the weight matrix; represents the bias vector;

[0141] Output layer: Predict the risk score of an asset using a regression task:

[0142] y pred =W out H3+b out ;

[0143] Where: represents the weight matrix, Expresses paranoia.

[0144] Preferably, the loss function of the deep neural network is as follows:

[0145]

[0146] Where: N represents the number of samples; λ represents the L2 regularization coefficient; α i Represents the confidence of the i-th sample; represents the true risk score of the i-th sample; represents the risk score of the i-th sample predicted by the model; W k represents the weight matrix of the kth layer.

[0147] Step 5: The system provides risk assessment reports to operators based on the assessment structure to help decision makers specify subsequent asset management strategies.

[0148] The NLP-based asset performance auxiliary evaluation method provided by the present invention aims to solve the limitations faced by traditional evaluation methods based on structured data, especially the technical bottlenecks in accurately predicting risks and comprehensively evaluating them. Traditional methods mainly rely on structured data of power grid equipment (such as operating status, fault history, etc.), but these data cannot fully capture the potential risks and complex equipment behaviors of the equipment, resulting in the evaluation results often being superficial and difficult to provide accurate risk warnings and performance evaluations. The present invention solves this problem by introducing unstructured data (such as equipment failure logs, operator records, etc.) and using NLP technology for in-depth analysis. First, the structured data is normalized and the unstructured data is cleaned and uniformly formatted to ensure data quality and consistency. Then, a context-aware word embedding model is used to convert unstructured text data into vector form, and the feature weights are dynamically adjusted in combination with the operating status and maintenance cycle of the equipment, so that the changes and risks of the equipment in different states can be reflected in real time. Then, the emotional information in the text is analyzed through a multi-level emotional perception model to capture the deep association between the text and the equipment status, and this information is effectively integrated with the structured data to construct a comprehensive performance evaluation feature library. Finally, based on the feature library, a deep neural network is used to conduct a comprehensive assessment of asset performance, thereby achieving a more accurate risk assessment and providing decision makers with a scientific asset management strategy through a risk assessment report. This method breaks through the limitations of traditional assessment methods through multi-source data fusion and situational awareness, and significantly improves the accuracy and comprehensiveness of power grid asset risk assessment.

Claims

1. The asset performance auxiliary evaluation method based on NLP is characterized by: The following steps are involved: Step 1: Obtain structured data and unstructured data in the power grid field, normalize the structured data, clean the unstructured data, and unify the format after removing irrelevant information; Step 2: Perform text cleaning based on the processed unstructured data, and convert the text data into vector form using a context-aware word embedding model, wherein the context-aware word embedding model dynamically adjusts feature weights based on the equipment operating status and maintenance cycle; Step 3: Capture the association between text and device status through a multi-level emotion perception model, and build a comprehensive performance evaluation feature library by fusing the output of the multi-level emotion perception model with the processed structured data; Step 4: Use a deep neural network to perform a comprehensive evaluation of asset performance using the comprehensive feature vectors of the comprehensive performance evaluation feature library, thereby achieving asset risk assessment; Step 5: The system provides risk assessment reports to operators based on the assessment structure to help decision makers specify subsequent asset management strategies.

2. The NLP-based asset performance auxiliary evaluation method according to claim 1 is characterized in that: The step 2 comprises the following steps: Input Layer: Text input: Input the operation and maintenance text data of power grid equipment, and obtain the basic word embedding of the text through the pre-trained BERT model. Suppose the input text is T = [w1,w2,...,w n ], each word w is transformed into i Convert to vector e(w i ); Equipment operation and maintenance status and maintenance cycle information: Assume that the equipment status is S = [s1, s2, ..., s m ], where each s i is a device status parameter, and the maintenance cycle information is M = [m1, m2, ..., m k ], indicating the maintenance cycle data of the equipment; Context-aware embedding generation: embedding e(w i ), introduce a weighted mechanism for equipment status and maintenance cycle, and adjust the word embedding according to the equipment status S and maintenance cycle M: Where: represents the word embedding after weighting by equipment status and maintenance cycle; W s and W m The learned weight representations are used to adjust the impact of equipment status and maintenance cycle on word embedding; Transformer encoding layer: The context-aware embedding is input into the Transformer encoding layer of the BERT model, and the contextual relationship of each layer is captured through the self-attention mechanism to generate accurate contextual semantic representation: Where: H i Represents the final embedding vector of each word, which is output after being processed by the Transformer encoding layer.

3. The NLP-based asset performance auxiliary evaluation method according to claim 1 is characterized in that: The step 3 comprises the following steps: Sentiment labeling: In each text segment, the sentiment label of the text is extracted through the sentiment analysis model. Let the output of sentiment analysis be E i ; Sentiment-enhanced representation: based on the sentiment label E of each text i , for the semantic representation H i For enhancement, an emotional enhancement network is used to adjust the weights of semantic representations according to emotional labels: H i,enhanced =H i +γ·f(E i ); Where: H i is the semantic representation obtained in step 2; γ represents a learnable enhancement coefficient; f(E i ) represents the emotion label E i Calculate the enhancement vector; Device state perception layer: associate the device state information with the sentiment information of the text, embed the state information into a high-dimensional space through a multi-layer perceptron, and obtain the device state embedding representation S emb ; Emotion and device state association layer: enhance the emotional representation of the text H i,enhanced and the device state embedding representation S emb Fusion is performed to establish the association between text sentiment and device status: H final =α·H i,enhanced +β·S emb ; Where: α and β are the learned weights; H final represents the final fusion feature; Fusion with structured data: H fusion =α′·H final +β′·D; Where: α′ and β′ are the learned weights; D represents structured data; Comprehensive performance evaluation feature library: By integrating the emotion perception model with the device status information, the comprehensive feature vector H is obtained. fusion .

4. The NLP-based asset performance auxiliary evaluation method according to claim 3 is characterized in that: The calculation steps of the enhancement vector are as follows: Emotional label embedding: embed the emotional label E i Convert to embedding vector e i ; Constructing a self-attention layer: Fusing the sentiment tag embedding with the semantic representation of the text: f(E i )=Self-Attention(H i ,e i ); Where: H i Represents the original representation of the text; e i The embedding vector representing the sentiment tag; the Self-Attention mechanism is used to adjust the semantic representation of the text according to the embedding of the sentiment tag; Output emotion enhancement vector: The final output f(E i ) is the weighted fusion vector of sentiment label and text representation.

5. The NLP-based asset performance auxiliary evaluation method according to claim 1 is characterized in that: The structure of the deep neural network is as follows: Input layer: receives the comprehensive feature vector H fusion , the dimension of the vector is d fusion ; The first fully connected layer: input feature vector X input , output a vector consisting of h1 neurons; use the ReLU activation function: H1=ReLU(W1X input +b1); Where: represents the weight matrix; represents the bias vector; The second fully connected layer: input feature vector H1, output vector consisting of h2 neurons; use ReLU activation function: H2=ReLU(W2H1+b2); Where: represents the weight matrix; represents the bias vector; The second fully connected layer: input feature vector H2, output vector consisting of h3 neurons; use ReLU activation function: H3=ReLU(W3H2+b3); Where: represents the weight matrix; represents the bias vector; Output layer: Predict the risk score of an asset using a regression task: and pred =In out H3+b out ; Where: represents the weight matrix, Expresses paranoia.

6. The NLP-based asset performance auxiliary evaluation method according to claim 5 is characterized in that: The loss function of the deep neural network is as follows: Where: N represents the number of samples; λ represents the L2 regularization coefficient; α i Represents the confidence of the i-th sample; represents the true risk score of the i-th sample; Represents the risk score of the i-th sample predicted by the model; W k represents the weight matrix of the kth layer.

Images