A topic labeling method and system based on dynamic LDA

Abstract

A kind of theme labeling method and system based on dynamic LDA, comprising: collecting historical power operation and maintenance work order data from multiple heterogeneous data sources and carrying out pretreatment operation, constructs power field professional dictionary;According to time stamp information data division is time slice, the power operation and maintenance work order data of each time slice is input as prior information into dynamic LDA model based on particle filtering, to train the model, optimize model parameters, the trained dynamic LDA model based on particle filtering corresponding to current time slice is stored, for the training of next time slice, and the theme of new power operation and maintenance work order data is labeled using the trained dynamic LDA model based on particle filtering.The present application realizes the accurate theme labeling of power field text.

Description

Technical Field

[0001] This invention belongs to the field of artificial intelligence, and specifically relates to a topic labeling method and system based on dynamic LDA. Background Technology

[0002] The safe and stable supply of power is of paramount importance. The power system is a vast and complex system, and its safe and stable operation depends on efficient operation and maintenance management. Power operation and maintenance work orders, as important carriers of records of power equipment operating status, fault information, and maintenance records, contain a wealth of knowledge about power system operation and maintenance. How to extract effective information from the massive amount of power operation and maintenance work order text data to provide decision support for the safe and stable operation of the power system is a key problem that urgently needs to be solved in the field of power system operation and maintenance.

[0003] Traditional power grid operation and maintenance work order analysis relies heavily on human experience, resulting in low efficiency, large errors, and difficulty in handling massive amounts of data, making it unsuitable for the higher demands placed on power grid operation and maintenance in the era of smart grids. In recent years, with the rapid development of artificial intelligence technology, there is a need to apply natural language processing (NLP) technology to power grid operation and maintenance work order analysis to achieve intelligent applications such as fault diagnosis, trend prediction, and knowledge extraction.

[0004] In the field of text mining, a large amount of data is unstructured, making it difficult to directly extract relevant and desired information. Topic models are important tools in text mining, capable of identifying topics in documents and uncovering hidden information within corpora. They have wide applications in topics such as topic aggregation, information extraction from unstructured text, and feature selection, and have gained significant attention in both industry and academia in recent years. Latent Dirichlet Allocation (LDA) is a representative model. Proposed by Blei, David M., Ng, Andrew Y., and Jordan in 2003, LDA is used to infer the topic distribution of documents. It can discover hidden topic patterns in corpora; annotate documents based on topics; and organize, sort, summarize, and retrieve documents using these annotations. However, traditional LDA models struggle to handle the dynamic and specialized characteristics of power operation and maintenance work order data, limiting their application in the power operation and maintenance field. Summary of the Invention

[0005] To address the shortcomings of existing technologies, this invention provides a topic labeling method and system based on dynamic LDA to solve the technical problem of accurate topic labeling of text data in the power industry.

[0006] To solve the above-mentioned technical problems, the present invention adopts the following technical solution.

[0007] This invention first discloses a topic annotation method based on dynamic LDA, which includes the following steps:

[0008] Step 1: Collect historical power operation and maintenance work order data from multiple heterogeneous data sources, convert the text data in the collected historical power operation and maintenance work order data into a unified format, perform preprocessing operations, and build a professional dictionary in the power field.

[0009] Step 2: Based on the timestamp information of the power operation and maintenance work order data, divide the power operation and maintenance work order data into different time slices, and use the power operation and maintenance work order data of each time slice as prior information to input the dynamic LDA model based on particle filtering to train the model. The model training uses the incremental Gibbs sampling algorithm to sample new data, update particle weights and model parameters, and realize the incremental update of the model.

[0010] Step 3: Optimize the model parameters, store the trained particle filter-based dynamic LDA model corresponding to the current time slice for training in the next time slice, and use the trained particle filter-based dynamic LDA model to perform topic labeling on the new power operation and maintenance work order data.

[0011] The present invention further includes the following preferred embodiments:

[0012] The data sources include internal data from the State Grid Corporation of China and publicly available external data. The internal data from the State Grid Corporation of China includes:

[0013] Technical standards, specifications, and procedures are textual data used to define the operation and maintenance specifications for power equipment;

[0014] The power equipment operation and maintenance manual includes the equipment's structure, operating principles, common faults, and troubleshooting methods.

[0015] The fault diagnosis case library contains detailed records of historical fault cases, including fault symptoms, cause analysis, and handling measures.

[0016] The externally disclosed data includes:

[0017] Electricity industry news, including power system operation status, application of new technologies, and failure events;

[0018] Technical forum discussion threads contain analyses and discussions of fault cases by power industry experts and engineers.

[0019] The preprocessing operation further includes:

[0020] First, data cleaning is performed to remove irrelevant information, including HTML tags, special characters, and punctuation marks, and the data format is standardized. Then, data deduplication is performed, and missing fields are filled in or deleted.

[0021] Then, a power industry-specific word segmentation tool was used for word segmentation. A new word discovery algorithm was used to identify new words, professional terms and equipment models in the power field and add them to the word segmentation dictionary.

[0022] Construct a stop dictionary, remove words whose contribution to the semantics of the text is below a threshold, and add stop words specific to the power operation and maintenance work order domain;

[0023] A stemming algorithm is used to convert words with different word forms into the same stemming form;

[0024] The word segmentation results are tagged with part-of-speech tags to identify the grammatical attributes of words; named entity recognition tools are used to identify named entities in the text and professional entities in the power field.

[0025] The step of using power maintenance work order data from each time slice as prior information input into the particle filter-based dynamic LDA model further includes:

[0026] Using a professional dictionary in the power industry as prior knowledge, the dictionary information is integrated into the model. When calculating word frequency, professional words in the dictionary are given higher weights; or, in the Dirichlet prior distribution of topic-word distribution, professional words in the dictionary are given higher prior probabilities; or, the dictionary information is used to construct a topic-word association matrix as prior information for topic-word distribution, guiding the model to learn the topic-word distribution of knowledge in the power industry.

[0027] Collect relevant information about power equipment, and use knowledge representation and reasoning techniques to construct a knowledge graph of power equipment. Based on the knowledge graph of power equipment, extract the semantic features of entities in power operation and maintenance work orders, and use these semantic features as input features of the PF-DTM model, integrating them into the bag-of-words representation of the document.

[0028] Utilize the relational information in the knowledge graph to establish association constraints between topics;

[0029] We collect and organize the experience and rules of power operation and maintenance experts in fault diagnosis and fault handling, and convert the expert experience into a rule form that can be understood and executed by computers. We then integrate the expert rules as constraints into the training process of the PF-DTM model to guide the model to learn the distribution and evolution of topics.

[0030] The incremental Gibbs sampling algorithm samples new data, updates particle weights and model parameters, and achieves incremental updates of the model, further including:

[0031] Step 2.1: For the data in the first time slice, train the standard PF-DTM model to obtain the initial topic-word distribution. Document-topic distribution θ, and particle set;

[0032] Step 2.2: For a new time slice t, the particle set and model parameters of the previous time slice t-1 are used as prior information. For each particle, the other particles and model parameters are kept unchanged. Based on the data of the current time slice and the prior information, the topic assignment of the particle is sampled and updated using the Gibbs sampling algorithm.

[0033] Step 2.3: Re-estimate the topic-word distribution based on the updated particle set. The parameters of the state transition function g are used to recalculate the weight of each particle based on the updated model parameters; the particles are then resampled based on their weights to generate a new set of particles for incremental updates in the next time slice.

[0034] Step 2.4: Repeat steps 2.2 and 2.3 until all time slices of data have been processed.

[0035] Step 2.3 further includes:

[0036] For each time slice, N particles are randomly generated, each particle representing a possible topic distribution θt; the weight of all particles is initialized to 1 / N, indicating that each particle has the same probability in the initial state;

[0037] For each time slice t, the probability of generating observation data for the topic distribution represented by each particle is calculated according to the observation equation, and used as the weight of each particle; resampling is performed according to the weight of the particles to generate new N particles;

[0038] Estimate the topic distribution θt for time slice t based on the resampled particles;

[0039] Based on the estimated topic distribution θt and the observed data, update the model parameters, including the topic-word distribution. And the parameters of the state transition function g;

[0040] Repeat the above steps until the model converges.

[0041] The optimization of model parameters further includes:

[0042] Select the time slice size Δt based on the data characteristics and analysis objectives;

[0043] The effects of different numbers of topics K were evaluated using perplexity and topic consistency metrics, and the optimal value for the number of topics was selected.

[0044] The number of particles N is selected based on model accuracy and computational efficiency;

[0045] Choose the state transition function g based on prior knowledge of the data.

[0046] This invention also discloses a topic annotation system based on dynamic LDA that utilizes the aforementioned topic annotation method based on dynamic LDA, comprising:

[0047] The data collection module is used to collect historical power operation and maintenance work order data from various heterogeneous data sources, convert the text data in the collected historical power operation and maintenance work order data into a unified format, perform preprocessing operations, and build a professional dictionary in the power field.

[0048] The dynamic LDA model training module is used to divide the power operation and maintenance work order data into different time slices according to the timestamp information of the power operation and maintenance work order data, and input the power operation and maintenance work order data of each time slice as prior information into the dynamic LDA model based on particle filtering to train the model. The model training uses the incremental Gibbs sampling algorithm to sample new data, update particle weights and model parameters, and realize the incremental update of the model.

[0049] The topic labeling module is used to optimize model parameters, store the trained particle filter-based dynamic LDA model corresponding to the current time slice for training in the next time slice, and use the trained particle filter-based dynamic LDA model to label new power operation and maintenance work order data.

[0050] Accordingly, this application also discloses a terminal, including a processor and a storage medium;

[0051] The storage medium is used to store instructions;

[0052] The processor is configured to operate according to the instructions to perform the steps of the aforementioned topic labeling method based on dynamic LDA.

[0053] Accordingly, this application also discloses a computer-readable storage medium storing a computer program that, when executed by a processor, implements the steps of the aforementioned topic labeling method based on dynamic LDA.

[0054] The beneficial effects of this invention are that, compared with the prior art, this invention provides a topic labeling method and system based on dynamic LDA, which combines the LDA model with the particle filtering algorithm and introduces an incremental learning mechanism, enabling it to effectively capture the evolution trend of power operation and maintenance work order topics over time. At the same time, it combines knowledge from fields such as power industry professional dictionaries, power equipment knowledge graphs, and power expert experience, further improving the accuracy and practicality of the model. Attached Figure Description

[0055] Figure 1This is a flowchart of the dynamic LDA model construction process in this invention. Detailed Implementation

[0056] To make the objectives, technical solutions, and advantages of the present invention clearer, the technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present invention.

[0057] The embodiments described in this application are merely some, not all, embodiments of the present invention. Based on the spirit of the present invention, other embodiments obtained by those skilled in the art without inventive effort are all within the protection scope of the present invention.

[0058] Power maintenance work order text data differs from general text data in that it has the following significant characteristics:

[0059] (1) The power operation and maintenance work order text is filled with a large number of power professional terms, equipment models, operating parameters, etc., such as "transformer", "circuit breaker", "overload", "grounding fault", etc. It is necessary to combine power system professional knowledge to accurately understand their meaning.

[0060] (2) The operating status of the power system changes constantly over time, and the thematic information reflected in the power operation and maintenance work orders also shows a dynamic evolution trend. For example, high temperatures in summer may lead to frequent overload faults in power lines, while low-temperature operation of equipment is more likely to occur in winter.

[0061] (3) Power operation and maintenance work orders have a wide range of data sources, including user repair work orders, system alarm work orders, inspection record work orders, etc. Work orders from different sources have significant differences in text format, content structure, professional terminology, etc.

[0062] (4) The power operation and maintenance work order data is huge, but the truly valuable information is often hidden in the massive text descriptions, which can only be extracted through effective text mining technology.

[0063] To address the shortcomings of existing technologies, this invention proposes a topic annotation method and system based on dynamic LDA. Building upon existing LDA topic mining techniques and incorporating the characteristics of the power industry, a dynamic LDA topic model is constructed. This model enables accurate topic annotation of text data in the power sector. Targeted analysis is conducted on the characteristics of power grid user maintenance work orders and system maintenance work orders. Multi-source heterogeneous data, including State Grid standards and specifications, expert systems, knowledge base systems, and social networks, are integrated to construct a power sector text corpus. Based on the LDA model and combined with time series analysis methods, a dynamic LDA topic model is constructed. Through incremental Gibbs sampling, the topic distribution and vocabulary distribution of time-slice text sets are dynamically estimated. By using the dynamic LDA model to annotate power sector text data and uncover potential topic evolution trends, the aim is to build an efficient and accurate power maintenance work order topic annotation system, providing technical support for power system fault diagnosis, trend prediction, and knowledge discovery.

[0064] Latent Dirichlet Allocation (LDA) is an unsupervised probabilistic topic model used to discover latent topic information from a collection of documents. Its basic idea is that each document is a mixture of multiple topics, and each topic is composed of multiple words with specific probabilities. The goal of an LDA model is to infer the latent topic structure by analyzing the co-occurrence relationships of words in a document and to provide the probability distribution of the document across different topics.

[0065] The mathematical representation of the LDA model is as follows:

[0066] Document set: D = {d1, d2, ..., d...} M}, where M represents the total number of documents, d i This represents the i-th document.

[0067] Set of topics: T = {t1, t2, ..., t} K}, where K represents the total number of topics, t k This represents the k-th topic.

[0068] Word set: W = {w1, w2, ..., w} V}, where V represents the total number of words, w v This indicates the vth word.

[0069] LDA model assumptions:

[0070] Each document d i It can be represented as a mixed distribution θ of K topics. i ={θ i1 θ i2 , ..., θ iK}, where θ ikDocument d represents document d i Belongs to topic t k The probability of.

[0071] Each topic t k This can be represented as a probability distribution of V words. in The word w v Belongs to topic t k The probability of.

[0072] The generation process of LDA models can be described as follows:

[0073] For each document d i :

[0074] From the Dirichlet distribution D ir (α) samples a topic distribution θ i .

[0075] For document d i Each word w in j :

[0076] From topic distribution θ i Sampling a topic z j .

[0077] From topic z j Corresponding word distribution Sample a word w j .

[0078] Repeat the above process until all documents are generated.

[0079] In the LDA model, the parameters α and β control the sparsity of the document-topic distribution and the topic-word distribution, respectively. The larger α is, the more uniform the topic distribution of the documents; the larger β is, the more uniform the word distribution of the topics.

[0080] The training of LDA models involves iteratively updating model parameters, enabling the model to better interpret the observed document data.

[0081] LDA models can be used for topic analysis of power operation and maintenance work orders, for example:

[0082] Fault type identification: Power maintenance work orders are classified by topic, such as "transformer fault", "line fault", "circuit breaker fault", etc., to help maintenance personnel quickly identify fault types and improve fault handling efficiency.

[0083] Example 1:

[0084] Work order 1: "10kV #3 line tripped, protection action information is instantaneous overcurrent, preliminary judgment is line short circuit fault."

[0085] Work order 2: "The user reported that the voltage in their home was too low. After investigation, it was found that the problem was caused by a fault in the transformer substation."

[0086] The LDA model can categorize these two work orders under the themes of "line fault" and "transformer fault" respectively, helping maintenance personnel to quickly identify the fault type.

[0087] Fault Cause Analysis: By analyzing the most frequent terms under each topic, common causes of different types of faults can be identified. For example, if the most frequent terms under the topic "line fault" include "tree obstruction," "foreign object," and "lightning strike," it indicates that the main cause of the line fault may be external environmental factors.

[0088] Example 2:

[0089] Work Order 3: "Heavy rain caused a tree-related trip on the 10kV #5 line. The line load has been switched to other power sources, and repairs are currently underway."

[0090] Work order 4: "Thunderstorms caused a lightning strike to trip the 10kV #8 line. Personnel have been dispatched to patrol the line and investigate the fault."

[0091] The LDA model can identify keywords related to line faults, such as "tree obstruction" and "lightning strike," to help maintenance personnel analyze the causes of faults.

[0092] Fault trend prediction: By analyzing the trend of the distribution of themes in different time periods, future fault development trends can be predicted, such as which types of faults may increase and which areas have a higher risk of faults, thereby providing support for power system operation and maintenance decisions.

[0093] Example 3: If the LDA model finds a significant increase in the probability of topics related to "transformer overload" recently, it indicates that the risk of transformer overload operation is increasing, and maintenance personnel need to take timely measures to avoid large-scale power outages.

[0094] Traditional LDA models assume that the document set is static, meaning all documents come from the same time period. However, power maintenance work order data is dynamic, and traditional LDA models struggle to capture the evolution of topics over time. Traditional LDA models treat words merely as independent symbols, ignoring semantic relationships between words and power industry expertise, which limits the model's ability to understand power maintenance work order text.

[0095] To address the limitations of traditional LDA models in handling dynamic text data, this invention employs a Dynamic Topic Model (DTM) to incorporate temporal information. This model assumes that documents from different time periods have different topic distributions, and that these topic distributions exhibit an evolutionary relationship. By incorporating temporal information, the DTM model effectively captures the evolutionary patterns of topics over time, such as the emergence of new topics, the disappearance of old topics, and the merging and splitting of topics. By considering these evolutionary relationships, the DTM model can better distinguish topics from different time periods, thereby improving the accuracy of topic identification. Analyzing the evolutionary trajectory of topics can uncover potential evolutionary patterns, such as periodic changes and trend shifts, providing support for decision-making.

[0096] This invention employs a dynamic LDA model for topic labeling of power maintenance work orders. First, data collection and preprocessing are performed, including collecting power maintenance work order data and conducting data cleaning, word segmentation, and stop word removal to prepare for subsequent topic modeling. Then, based on the timestamp information of the power maintenance work orders, the data is divided into different time slices, such as by day, week, or month, to facilitate the dynamic LDA model in capturing topic evolution trends. A suitable dynamic LDA model is selected and trained using the preprocessed power maintenance work order data. The evolution trends of topics in different time slices are analyzed, including the emergence of new topics, the disappearance of old topics, and the merging and splitting of topics, and the evolution patterns are explained using knowledge of power system operation and maintenance. Finally, the trained dynamic LDA model is used to label new power maintenance work orders, and the labeling results are applied to fault diagnosis, trend prediction, and knowledge discovery.

[0097] This invention employs a particle filtering-based dynamic LDA model, PF-DTM. Particle filtering is a non-parametric Bayesian inference method that can effectively handle nonlinear and non-Gaussian dynamic systems. By representing the topic distribution of each time slice as a set of particles and using the particle filtering algorithm to iteratively update the particles, the evolutionary patterns of topic appearance, disappearance, merging, and splitting can be accurately captured.

[0098] This invention introduces an incremental learning mechanism into the PF-DTM model, enabling the model to continuously update the topic model based on new data while retaining existing knowledge, without retraining the entire model. This invention employs an incremental Gibbs sampling algorithm for model training. In each new time slice, the particle set and model parameters from the previous time slice are used as prior information. The incremental Gibbs sampling algorithm is then used to sample new data, updating particle weights and model parameters, thus achieving incremental model updates. This incremental learning mechanism significantly improves the model's training efficiency and allows the model to track the evolution trend of the topic in real time.

[0099] See Figure 1 As shown, the topic annotation method based on dynamic LDA disclosed in this invention includes the following steps:

[0100] Step 1: Collect historical power operation and maintenance work order data from multiple heterogeneous data sources, convert the text data in the collected historical power operation and maintenance work order data into a unified format, perform preprocessing operations, and at the same time build a professional dictionary in the power field.

[0101] During the data preparation phase, data from heterogeneous data sources are collected and preprocessed to prepare for subsequent PF-DTM model construction and experimental analysis. Data sources include internal State Grid data and publicly available external data.

[0102] The State Grid's internal data includes textual data such as technical standards, specifications, and procedures, which define the operation and maintenance specifications of power equipment and contain rich domain knowledge, helping the model understand the professional terminology and fault phenomenon descriptions in work orders; power equipment operation and maintenance manuals, which contain information such as the structure, principles, common faults and handling methods of the equipment, can be used to build a power equipment knowledge graph and enhance the model's understanding of fault information; and a fault diagnosis case library, which contains detailed records of historical fault cases, including fault phenomena, cause analysis, and handling measures, can be used for model training and evaluation.

[0103] Publicly available external data includes power industry news, which contains information on power system operation, new technology applications, and failure events, helping models capture the evolution trends of themes; and technical forum discussion threads, which contain analysis and discussion of failure cases by power industry experts and engineers, which can be used to supplement domain knowledge and expert experience.

[0104] In the preprocessing stage, the collected text data is converted into a uniform format, such as txt or XML, to facilitate subsequent processing; for unstructured data containing information such as tables and images, it is extracted and converted into text form or structured data for model processing.

[0105] First, data cleaning is performed to remove irrelevant information, including HTML tags, special characters, punctuation marks, etc., and the data format is standardized, such as date and time formats, and units. Duplicate or highly similar work order records are removed through data deduplication. Finally, missing fields are filled in or deleted.

[0106] Then, power industry-specific word segmentation tools such as Jieba are used for word segmentation, combined with a professional dictionary in the power field to improve the accuracy of word segmentation. New word discovery algorithms are used to identify new words, professional terms, and equipment models in the power field, and these are added to the word segmentation dictionary to improve the model's sensitivity to new information.

[0107] Construct a stop word dictionary to remove words with a semantic contribution degree lower than the threshold in the text, such as "的", "是", "了", etc. Then, according to the characteristics of power operation and maintenance work orders, add domain-specific stop words, such as some common device operation status description words, etc.;

[0108] Adopt a stemming algorithm to convert words with different word forms into the same stem form. For example, convert "跳闸", "跳开", "跳合" into "跳" to reduce the vocabulary size and improve the model efficiency;

[0109] Use a词性标注工具 (lexical category tagging tool) to perform lexical category tagging on the word segmentation results to identify the grammatical attributes of words, such as nouns, verbs, adjectives, etc., providing a basis for subsequent feature selection; Use a named entity recognition tool to identify named entities in the text, such as person names, place names, organization names, and professional entities in the power field, such as device names, fault types, operation parameters, etc.

[0110] Step 2: According to the timestamp information of the power operation and maintenance work order data, divide the power operation and maintenance work order data into different time slices, and use the power operation and maintenance work order data of each time slice as prior information to input into the dynamic LDA model based on particle filter to train the model. In model training, an incremental Gibbs sampling algorithm is used to sample new data, update particle weights and model parameters, and achieve incremental update of the model.

[0111] To further improve the accuracy and interpretability of the model, the present invention integrates power domain knowledge into all aspects of the PF-DTM model, achieving multi-level integration of power domain knowledge.

[0112] First, collect and organize materials such as power industry standards, specifications, and device manuals, extract vocabulary such as professional terms, device models, and fault types, and construct a power domain professional dictionary. This dictionary can be used in preprocessing links such as word segmentation, lexical category tagging, and named entity recognition, and as prior knowledge to be integrated into the learning process of the topic-word distribution.

[0113] Use the power domain professional dictionary as prior knowledge to guide the model in learning the topic-word distribution When doing so, increase the attention to professional vocabulary. In specific embodiments, the dictionary information can be integrated into the model in various ways. For example, when calculating the word frequency of words, assign higher weights to professional vocabulary in the dictionary; or in the Dirichlet prior distribution of the topic-word distribution, set higher prior probabilities for professional vocabulary in the dictionary; or use the dictionary information to construct a topic-word association matrix as the prior information of the topic-word distribution to guide the model to learn a topic-word distribution more in line with power domain knowledge.

[0114] It should be noted that "词性标注工具" is a literal translation, and in the context of natural language processing, there are more specific and commonly used terms like "part-of-speech tagging tool" which might be more appropriate depending on the actual situation. Here, the literal translation is used for the purpose of directly translating the given text.Collect relevant information about power equipment, such as equipment type, function, parameters, connection relationships, and common faults. Utilize knowledge representation and reasoning techniques to construct a power equipment knowledge graph. Based on this knowledge graph, extract semantic features of entities in power operation and maintenance work orders, such as equipment type, fault type, and geographical location. These semantic features can serve as input features for the PF-DTM model, for example, by incorporating them into the bag-of-words representation of the document, enhancing the model's understanding of the semantic information in the work orders.

[0115] By leveraging relational information within knowledge graphs, such as connections between devices and fault propagation relationships, we can establish association constraints between topics. For instance, if two devices are connected in the knowledge graph, we can constrain the PF-DTM model to bring related topics closer together during evolution, or increase the probability of these two topics appearing simultaneously during particle sampling.

[0116] We collect and organize the experience and rules of power operation and maintenance experts in fault diagnosis and handling, and transform this expert experience into rule forms that can be understood and executed by computers, such as IF-THEN rules and decision trees. These expert rules are then incorporated as constraints into the training process of the PF-DTM model, guiding the model to learn topic distributions and evolution patterns that better reflect real-world conditions. For example:

[0117] According to expert rules, the probability of specific words under certain topics is adjusted, such as increasing the probability of words related to the cause of failure under that topic; according to expert rules, the topic evolution path is constrained, such as limiting the possibility of merging or splitting between certain topics; or according to expert rules, the topic allocation of particles is screened during particle sampling, such as removing particles that do not conform to expert rules.

[0118] During the model training phase, a professional dictionary of the power industry is integrated as prior knowledge into the topic-word distribution, guiding the model to learn a topic-word distribution that better aligns with power industry knowledge. Entity and relationship information from the power equipment knowledge graph is utilized to enhance the model's semantic understanding capabilities, for example, through semantic feature extraction and topic association constraints. Power expert experience is converted into rules or constraints and integrated into the model training process, guiding the model to learn topic distributions and evolutionary patterns that better reflect real-world situations. By integrating power industry knowledge into different stages of the model, deep knowledge fusion is achieved, significantly improving the model's domain adaptability and the accuracy of topic labeling.

[0119] The PF-DTM model incorporates a particle filter algorithm to handle the dynamic changes in topic distribution over time. Particle filtering uses a set of "particles" to represent the possible states of the topic distribution θt at various time slices. Each particle represents a possible topic distribution and has a corresponding weight, indicating how well that distribution fits the actual data. As new data arrives, the particles are updated based on the observed data; particles with higher weights represent a more reasonable topic distribution, while particles with lower weights are eliminated, thus tracking the evolution of the topic.

[0120] Based on the PF-DTM model, a topic feature extraction method for operation and maintenance work order processing is designed. The document-topic distribution θ learned by the PF-DTM model is used to represent the topic features of the work order, i.e., a K-dimensional vector, where each element represents the probability that the work order belongs to the corresponding topic. This topic feature representation method can effectively capture the topic information of the work order and be used for subsequent similarity calculation and retrieval.

[0121] The Particle Filter-based Dynamic LDA (PF-DTM) model integrates knowledge from the power industry and introduces an incremental learning mechanism to improve the model's accuracy, efficiency, and interpretability, enabling it to be effectively applied to the intelligent processing of power operation and maintenance work orders.

[0122] This invention introduces an incremental learning mechanism into the PF-DTM model, enabling the model to continuously update the topic model based on new data while retaining existing knowledge, without having to retrain the entire model.

[0123] Specifically, the particle set and model parameters of the previous time slice are used as prior information for the current time slice. The incremental Gibbs sampling algorithm is used to sample new data, update particle weights and model parameters, and realize incremental update of the model.

[0124] This invention employs the incremental Gibbs sampling algorithm to achieve incremental learning of the PF-DTM model. Its main steps are as follows:

[0125] Step 2.1: For the data in the first time slice, train the standard PF-DTM model to obtain the initial topic-word distribution. Document-topic distribution θ, and particle set.

[0126] Step 2.2: For a new time slice t, the particle set and model parameters of the previous time slice t-1 are used as prior information. For each particle, the other particles and model parameters are kept unchanged. Based on the data of the current time slice and the prior information, the topic assignment of the particle is sampled and updated using the Gibbs sampling algorithm.

[0127] The topic distribution θ of the current time slice t t The topic z was obtained by sampling from the middle.tj For each topic k, the probability of being selected in time slice (t) is θ. t,k ;

[0128] P(z tj =k)=θ t,k k = 1, 2, ..., K

[0129] Step 2.3: Re-estimate the topic-word distribution based on the updated particle set. and the parameters of the state transition function g; that is, through the topic z j Word distribution φ zj Generate word w tj Re-estimate the word w tj The generation probability, where φ zj,v This indicates that the word 'v' in the vocabulary appears in topic 'z'. j The probability of the following:

[0130] P(w tj =v|z tj =z j )=φ zj,v v = 1, 2, ..., V

[0131] Based on the updated model parameters, recalculate the weight of each particle; update the word w. tj With topic z j The joint probability is obtained by combining the likelihood of words with the prior knowledge of topics using Bayes' theorem:

[0132]

[0133] Where: P(w) tj |φ zj ) is the word w tj In topic z j The likelihood probability of z; P(z) tj |θ t ) is the topic z j Prior probability at time slice t;

[0134] Resampling is performed based on the topic distribution and particle weights to generate a new particle set for incremental updates in the next time slice:

[0135] Where w tj,k It is the weight of particle k, based on the observed words and the previous state; These are new particles obtained after resampling.

[0136] Step 2.4: Repeat steps 2.2 and 2.3 until all time slices of data have been processed.

[0137] The PF-DTM model uses a state-space model to describe the dynamic evolution of topic distribution. It uses a topic distribution θ to describe time slice t. t From the previous time slice θ t-1 The state equation for the evolution process is:

[0138] θ t =g(θ) t-1 ,ε t )

[0139] Where g(·) represents the state transition function, which can be a linear or nonlinear function, such as a linear Gaussian model or a nonlinear dynamic system model, used to capture the trend of topic evolution. ε t This represents state noise, which is usually assumed to follow a Gaussian distribution and is used to simulate random fluctuations in the evolution of a subject.

[0140] Using a description of a given topic distribution θ t and word distribution φ zj In the case of generating words w in time slice t tj The observation equation:

[0141] w tj ~Mult(θ) t ,φ zj )

[0142] Where w tj φ represents the j-th word in time slice t; zj Representing topic z j The corresponding word distribution, that is, the probability of different words appearing under this topic, is usually a vector of length V, where V is the size of the vocabulary.

[0143]

[0144] Where w tj,v This represents the frequency (0 or 1) of the j-th word (v) in time slice t, and represents the frequency (0 or 1) of the word (w). tj Is it equal to the v-th word in the vocabulary? φ zj,v This represents the probability of word v appearing in relation to topic zj.

[0145] This represents the product of all items at index v from 1 to V.

[0146] Step 3: Optimize the model parameters, store the trained particle filter-based dynamic LDA model corresponding to the current time slice for training in the next time slice, and use the trained particle filter-based dynamic LDA model to perform topic labeling on the new power operation and maintenance work order data.

[0147] The performance of the PF-DTM model is affected by multiple parameters, which need to be set and optimized appropriately according to the specific dataset and application scenario.

[0148] The time slice size (Δt) determines the temporal granularity at which the model analyzes the evolution of the subject. Choosing a smaller time slice allows for more precise capture of short-term changes in the subject, but may increase the model's computational complexity and the risk of overfitting. Choosing a larger time slice can reduce the model's computational complexity, but may result in the loss of some important short-term information. An appropriate time slice size, such as in days, weeks, or months, needs to be selected based on the characteristics of the data and the analysis objectives.

[0149] The number of topics (K) determines the number of potential topics the model can identify. Too few topics will prevent the model from capturing all the topic information in the data, while too many topics will make the model overly complex and prone to overfitting. Metrics such as perplexity and topic consistency can be used to evaluate the effectiveness of different numbers of topics and select the optimal value.

[0150] The number of particles (N) represents the number of samples used to approximate the posterior probability of the topic distribution. A higher number of particles results in a more accurate estimate of the posterior probability, but also increases computational complexity. A trade-off between model accuracy and computational efficiency must be struck based on the specific circumstances to select an appropriate number of particles.

[0151] The state transition function (g) describes the evolution of the topic distribution over time. An appropriate function form can be chosen based on prior knowledge of the data. For example: a linear Gaussian model is suitable for scenarios where topic evolution is relatively smooth; a nonlinear dynamic system model is suitable for scenarios where topic evolution is more complex and involves nonlinear relationships; and a Gaussian process model is suitable for scenarios where the evolutionary patterns of the topic are difficult to pre-assume.

[0152] The hyperparameters of the Dirichlet distribution include α and β. α controls the sparsity of the document-topic distribution; a larger α value indicates a more uniform distribution of topics, meaning the document tends to contain more topics. β controls the sparsity of the topic-word distribution; a larger β value indicates a more uniform distribution of words within topics, meaning the topic tends to contain more words. The α and β parameters can be optimized using methods such as cross-validation.

[0153] The training process of the PF-DTM model uses the particle filtering algorithm to iteratively update the particle weights based on the observed data, and finally obtains model parameters that accurately describe the evolution of the topic.

[0154] 1. For each time slice, randomly generate N particles, each representing a possible topic distribution θt. Initialize the weight of all particles to 1 / N, indicating that each particle has the same probability in the initial state.

[0155] 2. For each time slice t, calculate the weight of each particle based on the observed data (i.e., the power maintenance work order text data within that time slice). The weight is calculated by determining the probability of generating observed data for the topic distribution represented by each particle, according to the observation equation. Resampling is then performed based on the particle weights to generate N new particles. During resampling, particles with higher weights are more likely to be replicated multiple times, while particles with lower weights may be discarded.

[0156] 3. Estimate the topic distribution θt for time slice t based on the resampled particles. For example, the weighted average of all particles can be used as an estimate of the topic distribution.

[0157] 4. Update the model parameters, including topic-word distribution, based on the estimated topic distribution θt and the observed data. And the parameters of the state transition function g. Parameter updates can be performed using methods such as maximum likelihood estimation and expectation-maximization (EM) algorithms.

[0158] 5. Repeat steps 2-4 until the model converges.

[0159] Power operation and maintenance work orders can be divided into two main categories: power grid user operation and maintenance work orders and system operation and maintenance work orders. The two types of work orders differ in text features and topic distribution, and therefore require different topic annotation strategies.

[0160] For power grid user maintenance work orders, due to the colloquial language used, a dictionary oriented towards the user's language needs to be built for tasks such as word segmentation, part-of-speech tagging, and named entity recognition. Based on the user's geographical location information, user maintenance work orders are clustered to analyze the characteristics of power problems in different regions. For example, user maintenance work orders in urban areas can be clustered together to analyze the fault characteristics of urban distribution networks; user maintenance work orders in rural areas can be clustered together to analyze the fault characteristics of rural power grids. Text classification technology is used to identify the user's electricity consumption type, and different topic models are built for different types of users. For example, one topic model can be built for residential users, and another topic model can be built for commercial users to more accurately identify the electricity needs and fault types of different user types.

[0161] For system maintenance work orders, a specialized dictionary in the power industry needs to be built for tasks such as word segmentation, part-of-speech tagging, and named entity recognition, to improve the topic model's ability to recognize power-related technical terms. External knowledge bases, such as power equipment knowledge graphs, can be utilized to enhance the semantic understanding capabilities of the topic model. For example, entity and relationship information from the equipment knowledge graph can be integrated into the dynamic LDA model to improve the model's understanding of power maintenance work order text and the accuracy of topic recognition. Time-series information can be incorporated into the dynamic LDA model to analyze the evolution patterns of power system faults. For example, information on the changes in power equipment operating parameters over time can be used as input features of the dynamic LDA model to study the impact of different operating parameters on fault occurrence.

[0162] During the experimental verification phase, real power operation and maintenance work order data were used to conduct experimental verification of the constructed PF-DTM model and to compare and analyze it with other baseline models to evaluate the effectiveness and advancement of the model.

[0163] First, the preprocessed text data in the power sector is divided into training, validation, and test sets. Suitable baseline models are then selected for comparative experiments, such as the traditional LDA model, the Kalman filter-based dynamic LDA model (KF-DTM), and the variational inference-based dynamic LDA model (VI-DTM).

[0164] Choose the following evaluation metrics to assess the model's performance:

[0165] Perplexity: Measures a model’s ability to predict unseen data. The lower the perplexity, the better the model’s predictive performance.

[0166] Topic Coherence: This assesses the semantic consistency of topics, ensuring that words within each topic have good semantic relevance. Commonly used metrics include UMass, UCI, and NPMI.

[0167] Topic Diversity: Assess the degree of difference between different topics to avoid situations where topics are too similar or repetitive;

[0168] Topic Stability: Evaluates the stability of the model in identifying topics across different time slices to avoid topic drift.

[0169] The PF-DTM model and the baseline model are trained using the training set data. Model parameters, such as the number of topics and the time window size, are tuned using the validation set data, and the best-performing model is selected. The final model is evaluated using the test set data and compared with the baseline model. The performance of the PF-DTM model and the baseline model is compared in terms of perplexity, topic consistency, topic diversity, and topic stability.

[0170] This study analyzes whether the topics identified by the PF-DTM model conform to knowledge in the power industry. For example, it analyzes the semantic information of high-probability words under each topic and whether the evolution trend of the topics over time conforms to the operating rules of the power system. It also compares and analyzes the differences in topics identified by different models and the different models' ability to capture topic evolution trends.

[0171] In the results analysis and optimization phase, the performance of the PF-DTM model on different evaluation metrics and the impact of different model parameters on model performance are analyzed. The evolution trend of topics identified by the PF-DTM model over time is analyzed, including the emergence of new topics, the disappearance of old topics, and the merging and splitting of topics, and the evolution patterns are explained in conjunction with knowledge of power system operation and maintenance. Typical power operation and maintenance work order cases are selected to analyze the topic labeling results of the PF-DTM model for these cases, and how the model captures key information and evolution trends in the cases. Based on the experimental results, the model parameters are further optimized, such as adjusting the number of topics, time window size, and number of particles. More complex functions are attempted to model the evolution process of topics, such as nonlinear dynamic system models and Gaussian process models, to improve the model's ability to capture complex evolution trends. For example, deep learning models are attempted to learn vector representations of terms in the power field, or a more refined knowledge graph of power equipment is constructed to further improve the model's semantic understanding capabilities. In an optional embodiment, the PF-DTM model can be combined with other models, such as text classification models and sequence labeling models, to build a more comprehensive power operation and maintenance work order analysis system.

[0172] The beneficial effects of this invention are that, compared with the prior art, this invention provides a topic labeling method and system based on dynamic LDA, which combines the LDA model with the particle filtering algorithm and introduces an incremental learning mechanism, enabling it to effectively capture the evolution trend of power operation and maintenance work order topics over time. At the same time, it combines knowledge from fields such as power industry professional dictionaries, power equipment knowledge graphs, and power expert experience, further improving the accuracy and practicality of the model.

[0173] Specifically, by introducing a particle filtering algorithm, accurate modeling of the dynamic evolution of topic distribution is achieved, overcoming the limitation of traditional LDA models in handling dynamic text data. Simultaneously, the incremental Gibbs sampling algorithm is used for model training, enabling the model to efficiently handle the continuously accumulating characteristics of power operation and maintenance work order data, eliminating the need to retrain all historical data in each time slice, significantly improving model training efficiency. Power-specific information such as State Grid standards and specifications, expert systems, and knowledge base systems are converted into prior knowledge and integrated into the PF-DTM model training process, enhancing the model's ability to identify power-specific topics and strengthening its domain adaptability. Specifically, this invention explores three knowledge integration methods: integrating a power-specific dictionary into the topic-word distribution to improve the model's ability to recognize technical terms; utilizing entity and relational information from the power equipment knowledge graph to enhance the model's semantic understanding ability, for example, through semantic feature extraction and topic association constraints; and converting power expert experience into rules or constraints and integrating them into the model training process to guide the model to learn topic distributions and evolution patterns that better reflect reality. The expected experimental results show that, compared with traditional LDA models and other topic models, the PF-DTM model proposed in this invention achieves higher accuracy, recall, and F1 score in power operation and maintenance work order topic labeling tasks, verifying the effectiveness and superiority of the model. Furthermore, the PF-DTM model can identify more reasonable topic structures and evolution trends, providing more accurate topic information for subsequent tasks such as fault diagnosis, trend prediction, and knowledge discovery.

[0174] This invention can be a system, method, and / or computer program product. This invention also discloses a dynamic LDA-based topic tagging system based on the aforementioned dynamic LDA-based topic tagging method, comprising:

[0175] The data collection module is used to collect historical power operation and maintenance work order data from various heterogeneous data sources, convert the text data in the collected historical power operation and maintenance work order data into a unified format, perform preprocessing operations, and build a professional dictionary in the power field.

[0176] The dynamic LDA model training module is used to divide the power operation and maintenance work order data into different time slices according to the timestamp information of the power operation and maintenance work order data, and input the power operation and maintenance work order data of each time slice as prior information into the dynamic LDA model based on particle filtering to train the model. The model training uses the incremental Gibbs sampling algorithm to sample new data, update particle weights and model parameters, and realize the incremental update of the model.

[0177] The topic labeling module is used to optimize model parameters, store the trained particle filter-based dynamic LDA model corresponding to the current time slice for training in the next time slice, and use the trained particle filter-based dynamic LDA model to label new power operation and maintenance work order data.

[0178] Based on the spirit of this invention, those skilled in the art will readily conceive of obtaining a computer program product based on the aforementioned dynamic LDA-based topic labeling method. The computer program product may include a computer-readable storage medium on which computer-readable program instructions are loaded to cause a processor to implement various aspects of this disclosure. That is, this application also includes a terminal comprising a processor and a storage medium; the storage medium is used to store instructions; the processor is used to operate according to the instructions to perform the steps according to the aforementioned dynamic LDA-based topic labeling method.

[0179] Computer-readable storage media can be tangible devices capable of holding and storing instructions for use by an instruction execution device. Computer-readable storage media can be, for example, but not limited to, electrical storage devices, magnetic storage devices, optical storage devices, electromagnetic storage devices, semiconductor storage devices, or any suitable combination of the foregoing. More specific examples (a non-exhaustive list) of computer-readable storage media include: portable computer disks, hard disks, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), static random access memory (SRAM), portable compact disc read-only memory (CD-ROM), digital multifunction disc (DVD), memory sticks, floppy disks, mechanical encoding devices, such as punch cards or recessed protrusions storing instructions thereon, and any suitable combination of the foregoing. The computer-readable storage media used herein are not to be construed as transient signals themselves, such as radio waves or other freely propagating electromagnetic waves, electromagnetic waves propagating through waveguides or other transmission media (e.g., light pulses through fiber optic cables), or electrical signals transmitted through wires.

[0180] The computer-readable program instructions described herein can be downloaded from computer-readable storage media to various computing / processing devices, or downloaded via a network, such as the Internet, local area network, wide area network, and / or wireless network, to an external computer or external storage device. The network may include copper transmission cables, fiber optic transmission, wireless transmission, routers, firewalls, switches, gateway computers, and / or edge servers. A network adapter card or network interface in each computing / processing device receives the computer-readable program instructions from the network and forwards them to the computer-readable storage media in the respective computing / processing device.

[0181] Computer program instructions used to perform the operations of this disclosure may be assembly instructions, instruction set architecture (ISA) instructions, machine instructions, machine-dependent instructions, microcode, firmware instructions, status setting data, or source code or object code written in any combination of one or more programming languages, including object-oriented programming languages ​​such as Smalltalk, C++, etc., and conventional procedural programming languages ​​such as the "C" language or similar programming languages. The computer-readable program instructions may execute entirely on the user's computer, partially on the user's computer, as a standalone software package, partially on the user's computer and partially on a remote computer, or entirely on a remote computer or server. In cases involving a remote computer, the remote computer may be connected to the user's computer via any type of network—including a local area network (LAN) or a wide area network (WAN)—or may be connected to an external computer (e.g., via the Internet using an Internet service provider). In some embodiments, electronic circuitry, such as programmable logic circuitry, field-programmable gate arrays (FPGAs), or programmable logic arrays (PLAs), is personalized by utilizing the status information of the computer-readable program instructions to implement various aspects of this disclosure.

[0182] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit it. Although the present invention has been described in detail with reference to the above embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the specific implementation of the present invention. Any modifications or equivalent substitutions that do not depart from the spirit and scope of the present invention should be covered within the protection scope of the claims of the present invention.

Claims

1. A topic annotation method based on dynamic LDA, characterized in that, Includes the following steps: Step 1: Collect historical power operation and maintenance work order data from multiple heterogeneous data sources, convert the text data in the collected historical power operation and maintenance work order data into a unified format, perform preprocessing operations, and build a professional dictionary in the power field. Step 2: Based on the timestamp information of the power operation and maintenance work order data, divide the power operation and maintenance work order data into different time slices, and use the power operation and maintenance work order data of each time slice as prior information to input the dynamic LDA model based on particle filtering to train the model. The model training adopts the incremental Gibbs sampling algorithm to sample new data, update particle weights and model parameters, and realize the incremental update of the model. Step 3: Optimize the model parameters, store the trained particle filter-based dynamic LDA model corresponding to the current time slice for training in the next time slice, and use the trained particle filter-based dynamic LDA model to perform topic labeling on the new power operation and maintenance work order data. The data sources include internal data from the State Grid Corporation of China and publicly available external data. The internal data from the State Grid Corporation of China includes: Technical standards, specifications, and procedures are textual data used to define the operation and maintenance specifications for power equipment; The power equipment operation and maintenance manual includes the equipment's structure, operating principles, common faults, and troubleshooting methods. The fault diagnosis case library contains detailed records of historical fault cases, including fault symptoms, cause analysis, and handling measures. The externally disclosed data includes: Electricity industry news, including power system operation status, application of new technologies, and failure events; Technical forum discussion threads contain analyses and discussions of fault cases by power industry experts and engineers; The preprocessing operation further includes: First, data cleaning is performed to remove irrelevant information, including HTML tags, special characters, and punctuation marks, and the data format is standardized. Then, data deduplication is performed, and missing fields are filled in or deleted. Then, a power industry-specific word segmentation tool was used for word segmentation. A new word discovery algorithm was used to identify new words, professional terms and equipment models in the power field and add them to the word segmentation dictionary. Construct a stop dictionary, remove words whose contribution to the semantics of the text is below a threshold, and add stop words specific to the power operation and maintenance work order domain; A stemming algorithm is used to convert words with different word forms into the same stemming form; The word segmentation results are tagged with part-of-speech tags to identify the grammatical attributes of words; named entity recognition tools are used to identify named entities and professional entities in the power industry in the text. The step of using power maintenance work order data from each time slice as prior information input into the particle filter-based dynamic LDA model further includes: Using a professional dictionary in the power industry as prior knowledge, the dictionary information is integrated into the model. When calculating word frequency, professional words in the dictionary are given higher weights; or, in the Dirichlet prior distribution of topic-word distribution, professional words in the dictionary are given higher prior probabilities; or, the dictionary information is used to construct a topic-word association matrix as prior information for topic-word distribution, guiding the model to learn the topic-word distribution of knowledge in the power industry. Collect relevant information about power equipment, and use knowledge representation and reasoning techniques to construct a knowledge graph of power equipment. Based on the knowledge graph of power equipment, extract the semantic features of entities in power operation and maintenance work orders, use these semantic features as input features of the PF-DTM model, and integrate them into the bag-of-words representation of the document. Utilize the relational information in the knowledge graph to establish association constraints between topics; We collect and organize the experience and rules of power operation and maintenance experts in fault diagnosis and fault handling, and convert the expert experience into a rule form that can be understood and executed by computers. We then integrate the expert rules as constraints into the training process of the PF-DTM model to guide the model to learn the distribution and evolution of topics.

2. The topic annotation method based on dynamic LDA according to claim 1, characterized in that, The incremental Gibbs sampling algorithm samples new data, updates particle weights and model parameters, and achieves incremental updates of the model, further including: Step 2.1: For the data in the first time slice, train the standard PF-DTM model to obtain the initial topic-word distribution φ, document-topic distribution θ, and particle set; Step 2.2: For a new time slice t, the particle set and model parameters of the previous time slice t-1 are used as prior information. For each particle, the other particles and model parameters are kept unchanged. Based on the data of the current time slice and the prior information, the topic assignment of the particle is sampled and updated using the Gibbs sampling algorithm. Step 2.3: Based on the updated particle set, re-estimate the parameters of the topic-word distribution φ and the state transition function g; recalculate the weight of each particle based on the updated model parameters; resample based on the particle weights to generate a new particle set for incremental updates in the next time slice. Step 2.4: Repeat steps 2.2 and 2.3 until all time slices of data have been processed.

3. The topic annotation method based on dynamic LDA according to claim 2, characterized in that, Step 2.3 further includes: For each time slice, N particles are randomly generated, each particle representing a possible topic distribution θt; the weight of all particles is initialized to 1 / N, indicating that each particle has the same probability in the initial state; For each time slice t, the probability of generating observation data for the topic distribution represented by each particle is calculated according to the observation equation, and used as the weight of each particle; resampling is performed according to the weight of the particles to generate new N particles; Estimate the topic distribution θt for time slice t based on the resampled particles; Based on the estimated topic distribution θt and the observed data, update the model parameters, including the topic-word distribution φ and the parameters of the state transition function g; Repeat the above steps until the model converges.

4. The topic annotation method based on dynamic LDA according to claim 3, characterized in that, The optimization of model parameters further includes: Select the time slice size Δt based on the data characteristics and analysis objectives; The effects of different numbers of topics K were evaluated using perplexity and topic consistency metrics, and the optimal value for the number of topics was selected. The number of particles N is selected based on model accuracy and computational efficiency; Choose the state transition function g based on prior knowledge of the data.

5. A topic labeling system based on dynamic LDA according to any one of claims 1-4, characterized in that, include: The data collection module is used to collect historical power operation and maintenance work order data from various heterogeneous data sources, convert the text data in the collected historical power operation and maintenance work order data into a unified format, perform preprocessing operations, and build a professional dictionary in the power field. The dynamic LDA model training module is used to divide the power operation and maintenance work order data into different time slices according to the timestamp information of the power operation and maintenance work order data, and input the power operation and maintenance work order data of each time slice as prior information into the dynamic LDA model based on particle filtering to train the model. The model training uses the incremental Gibbs sampling algorithm to sample new data, update particle weights and model parameters, and realize the incremental update of the model. The topic labeling module is used to optimize model parameters, store the trained particle filter-based dynamic LDA model corresponding to the current time slice for training in the next time slice, and use the trained particle filter-based dynamic LDA model to label new power operation and maintenance work order data.

6. A terminal, comprising a processor and a storage medium; characterized in that: The storage medium is used to store instructions; The processor is configured to operate according to the instructions to perform the steps of the topic annotation method based on dynamic LDA according to any one of claims 1-4.

7. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the program is executed by the processor, it implements the steps of the topic annotation method based on dynamic LDA as described in any one of claims 1-4.

Images