CN112307218B - Construction method of fault diagnosis knowledge base for typical equipment of intelligent power plant based on knowledge graph - Google Patents
Construction method of fault diagnosis knowledge base for typical equipment of intelligent power plant based on knowledge graphDownload PDFInfo
- Publication number
- CN112307218B CN112307218BCN202011131638.0ACN202011131638ACN112307218BCN 112307218 BCN112307218 BCN 112307218BCN 202011131638 ACN202011131638 ACN 202011131638ACN 112307218 BCN112307218 BCN 112307218B
- Authority
- CN
- China
- Prior art keywords
- fault
- text
- fault diagnosis
- data
- word
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Active
Links
Images
Classifications
- G—PHYSICS
- G06—COMPUTING OR CALCULATING; COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F16/00—Information retrieval; Database structures therefor; File system structures therefor
- G06F16/30—Information retrieval; Database structures therefor; File system structures therefor of unstructured textual data
- G06F16/36—Creation of semantic tools, e.g. ontology or thesauri
- G06F16/367—Ontology
- G—PHYSICS
- G06—COMPUTING OR CALCULATING; COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F40/00—Handling natural language data
- G06F40/20—Natural language analysis
- G06F40/205—Parsing
- G06F40/216—Parsing using statistical methods
- G—PHYSICS
- G06—COMPUTING OR CALCULATING; COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F40/00—Handling natural language data
- G06F40/20—Natural language analysis
- G06F40/237—Lexical tools
- G06F40/242—Dictionaries
- G—PHYSICS
- G06—COMPUTING OR CALCULATING; COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F40/00—Handling natural language data
- G06F40/20—Natural language analysis
- G06F40/279—Recognition of textual entities
- G06F40/289—Phrasal analysis, e.g. finite state techniques or chunking
- G—PHYSICS
- G06—COMPUTING OR CALCULATING; COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/04—Architecture, e.g. interconnection topology
- G06N3/045—Combinations of networks
- G—PHYSICS
- G06—COMPUTING OR CALCULATING; COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N5/00—Computing arrangements using knowledge-based models
- G06N5/02—Knowledge representation; Symbolic representation
- G—PHYSICS
- G06—COMPUTING OR CALCULATING; COUNTING
- G06Q—INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES; SYSTEMS OR METHODS SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES, NOT OTHERWISE PROVIDED FOR
- G06Q50/00—Information and communication technology [ICT] specially adapted for implementation of business processes of specific business sectors, e.g. utilities or tourism
- G06Q50/06—Energy or water supply
- Y—GENERAL TAGGING OF NEW TECHNOLOGICAL DEVELOPMENTS; GENERAL TAGGING OF CROSS-SECTIONAL TECHNOLOGIES SPANNING OVER SEVERAL SECTIONS OF THE IPC; TECHNICAL SUBJECTS COVERED BY FORMER USPC CROSS-REFERENCE ART COLLECTIONS [XRACs] AND DIGESTS
- Y04—INFORMATION OR COMMUNICATION TECHNOLOGIES HAVING AN IMPACT ON OTHER TECHNOLOGY AREAS
- Y04S—SYSTEMS INTEGRATING TECHNOLOGIES RELATED TO POWER NETWORK OPERATION, COMMUNICATION OR INFORMATION TECHNOLOGIES FOR IMPROVING THE ELECTRICAL POWER GENERATION, TRANSMISSION, DISTRIBUTION, MANAGEMENT OR USAGE, i.e. SMART GRIDS
- Y04S10/00—Systems supporting electrical power generation, transmission or distribution
- Y04S10/50—Systems or methods supporting the power network operation or management, involving a certain degree of interaction with the load-side end user applications
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Physics & Mathematics (AREA)
- General Physics & Mathematics (AREA)
- General Engineering & Computer Science (AREA)
- Health & Medical Sciences (AREA)
- Computational Linguistics (AREA)
- Artificial Intelligence (AREA)
- General Health & Medical Sciences (AREA)
- Business, Economics & Management (AREA)
- Audiology, Speech & Language Pathology (AREA)
- Data Mining & Analysis (AREA)
- Software Systems (AREA)
- Life Sciences & Earth Sciences (AREA)
- Computing Systems (AREA)
- Mathematical Physics (AREA)
- Evolutionary Computation (AREA)
- Economics (AREA)
- Water Supply & Treatment (AREA)
- Marketing (AREA)
- Probability & Statistics with Applications (AREA)
- Biophysics (AREA)
- Public Health (AREA)
- Molecular Biology (AREA)
- Human Resources & Organizations (AREA)
- Biomedical Technology (AREA)
- Primary Health Care (AREA)
- Strategic Management (AREA)
- Tourism & Hospitality (AREA)
- General Business, Economics & Management (AREA)
- Animal Behavior & Ethology (AREA)
- Databases & Information Systems (AREA)
- Testing And Monitoring For Control Systems (AREA)
Abstract
Translated fromChinese
Description
Translated fromChinese技术领域technical field
本发明属于电厂生产设备运行故障诊断领域，包含故障诊断知识图谱的设计与构建方法，以及知识库的应用方案，知识库的更新策略。The invention belongs to the field of fault diagnosis of power plant production equipment operation, and includes a design and construction method of a fault diagnosis knowledge map, an application scheme of a knowledge base, and an update strategy of the knowledge base.
背景技术Background technique
智能电厂是在信息化与工业化深度融合的背景下提出的，旨在提升电力行业的智能化水平，实现无人巡检、自动故障诊断与处理、大数据分析与智能控制等技术提升。其中，故障诊断与处理对维持发电过程稳定进行、保障生产安全至关重要。对智能电厂关键设备的故障诊断，关键是构造故障诊断知识库，实现自动、可靠的故障诊断。通过对故障案例的收集和分析，利用知识图谱等前沿技术，有效处理自然语言，构建故障诊断知识库，可为实现智能电厂典型设备故障诊断提供基础。The smart power plant is proposed in the context of the deep integration of informatization and industrialization. It aims to improve the intelligence level of the power industry and realize technological advancements such as unmanned inspection, automatic fault diagnosis and processing, big data analysis and intelligent control. Among them, fault diagnosis and treatment are very important to maintain the stability of the power generation process and ensure the safety of production. The key to the fault diagnosis of key equipment in smart power plants is to construct a fault diagnosis knowledge base to achieve automatic and reliable fault diagnosis. Through the collection and analysis of fault cases, the use of cutting-edge technologies such as knowledge graphs, the effective processing of natural language, and the construction of a fault diagnosis knowledge base can provide a basis for the realization of fault diagnosis of typical equipment in smart power plants.
知识图谱技术在2012年由Google公司率先应用于其搜索引擎中，极大提升了其搜索结果的质量。近年来，知识图谱与各行业开始进行深度融合，旨在解决行业痛点问题、降低人力成本。例如，临床数据相对受限的医疗领域也建立了相应的医学知识图谱，用于诊断常见的儿科疾病与部分危急重症，这种AI诊断模型对儿科疾病的临床平均准确率达90％，其表现可媲美低年资的主治医生。通过构建故障诊断知识库，可以实现智能电厂日常运行过程中自动、可靠、高效的故障诊断，从而形成高度智能化的决策，主动为电厂操作人员提供具体、有效的指导建议。当下，积累下来的故障诊断知识大多为非结构化数据，直接应用起来较为困难，需要重新梳理成各部分有机联系的知识图谱形式，从而对知识的存储、检索、推理、应用都具有重要意义。The knowledge graph technology was first applied to its search engine by Google in 2012, which greatly improved the quality of its search results. In recent years, knowledge graphs have begun to be deeply integrated with various industries, aiming to solve industry pain points and reduce labor costs. For example, medical fields with relatively limited clinical data have also established corresponding medical knowledge maps to diagnose common pediatric diseases and some critical illnesses. This AI diagnostic model has an average clinical accuracy rate of 90% for pediatric diseases, and its performance Comparable to junior attending physicians. By building a fault diagnosis knowledge base, automatic, reliable and efficient fault diagnosis in the daily operation of smart power plants can be realized, so as to form highly intelligent decision-making, and actively provide specific and effective guidance and suggestions for power plant operators. At present, most of the accumulated fault diagnosis knowledge is unstructured data, which is difficult to apply directly. It needs to be reorganized into the form of knowledge graph with organic connection of various parts, which is of great significance to the storage, retrieval, reasoning and application of knowledge.
近年来自然语言处理技术蓬勃发展，例如情感分析、文本摘要等有监督任务的表现都在不断提高。seq2seq等端到端的机器翻译模型也显著改进了当下的机器翻译效果。然而对于工业故障诊断文本的特征提取，常常是没有标签或者获取标签是十分昂贵的。并且，已有的故障诊断知识图谱方法通常以某种设备或具体故障术语为节点，以参数、表征、原因等几种特定的类型为边构建三元组，这种方法受到三元组自身的表意能力限制，描述具有复杂原因、复杂解决方案的能力受限。因此本发明采用了自监督学习的思想结合编码器——解码器模型来获取知识图谱中节点中的文本的特征向量，为之后需要高效计算和推理的任务提供支持。In recent years, natural language processing technology has flourished, and the performance of supervised tasks such as sentiment analysis and text summarization has been continuously improved. End-to-end machine translation models such as seq2seq have also significantly improved current machine translation performance. However, for feature extraction of industrial fault diagnosis texts, there are often no labels or it is very expensive to obtain labels. In addition, the existing fault diagnosis knowledge graph method usually uses a certain equipment or specific fault term as a node, and uses several specific types such as parameters, representations, and causes as edges to construct triples. This method is affected by the triple itself. Limitation of ideographic ability, limited ability to describe complex causes and complex solutions. Therefore, the present invention adopts the idea of self-supervised learning combined with the encoder-decoder model to obtain the feature vector of the text in the nodes in the knowledge graph, so as to provide support for subsequent tasks requiring efficient calculation and reasoning.
发明内容SUMMARY OF THE INVENTION
本发明的目的在于将机器学习和知识图谱技术应用于智能电厂典型设备故障诊断领域，通过设计针对故障诊断领域的知识图谱架构，构建智能电厂故障诊断知识库，为智能电厂故障诊断中涉及到的知识存储、检索、推理、应用、更新提供支持。The purpose of the present invention is to apply machine learning and knowledge graph technology to the field of fault diagnosis of typical equipment in intelligent power plants, and to construct a knowledge base for fault diagnosis of intelligent power plants by designing a knowledge graph architecture for the field of fault diagnosis, which is the most important part of fault diagnosis of intelligent power plants. Knowledge storage, retrieval, reasoning, application, and updating provide support.
本发明的目的是通过以下技术方案来实现的：基于知识图谱的智能电厂典型设备故障诊断知识库构建方法，该方法包括以下步骤：The object of the present invention is achieved through the following technical solutions: a method for constructing a fault diagnosis knowledge base for typical equipment of an intelligent power plant based on a knowledge graph, the method comprises the following steps:
步骤1)收集原始数据。数据的来源包括互联网和合作电厂。合作电厂的数据质量高但数量少；互联网上数据多但质量较差。足够充足的数据来源一方面可以增大知识库的规模，另一方面也为之后的训练提供充足的训练样本。Step 1) Collect raw data. Sources of data include the Internet and cooperative power plants. The data of the cooperative power plants are of high quality but low quantity; the data on the Internet is abundant but of poor quality. Sufficient data sources can increase the size of the knowledge base on the one hand, and provide sufficient training samples for subsequent training on the other hand.
步骤2)对多模态数据进行针对性的预处理，将非文本数据转换为文本数据。Step 2) Targeted preprocessing is performed on the multimodal data, and the non-text data is converted into text data.
步骤3)对文本数据进行处理，构建“双层——三要素”的知识图谱。“双层”指设备层、故障层。既便于故障诊断的落地应用，也利于使用设备检索故障。设备层基于专家提供、领域术语词典、TF-IDF算法提取出的关键词构建。故障层包括故障诊断的“三要素”：故障描述、故障诊断、处理意见。从而得到了故障诊断知识图谱；Step 3) Process the text data to construct a "two-layer-three-element" knowledge graph. "Double layer" refers to the equipment layer and the fault layer. It is not only convenient for the landing application of fault diagnosis, but also conducive to the use of equipment to retrieve faults. The device layer is constructed based on keywords provided by experts, domain term dictionary, and keywords extracted by TF-IDF algorithm. The fault layer includes the "three elements" of fault diagnosis: fault description, fault diagnosis, and handling opinions. Thus, the fault diagnosis knowledge graph is obtained;
步骤4)对故障描述文本、故障诊断文本、处理意见文本进行进一步处理，包括分句、分词、BPE处理，构建用于双向GRU网络提取文本特征的训练集；Step 4) further processing the fault description text, fault diagnosis text, and processing opinion text, including sentence segmentation, word segmentation, and BPE processing, and constructing a training set for bidirectional GRU network extraction of text features;
步骤5)构建并训练基于双向GRU网络、注意力机制的编码器——解码器模型，从编码器输出的状态得到无标签文本的特征向量。冻结训练好的网络参数、并存储获得的特征向量；Step 5) Build and train an encoder-decoder model based on bidirectional GRU network and attention mechanism, and obtain the feature vector of unlabeled text from the state output by the encoder. Freeze the trained network parameters and store the obtained feature vector;
步骤6)应用得到的特征向量，结合设备层与领域词典提供的关键术语，将故障现场图、过程数据生成的文本描述与原有文本数据的故障描述进行对齐。可消除已有知识库的冗余，或基于故障诊断知识图谱进行推理、检索；Step 6) Applying the obtained feature vector and combining the key terms provided by the device layer and the domain dictionary, align the fault scene diagram and the text description generated by the process data with the fault description of the original text data. It can eliminate the redundancy of the existing knowledge base, or perform reasoning and retrieval based on the fault diagnosis knowledge graph;
其中，对于新增故障诊断知识、采用2)、3)、4)的过程进行处理后，如果其中没有出现词典未收录的关键术语，则使用5)中存储的参数得到编码结果，使用解码器结合柱搜索算法得到解码结果，解码结果和原子句进行比对，如一致则通过检验，将新知识并入原有知识图谱，实现知识图谱的更新。Among them, after the process of 2), 3), and 4) is used for the newly added fault diagnosis knowledge, if there is no key term not included in the dictionary, the parameters stored in 5) are used to obtain the encoding result, and the decoder is used. Combined with the column search algorithm, the decoding results are obtained, and the decoding results are compared with the atomic sentences. If they are consistent, the test is passed, and the new knowledge is merged into the original knowledge map to realize the update of the knowledge map.
进一步地，所述步骤2)具体为：Further, the step 2) is specifically:
针对图片数据，对故障描述对应的故障现场图足够充足的常发故障，采用基于GAN的图像文本生成技术生成故障描述；对于样本数极少的偶发故障，人工生成故障描述文本。针对传感器收集到的包含故障时段的生产数据，在已知数据正常范围的情况下，据此确定正常数据所在的3-Sigma的阈值范围，同时检测时序数据异常点，超出这个范围的数据，都归属于异常数据。将异常数据变量和其相关变量的高预警或低预警转化为文本形式的故障描述。For the picture data, if the fault scene map corresponding to the fault description is sufficient and frequent faults, the GAN-based image text generation technology is used to generate the fault description; for the occasional fault with a small number of samples, the fault description text is manually generated. For the production data collected by the sensor including the fault period, when the normal range of the data is known, the 3-Sigma threshold range where the normal data is located is determined accordingly, and the abnormal points of the time series data are detected at the same time. Attributable to abnormal data. Convert abnormal data variables and high or low warnings of their related variables into textual fault descriptions.
进一步地，所述步骤3)具体为：Further, the step 3) is specifically:
首先从文本数据中提取出故障描述、故障诊断、处理意见。对于设备层，将设备层划分为两个子层，电厂中典型设备整体的名称为顶层节点，典型设备的具体部件为底层节点，典型设备整体的名称与其相应的具体部件以“包含”为关系进行连接。其中，除了源于领域词典和专家提供的关键词外，还引入TF-IDF算法提取的关键词，并在结果中去除停用词，如下：Firstly, the fault description, fault diagnosis and processing opinions are extracted from the text data. For the equipment layer, the equipment layer is divided into two sub-layers. The name of the whole typical equipment in the power plant is the top node, and the specific components of the typical equipment are the bottom nodes. connect. Among them, in addition to the keywords derived from the domain dictionary and experts, the keywords extracted by the TF-IDF algorithm are also introduced, and stop words are removed from the results, as follows:
TF-IDF＝TF*IDF#(20)TF-IDF=TF*IDF#(20)
其中TF为词频，IDF为逆文档频率。语料库是指所收集到的故障描述文档的集合。取TF-IDF值排名较大靠前的词也作为设备层候选节点。where TF is the term frequency and IDF is the inverse document frequency. Corpus refers to a collection of collected fault description documents. The word with the larger TF-IDF value is also used as the candidate node of the device layer.
故障层以故障描述文本、故障诊断文本、处理意见文本为节点，对应的故障描述文本与故障诊断文本以“诊断”为关系进行连接；对应的故障描述文本与处理意见文本以“处理”为关系进行连接，形成故障描述层。而后将故障层中故障描述文本节点与其中涉及到的具体部件节点相连接，从而形成“两层——三要素”的塔形故障诊断知识图谱架构。The fault layer takes fault description text, fault diagnosis text, and processing opinion text as nodes, and the corresponding fault description text and fault diagnosis text are connected by the relationship of "diagnosis"; the corresponding fault description text and processing opinion text are related to "processing" Make connections to form a fault description layer. Then, the fault description text nodes in the fault layer are connected with the specific component nodes involved, so as to form a tower-shaped fault diagnosis knowledge graph architecture of "two layers - three elements".
进一步地，所述步骤4)具体为：首先对文本进行分句，将文本从逗号、句号、冒号处分割成一个个子句。对每一个子句，使用结巴(jieba)分词工具进行分词。然后进行BPE处理，得到BPE处理后的子句和词典，以每一个处理后的子句作为单个训练样本。Further, the step 4) is specifically as follows: firstly, the text is divided into clauses, and the text is divided into clauses from commas, periods and colons. For each clause, use the jieba word segmentation tool for word segmentation. Then perform BPE processing to obtain BPE-processed clauses and dictionaries, and use each processed clause as a single training sample.
进一步地，所述步骤5)具体为：采用编码器——解码器框架和注意力机制构建使用GRU(门控循环单元)的双向循环神经网络(Bi-direction Recurrent Neural Network)。通过每一个子句同时作为源语句和目标语句，自监督获取每一个子句的特征向量。其中GRU具体包括：Further, the step 5) is specifically: using an encoder-decoder framework and an attention mechanism to construct a Bi-direction Recurrent Neural Network (Bi-direction Recurrent Neural Network) using GRU (Gated Recurrent Unit). By using each clause as both a source sentence and a target sentence, self-supervision obtains the feature vector of each clause. The GRU specifically includes:
首先将每个用于编码器输入端的子句样本中的每个词采用独热编码转换为一维向量，每一个向量的长度和BPE处理得到的词典大小相同，其中只有该词对应的位置为1，其余位置为0。然后使用embedding层进行降维映射，映射矩阵大小为K*V，其中K为设定的词向量维度，V为词典大小。将映射矩阵与子句独热编码形成的矩阵相乘从而对词向量降维，得到词向量组x＝{x1,x2,…,xt,…,xT}，T表示子句中对应的词数量。在模型训练过程中，对于每个字句样本对应的期望输出语句，采用相同的方法进行处理，得到u＝{u1,u2,…,ut,…,uT}。First, each word in each clause sample used for the input of the encoder is converted into a one-dimensional vector by one-hot encoding. The length of each vector is the same as the size of the dictionary obtained by BPE processing, and only the corresponding position of the word is 1, the rest of the positions are 0. Then use the embedding layer to perform dimension reduction mapping. The size of the mapping matrix is K*V, where K is the set word vector dimension and V is the dictionary size. Multiply the mapping matrix and the matrix formed by the one-hot encoding of the clause to reduce the dimension of the word vector, and obtain the word vector group x={x1 ,x2 ,...,xt ,...,xT }, T represents the number of words in the clause the corresponding number of words. In the model training process, the expected output sentence corresponding to each sentence sample is processed in the same way, and u={u1 ,u2 ,...,ut ,...,uT } is obtained.
通过上一刻的状态ht-1和当前节点的输入xt来获取重置门r和更新门z：The reset gate r and update gate z are obtained by the state ht-1 of the last moment and the input xt of the current node:
r＝sigmoid(wr*[ht-1,xt]) (21)r=sigmoid(wr *[ht-1 ,xt ]) (21)
z＝sigmoid(wz*[ht-1,xt]) (22)z=sigmoid(wz *[ht-1 ,xt ]) (22)
其中，x＝{x1,x2,…,xt,…,xT}为每一个子句样本经过上述独热编码和映射过程得到的词向量组，t即当前时刻，表示当前所输入的词在子句中所处的位置。sigmoid函数将数值映射到0-1范围内。wr、wz均是需要学习的参数。Among them, x={x1 ,x2 ,...,xt ,...,xT } is the word vector group obtained by each clause sample through the above one-hot encoding and mapping process, t is the current moment, indicating the current input the position of the word in the clause. The sigmoid function maps numeric values to the range 0-1. Both wr and wz are parameters that need to be learned.
接着获得当前时刻的状态与输出Then get the state and output of the current moment
其中w为需要学习的参数。ht为当前时间步的输出。where w is the parameter to be learned. ht is the output of the current time step.
对于编码器部分，双向循环神经网络分别在时间维度上以前向和后向处理输入序列，并将每个时间步的输出拼接作为最终的特征向量输出。For the encoder part, the bidirectional recurrent neural network processes the input sequence forward and backward separately in the time dimension, and concatenates the output of each time step as the final feature vector output.
其中xt为独热编码后降维得到的词向量，
对于解码器部分，应用注意力机制，每一个时刻，根据由公式(29)计算出的第t个词的上下文向量ct,目标序列第t个词向量ut和t时刻隐藏状态zt，计算出下一个隐层状态zt+1：For the decoder part, the attention mechanism is applied. At each moment, according to the context vector ct of the t-th word calculated by formula (29), the t-th word vector ut of the target sequence and the hidden state zt at time t , Calculate the next hidden layer state zt+1 :
其中权重aij表示目标词i对源词j的注意力大小，align为对齐模型，用于衡量目标词i对源词j的匹配程度。The weight aij represents the attention of the target word i to the source word j, and align is the alignment model, which is used to measure the matching degree of the target word i to the source word j.
将zt+1通过softmax归一化，得到目标序列第t+1个词的概率分布pt+1,使用交叉熵函数得到t+1的代价，对所有时刻取平均得到总的损失函数：Normalize zt+1 by softmax to get the probability distribution pt+1 of the t+1th word in the target sequence, use the cross entropy function to get the cost of t+1, and average all the moments to get the total loss function:
pt+1＝softmax(wszt+1+b) (32)pt+1 =softmax(ws zt+1 +b) (32)
其中avg为求平均函数，cross_entropy为交叉熵函数。ws,b为需要学习的参数。冻结训练好的网络参数、并存储获得的特征向量；Where avg is the averaging function, and cross_entropy is the cross entropy function. ws , b are parameters that need to be learned. Freeze the trained network parameters and store the obtained feature vectors;
进一步地，所述步骤6)具体为：使用步骤5)中训练得到的编码器，对故障现场图、过程数据生成的文本描述进行编码，得到特征向量，与已有的故障描述文本的特征向量计算余弦相似度，如下：Further, the step 6) is specifically: using the encoder trained in step 5) to encode the text description generated by the fault site map and the process data to obtain a feature vector, and the feature vector of the existing fault description text. Calculate the cosine similarity as follows:
其中，A、B分别表示新计算的和已有的故障描述文本的特征向量，两向量维度相同。n表示A向量和B向量的维度。Among them, A and B respectively represent the feature vectors of the newly calculated and existing fault description texts, and the two vectors have the same dimension. n represents the dimension of A vector and B vector.
将相似度最高的一组进行对齐。通过将已有的故障描述文本的特征向量两两计算相似度，可以对相似度高的故障描述节点进行合并，消除冗余，从而获得智能电厂典型设备故障诊断知识库，用于后续应用。Align the group with the highest similarity. By calculating the similarity of the feature vectors of the existing fault description texts, the fault description nodes with high similarity can be merged to eliminate redundancy, so as to obtain the fault diagnosis knowledge base of typical equipment of smart power plants for subsequent applications.
进一步地，还包括构建用于工业落地应用的GUI界面步骤，所述GUI功能包括磨煤机故障诊断、查询历史、近期检修情况等。Further, it also includes the step of constructing a GUI interface for industrial landing applications, and the GUI functions include coal mill fault diagnosis, query history, recent maintenance and the like.
进一步地，磨煤机故障诊断可通过故障描述或传感器数据进行故障诊断，返回故障诊断结果、检修建议、检修故障图，具体为：使用编码器得到特征向量后，知识图谱中进行相似度比较，返回相似度最高的故障描述文本对应的故障诊断文本、处理意见文本和故障图。Further, the fault diagnosis of the coal mill can be carried out through fault description or sensor data, and the fault diagnosis results, maintenance suggestions, and maintenance fault diagrams can be returned. Return the fault diagnosis text, processing opinion text and fault graph corresponding to the fault description text with the highest similarity.
进一步地，对于新增的故障诊断知识，对其文本进行编码获取特征向量后获取特征向量，采用柱搜索算法进行解码。在解码的过程中，不断通过pt+1采样ut+1。对于柱搜索算法，使用广度优先策略建立搜索树，在树的每一层，按照生成词的log概率之和为启发代价对节点进行排序，然后仅留下预先确定的个数的节点，直到获得句子结束标记或超过最大生成长度为止。Further, for the newly added fault diagnosis knowledge, the text is encoded to obtain the feature vector, and then the feature vector is obtained, and the column search algorithm is used for decoding. During the decoding process, ut+1 is continuously sampled by pt+1. For the column search algorithm, a breadth-first strategy is used to build a search tree. At each layer of the tree, the nodes are sorted according to the sum of the log probabilities of the generated words as the heuristic cost, and then only a predetermined number of nodes are left until the End-of-sentence marker or until the maximum generated length is exceeded.
本发明设计并构建了针对于电厂典型设备故障诊断这一应用场景的知识图谱形式的知识库，将非结构化的多模态数据统一为结构化的知识图谱形式。并且使用双向GRU网络对知识图谱中节点中的文本描述进行编码，为基于知识图谱的高性能推理、应用任务提供基础。本发明对于提升智能电厂典型设备故障诊断具有重要意义。The present invention designs and builds a knowledge base in the form of a knowledge graph for the application scenario of fault diagnosis of typical equipment in a power plant, and unifies the unstructured multimodal data into a structured knowledge graph form. And the bidirectional GRU network is used to encode the text description in the nodes in the knowledge graph, which provides the basis for high-performance reasoning and application tasks based on the knowledge graph. The invention has important significance for improving the fault diagnosis of typical equipment of the intelligent power plant.
附图说明Description of drawings
图1：本发明流程示意图；Fig. 1: the schematic flow chart of the present invention;
图2：多模态数据示意图；Figure 2: Schematic diagram of multimodal data;
图3：磨煤机D磨碗差压异常示例图；其中a、b为短周期示意图，其中b带预警标记，c为长周期示意图；Figure 3: An example of abnormal differential pressure in the grinding bowl of coal mill D; where a and b are short-cycle schematic diagrams, where b is marked with an early warning mark, and c is a long-period schematic diagram;
图4：磨煤机的短周期风粉混合物压力(a)、一次风压力(b)、磨煤机电流图(c)和长周期风粉混合物压力(d)、一次风压力(e)、磨煤机电流图(f)；Figure 4: Short-cycle air-powder mixture pressure (a), primary air pressure (b), coal mill current diagram (c) and long-cycle air-powder mixture pressure (d), primary air pressure (e), Coal mill current diagram (f);
图5：“双层——三要素”知识图谱架构图；Figure 5: "Double-layer - three elements" knowledge graph architecture diagram;
图6：基于双向RNN的编码器——解码器结构图；Figure 6: Encoder-decoder structure diagram based on bidirectional RNN;
图7：GRU单元结构图；Figure 7: GRU unit structure diagram;
图8：故障原因树示意图；Figure 8: Schematic diagram of fault cause tree;
图9：GUI功能示意图，其中，a为主界面、b为磨煤机故障诊断，c为磨煤机故障查询历史；d为磨煤机近期检修情况。Figure 9: Schematic diagram of GUI functions, in which a is the main interface, b is the fault diagnosis of the coal mill, c is the fault query history of the coal mill, and d is the recent maintenance of the coal mill.
具体实施方式Detailed ways
下面结合附图和具体实例，对本发明作进一步详细说明。The present invention will be described in further detail below in conjunction with the accompanying drawings and specific examples.
本发明收集来自互联网和浙江省某电厂提供的发电过程故障诊断描述案例，包含文本、图像、数值形式的传感器数据等多模态数据。通过将非结构化数据以特定的结构组织成结构化的知识图谱形式的知识库，提升故障的自动化诊断能力。The invention collects fault diagnosis description cases of power generation process from the Internet and a power plant in Zhejiang Province, including text, images, sensor data in the form of numerical values and other multimodal data. By organizing unstructured data into a knowledge base in the form of a structured knowledge graph in a specific structure, the automatic fault diagnosis capability is improved.
本发明的基于知识图谱的智能电厂典型设备故障诊断知识库构建包括以下步骤：The construction of the fault diagnosis knowledge base for the typical equipment of the intelligent power plant based on the knowledge graph of the present invention comprises the following steps:
步骤1)收集原始数据。数据的来源包括互联网和合作电厂。合作电厂提供的故障案例，质量较高，为word或pdf形式的文件，本身为半结构化形式，包含故障描述、故障诊断、处理意见、配套的数据、现场图等，可以直接进行提取。互联网上抓取的故障描述用于进一步扩增知识库。足够充足的数据来源一方面可以增大知识库的规模，另一方面也为之后的训练提供充足的训练样本。Step 1) Collect raw data. Sources of data include the Internet and cooperative power plants. The fault cases provided by the cooperative power plants are of high quality and are documents in word or pdf format, which are semi-structured and contain fault descriptions, fault diagnosis, handling opinions, supporting data, and site maps, etc., which can be directly extracted. Fault descriptions scraped from the Internet are used to further expand the knowledge base. Sufficient data sources can increase the size of the knowledge base on the one hand, and provide sufficient training samples for subsequent training on the other hand.
步骤2)对多模态数据，如图2所示，进行针对性的预处理，将非文本数据转换为文本数据，对图片形式的数据，如果同一故障种类下对应的图片样本数量较多，则训练基于GAN的图像文本生成技术生成故障描述；对于样本数极少的偶发故障，人工生成故障描述文本。Step 2) For multi-modal data, as shown in Figure 2, carry out targeted preprocessing, convert non-text data into text data, and for data in the form of pictures, if the number of corresponding picture samples under the same fault type is large, Then, the GAN-based image text generation technology is trained to generate fault descriptions; for occasional faults with a very small number of samples, fault description texts are artificially generated.
在知识图谱的构建过程中，针对传感器收集到的包含故障时段的生产数据，在已知每一个数据变量正常范围的情况下，据此确定每一个正常数据变量所在的3-Sigma的阈值范围，同时检测时序数据异常点，超出这个范围的数据变量，都归属于异常数据变量。某一异常数据变量与其他变量两两求相关系数，相关系数最高的几个变量认为其为该异常数据变量的相关变量。将异常数据变量和其相关变量的高预警或低预警转化为文本形式的故障描述。During the construction of the knowledge graph, for the production data collected by the sensor including the fault period, when the normal range of each data variable is known, the 3-Sigma threshold range where each normal data variable is located is determined accordingly. At the same time, abnormal points of time series data are detected, and data variables beyond this range are all attributed to abnormal data variables. The correlation coefficient between an abnormal data variable and other variables is calculated in pairs, and the variables with the highest correlation coefficient are considered as the relevant variables of the abnormal data variable. Convert abnormal data variables and high or low warnings of their related variables into textual fault descriptions.
3-Sigma原则定义如下：假设一组检测数据只含有随机误差，对原始数据进行计算处理得到标准差，然后按一定的概率确定一个区间，认为误差超过这个区间的就属于异常值。使用3-Sigma的前提是数据服从正态分布，满足这个条件之后，在3-Sigma范围(μ–3σ,μ+3σ)内99.73％的为正常数据，其中σ代表标准差，μ代表均值，x＝μ为图形的对称轴。The 3-Sigma principle is defined as follows: Assuming that a set of test data only contains random errors, the standard deviation is obtained by calculating and processing the original data, and then an interval is determined according to a certain probability, and the error exceeding this interval is considered to be an outlier. The premise of using 3-Sigma is that the data obey a normal distribution. After meeting this condition, 99.73% of the data within the range of 3-Sigma (μ–3σ, μ+3σ) are normal data, where σ represents the standard deviation, μ represents the mean, x=μ is the symmetry axis of the graph.
例如，对于一例“磨煤机D磨碗差压这一变量异常”的故障，如附图3所示，使用正常期间的数据计算故障测点和其他测点相关系数，最相关的为“磨煤机D出口选择后的风粉混合物压力”、“磨煤机D一次风压力”和“#6D磨煤机电流”三个变量，如下表所示。For example, for a fault case of "the variable differential pressure of the grinding bowl of coal mill D is abnormal", as shown in Figure 3, the correlation coefficient between the fault measuring point and other measuring points is calculated using the data during the normal period, and the most relevant one is "grinding mill D". The three variables of "air-pulverized mixture pressure after coal mill D outlet selection", "coal mill D primary air pressure" and "#6D coal mill current" are shown in the following table.
表1与故障点的相关系数表Table 1. Correlation coefficient table with fault point
如附图4所示，磨煤机D磨碗差压高预警曲线图，磨煤机D出口选择后的风粉混合物压力曲线图，磨煤机D一次风压力曲线图，#6D磨煤机电流曲线图及其相应的预警图如附图4所示。使用3-Sigma原则，对于本测试，取前30000min为正常时间，这个时间段内的数据没有发生大幅波动。则μ＝2.519,σ＝0.816，因此磨煤机D磨碗差压阈值的上下限分别为4.966和0.0720，发现当前数据超上限，如附图中标识的点区域所示。同样对于另外三个变量进行检验，发现并未超限。As shown in Figure 4, the high differential pressure warning curve of the grinding bowl of coal mill D, the pressure curve of the air-powder mixture after the outlet of coal mill D is selected, the primary air pressure curve of coal mill D, and the #6D coal mill The current curve diagram and its corresponding early warning diagram are shown in Figure 4. Using the 3-Sigma principle, for this test, the first 30,000 minutes are taken as the normal time, and the data during this time period does not fluctuate significantly. Then μ=2.519, σ=0.816, so the upper and lower limits of the differential pressure threshold of the grinding bowl of coal mill D are 4.966 and 0.0720, respectively. It is found that the current data exceeds the upper limit, as shown in the dotted area in the attached drawing. The other three variables were also tested and found that they did not exceed the limit.
则对于该故障，由数据生成的故障描述则为：“磨煤机D磨碗差压高预警，磨煤机D出口选择后的风粉混合物压力正常，磨煤机D一次风压力正常，磨煤机电流正常”。For this fault, the fault description generated by the data is: "The differential pressure of the grinding bowl of coal mill D is high, the pressure of the air-powder mixture after the outlet of coal mill D is selected is normal, the primary air pressure of coal mill D is normal, the The coal machine current is normal".
步骤3)对文本数据进行处理，如附图5所示，构建“双层——三要素”的知识图谱。如附图5所示，“双层”指设备层、故障层。故障层包括故障诊断的“三要素”：故障描述、故障诊断、处理意见。从而得到了故障诊断知识图谱。Step 3) Process the text data, as shown in Fig. 5, to construct a "two-layer-three-element" knowledge graph. As shown in Fig. 5, "double layer" refers to the equipment layer and the fault layer. The fault layer includes the "three elements" of fault diagnosis: fault description, fault diagnosis, and handling opinions. Thus, the fault diagnosis knowledge graph is obtained.
对于互联网渠道收集到的包含故障诊断描述的文本，基于因果关系的关联词进行初步定位，人工在定位区提炼故障描述、故障诊断、处理意见；对于工厂提供的故障诊断文本，本身具有一定结构，提取出故障描述、故障诊断、处理意见的部分即可。For the texts containing fault diagnosis descriptions collected from Internet channels, preliminary positioning is carried out based on the causal correlative words, and the fault description, fault diagnosis, and processing opinions are manually extracted in the positioning area; for the fault diagnosis text provided by the factory, it has a certain structure itself, and extraction The parts of fault description, fault diagnosis, and handling opinions are sufficient.
根据三元组理论，将两层分别表示为G＝(E,R,S)的形式，其中E表示知识图谱中的节点、R表示知识图谱的关系、S表示知识图谱中的三元组。According to the triple theory, the two layers are represented in the form of G=(E, R, S), where E represents the nodes in the knowledge graph, R represents the relationship of the knowledge graph, and S represents the triples in the knowledge graph.
设备层划分为两个子层，电厂中典型设备整体的名称为顶层节点，典型设备的具体部件为底层节点，典型设备整体的名称与其相应的具体部件以“包含”为关系进行连接。例如，典型设备“磨煤机”与其具体部件“磨碗”之间采用“包含”为关系进行连接。除了源于领域词典和专家提供外，为了防止某些不频繁术语被遗漏，引入TF-IDF算法提取关键词，并在结果中去除停用词，如下：The equipment layer is divided into two sub-layers. The name of the whole typical equipment in the power plant is the top node, and the specific components of the typical equipment are the bottom nodes. For example, a typical device "coal mill" and its specific component "grinding bowl" are connected with a "contains" relationship. In addition to being derived from domain dictionaries and provided by experts, in order to prevent some infrequent terms from being missed, the TF-IDF algorithm is introduced to extract keywords, and stop words are removed from the results, as follows:
TF-IDF＝TF*IDF (37)TF-IDF=TF*IDF (37)
其中TF为词频，IDF为逆文档频率。语料库是指所收集到的故障描述文档的集合。取TF-IDF排名靠前的词也作为设备层候选节点。where TF is the term frequency and IDF is the inverse document frequency. Corpus refers to a collection of collected fault description documents. The top words in TF-IDF are also used as device layer candidate nodes.
故障层以故障描述文本、故障诊断文本、处理意见文本为节点，对应的故障描述文本与故障诊断文本以“诊断”为关系进行连接；对应的故障描述文本与处理意见文本以“处理”为关系进行连接，形成故障描述层。而后将故障层中故障描述文本节点与其中涉及到的具体部件节点相连接，从而形成“两层——三要素”的塔形故障诊断知识图谱架构。使用Neo4j数据库存储知识图谱。The fault layer takes fault description text, fault diagnosis text, and processing opinion text as nodes, and the corresponding fault description text and fault diagnosis text are connected by the relationship of "diagnosis"; the corresponding fault description text and processing opinion text are related to "processing" Make connections to form a fault description layer. Then, the fault description text nodes in the fault layer are connected with the specific component nodes involved, so as to form a tower-shaped fault diagnosis knowledge graph architecture of "two layers - three elements". Use Neo4j database to store knowledge graph.
步骤4)对故障描述文本、故障诊断文本、处理意见文本进行进一步处理，构建用于双向GRU网络提取文本特征的训练集。首先对文本进行分句，将文本从逗号、句号、冒号处分割,形成子句。对每一个子句，使用结巴(jieba)分词工具进行分词。然后进行BPE处理，得到BPE处理后的子句和词典，以每一个处理后的子句作为单个训练样本。Step 4) Further processing the fault description text, fault diagnosis text and processing opinion text to construct a training set for extracting text features by the bidirectional GRU network. First, the text is segmented, and the text is split from commas, periods, and colons to form clauses. For each clause, use the jieba word segmentation tool for word segmentation. Then perform BPE processing to obtain BPE-processed clauses and dictionaries, and use each processed clause as a single training sample.
步骤5)构建使用GRU(门控循环单元)的双向循环神经网络(Bi-directionRecurrent Neural Network),如附图6，附图7所示，采用编码器——解码器框架和注意力机制。通过每一个子句同时作为源语句和目标语句，以自监督的方式获取每一个子句的特征向量。其中GRU具体包括：Step 5) Construct a Bi-direction Recurrent Neural Network (Bi-direction Recurrent Neural Network) using GRU (Gated Recurrent Unit), as shown in Figure 6 and Figure 7, using an encoder-decoder framework and an attention mechanism. The feature vector of each clause is obtained in a self-supervised manner by using each clause as both a source sentence and a target sentence. The GRU specifically includes:
首先将每个用于编码器输入端的子句样本中的每个词采用独热编码转换为一维向量，每一个向量的长度和BPE处理得到的词典大小相同，其中只有该词对应的位置为1，其余位置为0。然后使用embedding层进行降维映射，映射矩阵大小为K*V，其中K为人为设定的词向量维度，V为词典大小。将映射矩阵与子句独热编码形成的矩阵相乘从而对词向量降维，得到向量x＝{x1,x2,…,xt,…,xT}，T表示子句中对应的词数量。在模型训练过程中，对于每个字句样本对应的期望输出语句，采用相同的方法进行处理，得到u＝{u1,u2,…,ut,…,uT}。First, each word in each clause sample used for the input of the encoder is converted into a one-dimensional vector by one-hot encoding. The length of each vector is the same as the size of the dictionary processed by BPE, and only the corresponding position of the word is 1, the rest of the positions are 0. Then use the embedding layer for dimensionality reduction mapping, and the size of the mapping matrix is K*V, where K is the artificially set word vector dimension, and V is the dictionary size. Multiply the mapping matrix and the matrix formed by the one-hot encoding of the clause to reduce the dimension of the word vector, and obtain the vector x={x1 ,x2 ,...,xt ,...,xT }, T represents the corresponding value in the clause number of words. In the model training process, the expected output sentence corresponding to each sentence sample is processed in the same way to obtain u={u1 ,u2 ,...,ut ,...,uT }.
通过上一刻的节点状态ht-1和当前节点的输入xt来获取重置门r和更新门z：The reset gate r and update gate z are obtained from the node state ht-1 at the last moment and the input xt of the current node:
r＝sigmoid(wr*[ht-1,xt]) (38)r=sigmoid(wr *[ht-1 ,xt ]) (38)
z＝sigmoid(wz*[ht-1,xt]) (39)z=sigmoid(wz *[ht-1 ,xt ]) (39)
其中，x＝{x1,x2,…,xt,…,xT}为每一个子句样本经过上述独热编码和映射过程得到的向量组，t即当前时刻，表示当前所输入的词在子句中所处的位置。σ为sigmoid函数，将数值映射到0-1范围内。wr,wz为可学习的参数。Among them, x={x1 ,x2 ,...,xt ,...,xT } is the vector group obtained by each clause sample through the above one-hot encoding and mapping process, t is the current moment, representing the current input The position of the word in the clause. σ is a sigmoid function that maps values to a range of 0-1. wr , wz are learnable parameters.
接着获得当前时刻的状态与输出Then get the state and output of the current moment
其中w为可学习的参数。ht为当前时刻的输出。where w is a learnable parameter. ht is the output at the current moment.
对于编码器部分，双向循环神经网络分别在时间维度上以前向和后向处理输入序列，并将每个时间步的输出拼接进行输出。For the encoder part, a bidirectional recurrent neural network processes the input sequence forward and backward separately in the time dimension, and concatenates the outputs at each time step for output.
其中xt为独热编码后降维得到的词向量，
对于解码器部分，应用注意力机制，每一个时刻，根据第t个词的上下文向量ct,目标序列第t个词向量ut和t时刻隐藏状态zt，计算出下一个隐层状态zt+1：For the decoder part, the attention mechanism is applied. At each moment, the next hidden layer state z is calculated according to the context vector ct of the t-th word, the t-th word vector ut of the target sequence and the hidden state zt at time t.t+1 :
其中权重aij表示目标词i对源词j的注意力大小，align为对齐模型，用于衡量目标词i对源词j的匹配程度。The weight aij represents the attention of the target word i to the source word j, and align is the alignment model, which is used to measure the matching degree of the target word i to the source word j.
将zt+1通过softmax归一化，得到目标序列第t+1个词的概率分布pt+1,使用交叉熵函数得到t+1的代价，对所有时刻取平均得到总的损失函数：Normalize zt+1 by softmax to get the probability distribution pt+1 of the t+1th word in the target sequence, use the cross entropy function to get the cost of t+1, and average all the moments to get the total loss function:
pt+1＝softmax(wszt+1+b) (49)pt+1 =softmax(ws zt+1 +b) (49)
其中avg为求平均函数，crossentropy为交叉熵函数。ws,b为需要学习的参数。对网络进行训练，冻结训练好的网络参数、并存储获得的特征向量。实验中构建了包含43756个样本的训练集，最终loss下降到0.25，说明编码器提取的特征向量是有效的，因为解码器能够准确地使用该特征向量重构出原有输入。Where avg is the averaging function, and crossentropy is the cross entropy function. ws , b are parameters that need to be learned. Train the network, freeze the trained network parameters, and store the obtained feature vectors. In the experiment, a training set containing 43,756 samples was constructed, and the final loss dropped to 0.25, indicating that the feature vector extracted by the encoder is effective, because the decoder can accurately use the feature vector to reconstruct the original input.
步骤6)对于知识图谱中的故障描述节点，使用之前存储的特征向量，使用余弦相似度进行比对，将相似度高，且关键术语基本一致的故障描述节点进行合并，消除已有知识库的冗余。Step 6) For the fault description nodes in the knowledge graph, use the previously stored feature vector, use the cosine similarity to compare, and merge the fault description nodes with high similarity and basically the same key terms to eliminate the existing knowledge base. redundancy.
在推理检索的过程中，如果输入是文本形式的故障描述，使用步骤5)中训练好的编码器对输入文本进行编码，编码中既包含关键词的信息，也包含汉语语序信息，就可以使用余弦相似度在知识图谱中检索最相似的文本描述，得到其故障诊断、处理意见、故障现场图。并将得到的故障诊断再次作为故障描述进行检索，直到检索不出高相似度结果为止，如附图8所示，将多次检索到的结果以树的形式输出，实现多层深度追因。通过关联的设备层将涉及到的关键设备部件以标签的形式返回，便于使用者进行梳理总结。In the process of inference retrieval, if the input is a fault description in the form of text, use the encoder trained in step 5) to encode the input text. The encoding contains both keyword information and Chinese word order information, you can use Cosine similarity retrieves the most similar text descriptions in the knowledge graph, and obtains its fault diagnosis, handling advice, and fault scene map. The obtained fault diagnosis is retrieved as a fault description again until no high similarity results can be retrieved. As shown in FIG. 8 , the results retrieved for multiple times are output in the form of a tree to realize multi-layer deep tracing. The key equipment components involved are returned in the form of labels through the associated equipment layer, which is convenient for users to sort out and summarize.
如果输入的是一段数据，使用步骤2)中由3-Sigma原则处理正常期间的数据生成的阈值进行检测，从而转换成故障描述文本。If the input is a piece of data, use the threshold value generated by the 3-Sigma principle to process the data during the normal period in step 2) for detection, thereby converting it into fault description text.
作为优选方案，还可以设计用于工业落地应用的GUI界面。如附图9所示，功能包括磨煤机故障诊断、查询历史、近期检修情况。可通过故障描述或传感器数据进行故障诊断，返回故障诊断结果、检修建议、检修故障图。As a preferred solution, a GUI interface for industrial landing applications can also be designed. As shown in Figure 9, the functions include coal mill fault diagnosis, query history, and recent maintenance. Fault diagnosis can be carried out through fault description or sensor data, and the fault diagnosis results, maintenance suggestions, and maintenance fault diagrams can be returned.
作为另一优选方案，对于新增故障诊断知识、采用步骤2)、3)、4)的过程进行处理后，如果其中没有出现词典未收录的关键术语，则使用步骤5)中存储的参数得到编码结果，使用解码器结合柱搜索算法得到解码结果，在解码的过程中，不断通过pt+1采样ut+1，使用广度优先策略建立搜索树，在树的每一层，按照生成词的log概率之和为启发代价对节点进行排序，然后仅留下预先确定的个数的节点，直到获得句子结束标记或超过最大生成长度为止。解码结果和原子句进行比对，如一致则通过检验，将新知识并入原有知识图谱，实现知识图谱的更新。As another preferred solution, after processing the process of adding fault diagnosis knowledge and adopting steps 2), 3) and 4), if there is no key term not included in the dictionary, use the parameters stored in step 5) to obtain The coding result is obtained by using the decoder combined with the column search algorithm to obtain the decoding result. During the decoding process, ut+1 is continuously sampled through
Claims (8)
Translated fromChinese
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202011131638.0ACN112307218B (en) | 2020-10-21 | 2020-10-21 | Construction method of fault diagnosis knowledge base for typical equipment of intelligent power plant based on knowledge graph |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202011131638.0ACN112307218B (en) | 2020-10-21 | 2020-10-21 | Construction method of fault diagnosis knowledge base for typical equipment of intelligent power plant based on knowledge graph |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN112307218A CN112307218A (en) | 2021-02-02 |
| CN112307218Btrue CN112307218B (en) | 2022-08-05 |
Family
ID=74326825
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202011131638.0AActiveCN112307218B (en) | 2020-10-21 | 2020-10-21 | Construction method of fault diagnosis knowledge base for typical equipment of intelligent power plant based on knowledge graph |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN112307218B (en) |
Families Citing this family (33)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US11960993B2 (en)* | 2020-11-10 | 2024-04-16 | Equifax Inc. | Machine-learning techniques involving monotonic recurrent neural networks |
| CN113064999B (en)* | 2021-03-19 | 2023-12-15 | 南方电网调峰调频发电有限公司信息通信分公司 | Knowledge graph construction algorithm, system, equipment and medium based on IT equipment operation and maintenance |
| CN113094512B (en)* | 2021-04-08 | 2024-05-24 | 达观数据有限公司 | Fault analysis system and method in industrial production and manufacturing |
| CN113205485B (en)* | 2021-04-08 | 2024-08-06 | 惠州市惠发电梯工程有限公司 | Elevator fault self-diagnosis image display method, device, computer and storage medium |
| CN113112164A (en)* | 2021-04-19 | 2021-07-13 | 特变电工股份有限公司新疆变压器厂 | Transformer fault diagnosis method and device based on knowledge graph and electronic equipment |
| CN113254468B (en)* | 2021-04-20 | 2023-03-31 | 西安交通大学 | Equipment fault query and reasoning method |
| CN113283027B (en)* | 2021-05-20 | 2024-04-02 | 南京航空航天大学 | A mechanical fault diagnosis method based on knowledge graph and graph neural network |
| CN113359664B (en)* | 2021-05-31 | 2023-07-07 | 海南文鳐科技有限公司 | Fault diagnosis and maintenance system, method, equipment and storage medium |
| CN113240098B (en)* | 2021-06-16 | 2022-05-17 | 湖北工业大学 | Fault prediction method and device based on hybrid gated neural network and storage medium |
| CN113312411B (en)* | 2021-06-22 | 2023-02-10 | 中国电子科技集团公司第十四研究所 | A Method for Diagnosing Causes of Equipment Faults Based on Knowledge Graph and Rule Constraints |
| CN113569026B (en)* | 2021-07-26 | 2024-12-27 | 卡奥斯工业智能研究院(青岛)有限公司 | Equipment fault question and answer method, device, computer equipment and medium |
| CN114218932B (en)* | 2021-11-26 | 2024-02-20 | 中国航空综合技术研究所 | Aviation fault text abstract generation method and device based on fault causal map |
| CN113849663B (en)* | 2021-11-29 | 2022-04-01 | 湖南乐泊科技有限公司 | Fault monitoring method, system and computer medium for camera collaborative knowledge map |
| WO2023103027A1 (en)* | 2021-12-10 | 2023-06-15 | 北京腾锐视讯科技有限公司 | Fault arc detection apparatus |
| CN114970508B (en)* | 2022-05-17 | 2025-02-25 | 国网浙江省电力有限公司电力科学研究院 | Electric power text knowledge discovery method and equipment based on multi-source data fusion |
| CN114724639B (en)* | 2022-06-10 | 2022-09-16 | 国家超级计算天津中心 | Preprocessing acceleration method, device, equipment and storage medium |
| CN115270881A (en)* | 2022-08-01 | 2022-11-01 | 国网宁夏电力有限公司固原供电公司 | A line fault identification method based on multimodal information fusion knowledge graph |
| CN115544265A (en)* | 2022-09-13 | 2022-12-30 | 南京航空航天大学 | Bearing fault diagnosis method based on bearing fault knowledge graph |
| CN115638875B (en)* | 2022-11-14 | 2023-08-18 | 国家电投集团河南电力有限公司技术信息中心 | Power plant equipment fault diagnosis method and system based on map analysis |
| CN115640380A (en)* | 2022-12-14 | 2023-01-24 | 北京航空航天大学 | Information Extraction Method of Diagnosability Hierarchical Elements for Fault Diagnosis Algorithm Recommendation |
| CN115630143B (en)* | 2022-12-21 | 2023-04-07 | 中科航迈数控软件(深圳)有限公司 | Recommendation method and device for fault handling scheme, terminal equipment and storage medium |
| CN116360388B (en)* | 2023-01-18 | 2023-09-08 | 北京控制工程研究所 | Reasoning method and device of performance-fault relation map based on graph neural network |
| CN116882966B (en)* | 2023-06-27 | 2024-04-19 | 广东慧云科技股份有限公司 | A method and device for judging faults of inspection results of operation and maintenance equipment |
| CN117216629A (en)* | 2023-08-24 | 2023-12-12 | 上海发电设备成套设计研究院有限责任公司 | Fault diagnosis method, device, equipment and storage medium |
| CN117149498B (en)* | 2023-10-27 | 2024-03-01 | 华能信息技术有限公司 | Power plant fault diagnosis method and system |
| CN117443547A (en)* | 2023-11-15 | 2024-01-26 | 华能国际电力股份有限公司上海石洞口第一电厂 | A coal mill fault early warning method based on neural network |
| CN117271700B (en)* | 2023-11-23 | 2024-02-06 | 武汉蓝海科创技术有限公司 | Construction system of equipment use and maintenance knowledge base integrating intelligent learning function |
| CN117576710B (en)* | 2024-01-15 | 2024-05-28 | 西湖大学 | Method and device for generating natural language text based on graph for big data analysis |
| CN117666546B (en)* | 2024-01-31 | 2024-05-03 | 中核武汉核电运行技术股份有限公司 | Distributed control system fault diagnosis method and device |
| CN118535947A (en)* | 2024-04-23 | 2024-08-23 | 中国民用航空飞行学院 | Aircraft equipment fault diagnosis method based on KG-BN |
| CN118761664B (en)* | 2024-06-18 | 2025-02-11 | 江门市大晖科技有限公司 | A quality inspection and traceability method and device based on food mixer manufacturing |
| CN119149287A (en)* | 2024-11-18 | 2024-12-17 | 苏州阿普奇物联网科技有限公司 | Method for on-site operation and maintenance investigation and processing of industrial personal computer faults and electronic equipment |
| CN119484998A (en)* | 2025-01-09 | 2025-02-18 | 深圳市普雷德科技有限公司 | Low-light-level night vision imaging device and method |
Citations (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN108509483A (en)* | 2018-01-31 | 2018-09-07 | 北京化工大学 | The mechanical fault diagnosis construction of knowledge base method of knowledge based collection of illustrative plates |
| CN109284396A (en)* | 2018-09-27 | 2019-01-29 | 北京大学深圳研究生院 | Medical knowledge graph construction method, device, server and storage medium |
| CN109669994A (en)* | 2018-12-21 | 2019-04-23 | 吉林大学 | A kind of construction method and system of health knowledge map |
| CN110781671A (en)* | 2019-10-29 | 2020-02-11 | 西安科技大学 | Knowledge mining method for intelligent IETM fault maintenance record text |
| CN111598252A (en)* | 2020-04-30 | 2020-08-28 | 西安理工大学 | A problem-solving method for college computer basic knowledge based on deep learning |
Family Cites Families (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN113646711B (en)* | 2019-03-07 | 2024-09-24 | 霍尼韦尔国际公司 | System and method for establishing relationships between building automation system components |
| CN111737496A (en)* | 2020-06-29 | 2020-10-02 | 东北电力大学 | A method for constructing fault knowledge graph of power equipment |
- 2020
- 2020-10-21CNCN202011131638.0Apatent/CN112307218B/enactiveActive
Patent Citations (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN108509483A (en)* | 2018-01-31 | 2018-09-07 | 北京化工大学 | The mechanical fault diagnosis construction of knowledge base method of knowledge based collection of illustrative plates |
| CN109284396A (en)* | 2018-09-27 | 2019-01-29 | 北京大学深圳研究生院 | Medical knowledge graph construction method, device, server and storage medium |
| CN109669994A (en)* | 2018-12-21 | 2019-04-23 | 吉林大学 | A kind of construction method and system of health knowledge map |
| CN110781671A (en)* | 2019-10-29 | 2020-02-11 | 西安科技大学 | Knowledge mining method for intelligent IETM fault maintenance record text |
| CN111598252A (en)* | 2020-04-30 | 2020-08-28 | 西安理工大学 | A problem-solving method for college computer basic knowledge based on deep learning |
Non-Patent Citations (1)
| Title |
|---|
| "面向电网调度故障处理的知识图谱框架与";乔骥等;《中国电机工程学报 》;20200920;5837-5848* |
Also Published As
| Publication number | Publication date |
|---|---|
| CN112307218A (en) | 2021-02-02 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN112307218B (en) | Construction method of fault diagnosis knowledge base for typical equipment of intelligent power plant based on knowledge graph | |
| CN114168745B (en) | Knowledge graph construction method for production process of ethylene oxide derivative | |
| CN111177394B (en) | A classification method of knowledge graph relational data based on syntactic attention neural network | |
| CN110287323B (en) | Target-oriented emotion classification method | |
| CN115238029A (en) | Construction method and device of power failure knowledge graph | |
| CN116644192B (en) | Knowledge graph construction method based on aircraft component reliability | |
| CN114925157B (en) | Nuclear power station maintenance experience text matching method based on pre-training model | |
| CN113553440A (en) | A Medical Entity Relationship Extraction Method Based on Hierarchical Reasoning | |
| CN114970508A (en) | Method and device for power text knowledge discovery based on data multi-source fusion | |
| CN114897167B (en) | Method and device for constructing knowledge graph in biological field | |
| CN118820778A (en) | A DCS intelligent decision-making method and system integrating large language model and knowledge graph | |
| CN109241199B (en) | Financial knowledge graph discovery method | |
| CN110765277A (en) | Online equipment fault diagnosis platform of mobile terminal based on knowledge graph | |
| CN119336831B (en) | Intelligent electric power data searching method based on resource map | |
| CN116401624B (en) | Non-autoregressive mathematical problem solver based on multi-tree structure | |
| CN119127970B (en) | Standard substance standard substance retrieval and ordering method and system based on search engine | |
| CN111968700A (en) | Method and system for extracting rice phenomics knowledge map relation based on BERT | |
| CN114912435A (en) | Method and device for power text knowledge discovery based on frequent itemset algorithm | |
| CN116010619A (en) | Knowledge extraction method in complex equipment knowledge graph construction process | |
| CN112559741A (en) | Nuclear power equipment defect recording text classification method, system, medium and electronic equipment | |
| CN118467229A (en) | A database alarm intelligent diagnosis method and system based on multimodal knowledge graph fusion and small sample learning | |
| CN118916497A (en) | Overseeing-based convolution unsupervised cross-modal retrieval method, system, medium and device | |
| CN115129892A (en) | Method and device for constructing knowledge graph of fault handling in distribution network | |
| CN119961388A (en) | Knowledge question answering optimization system in the power field based on large model retrieval enhancement generation and instruction supervision fine-tuning | |
| CN119005321A (en) | Electromechanical equipment predictive maintenance method integrating time sequence knowledge graphs |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| GR01 | Patent grant | ||
| GR01 | Patent grant |