A remote fault diagnosis method and system for train traction drive systems

Abstract

This invention relates to the field of fault diagnosis technology and discloses a remote fault diagnosis method and system for train traction drive systems. The method involves constructing a fault lexicon, segmenting fault corpora based on the lexicon, constructing a fault knowledge graph ontology, extracting information from the segmented fault corpora using a preset method to obtain entity and relation information, constructing a knowledge graph of train traction drive system faults, connecting the graph database to a front-end interface based on the constructed knowledge graph, building a remote auxiliary decision-making system for fault diagnosis during train operation, and performing fault diagnosis based on this system. This addresses the problems of existing remote fault diagnosis methods, such as high reliance on expert experience, low utilization of unstructured knowledge, resulting in low diagnostic efficiency and time-consuming maintenance decisions.

Description

Technical Field

[0001] This invention relates to the field of fault diagnosis technology, and in particular to a remote fault diagnosis method and system for train traction transmission systems. Background Technology

[0002] With the continuous advancement of technology and the ever-increasing demand for transportation, rail trains are playing an increasingly important role in modern transportation networks. The traction drive system, providing driving force, is one of the core components of a rail train and is crucial for ensuring its safe and stable operation. Due to the complexity of the system, a serious malfunction can lead to significant property damage and personal injury. Currently, online fault diagnosis methods for systems have achieved many successes, but poor interpretability of diagnostic results and the inability to utilize unstructured knowledge limit their application in practical engineering scenarios. Therefore, it is necessary to combine remote fault diagnosis methods to achieve efficient and rapid diagnosis of system faults. Traditional remote fault diagnosis methods first integrate data and textual information describing the fault, and then conduct multi-stage consultations. Depending on the diagnostic difficulty, this is typically divided into two stages: the first stage involves on-site engineers diagnosing the fault based on their experience, fault logs, and maintenance manuals; if the fault cannot be resolved, the second stage involves multiple technical experts consulting based on their experience and knowledge, ultimately providing a diagnostic conclusion, making maintenance decisions for the problematic component, and resolving the fault. Compared to online diagnostic methods, this method is more practical and interpretable, and is widely used in engineering. However, fault logs and maintenance manuals primarily store information in text and tabular formats. This flat storage method lacks intuitiveness and makes searching difficult, hindering the efficient utilization of unstructured knowledge. Furthermore, expert consultations heavily rely on human experience, which is often difficult to quantify objectively and is easily influenced by factors such as the expert's age and mood, leading to lengthy and inefficient diagnoses. In addition, as the system continues to operate, new operation and maintenance records are constantly generated, causing an explosive growth in data and knowledge, increasing the difficulty of updating expert knowledge. Therefore, existing remote fault diagnosis methods have many limitations, insufficient utilization of unstructured knowledge, and low efficiency in fault diagnosis. Summary of the Invention

[0003] This invention provides a remote fault diagnosis method and system for train traction drive systems to solve the problems of high dependence on expert experience and low utilization of unstructured knowledge in existing remote fault diagnosis methods, resulting in low diagnosis efficiency and long operation and maintenance decision-making time.

[0004] To achieve the above objectives, the present invention employs the following technical solution:

[0005] In a first aspect, this application provides a remote fault diagnosis method for a train traction drive system, comprising:

[0006] S1: Collect fault data of the train traction transmission system, and clean and remove redundancy from the fault data;

[0007] S2: Construct a fault lexicon, and perform word segmentation on the fault corpus based on the fault lexicon;

[0008] S3: Construct the ontology of the fault knowledge graph, wherein the ontology of the knowledge graph includes the types of fault entities and the types of relationships between entities;

[0009] S4: Extract information from the faulty corpus after word segmentation according to the preset method to obtain entity information and relation information;

[0010] S5: Construct a knowledge graph of train traction drive system faults based on the entity information and relationship information;

[0011] S6: Based on the constructed knowledge graph, connect the graph database with the front-end interface to build a remote auxiliary decision-making system for fault diagnosis during train journey, and perform fault diagnosis based on the remote auxiliary decision-making system for fault diagnosis.

[0012] Secondly, this application provides a remote fault diagnosis system for a train traction transmission system, including a processor and a memory;

[0013] Memory, used to store computer programs;

[0014] A processor, when executing a program stored in memory, implements the steps of the method.

[0015] Beneficial effects:

[0016] This invention provides a remote fault diagnosis method for train traction drive systems. The method involves constructing a fault lexicon and segmenting fault corpora based on this lexicon. It also involves building a fault knowledge graph ontology, which includes the types of fault entities and the types of relationships between entities. Information is extracted from the segmented fault corpora using a preset method to obtain entity and relationship information. Based on this entity and relationship information, a knowledge graph of train traction drive system faults is constructed. Finally, the knowledge graph is connected to a front-end interface to build a remote auxiliary decision-making system for fault diagnosis during train operation. Fault diagnosis is then performed based on this system. This approach addresses the problems of existing remote fault diagnosis methods, such as high reliance on expert experience and low utilization of unstructured knowledge, leading to low diagnostic efficiency and time-consuming maintenance decisions. It helps improve the efficiency of on-site fault diagnosis and handling, providing a new and feasible solution for maintaining the safe and stable operation of rail trains. Attached Figure Description

[0017] Figure 1This is a flowchart of a remote fault diagnosis method for a train traction drive system based on a knowledge graph, according to a preferred embodiment of the present invention.

[0018] Figure 2 This is a schematic diagram of the dual BiLSTM information extraction structure of a preferred embodiment of the present invention;

[0019] Figure 3 This is a schematic diagram of the Neo4j knowledge graph visualization according to a preferred embodiment of the present invention. Detailed Implementation

[0020] The technical solution of the present invention will be clearly and completely described below. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0021] Unless otherwise defined, the technical or scientific terms used in this invention shall have the ordinary meaning understood by one of ordinary skill in the art to which this invention pertains. The terms "first," "second," and similar terms used in this invention do not indicate any order, quantity, or importance, but are merely used to distinguish different components. Similarly, the terms "an" or "a" and similar terms do not indicate a quantity limitation, but rather indicate the presence of at least one. The terms "connected" or "linked" and similar terms are not limited to physical or mechanical connections, but can include electrical connections, whether direct or indirect. "Up," "down," "left," "right," etc., are used only to indicate relative positional relationships; when the absolute position of the described object changes, the relative positional relationship also changes accordingly.

[0022] It should be understood that the remote fault diagnosis method for train traction drive systems of this application can be applied to scenarios involving rail train traction drive systems.

[0023] Please see Figure 1 This application provides a remote fault diagnosis method for train traction drive systems, comprising:

[0024] S1: Collect fault data of the train traction transmission system, and clean and remove redundancy from the fault data;

[0025] S2: Construct a fault lexicon, and perform word segmentation on the fault corpus based on the fault lexicon;

[0026] S3: Construct the ontology of the fault knowledge graph, wherein the ontology of the knowledge graph includes the types of fault entities and the types of relationships between entities;

[0027] S4: Extract information from the faulty corpus after word segmentation according to the preset method to obtain entity information and relation information;

[0028] S5: Construct a knowledge graph of train traction drive system faults based on the entity information and relationship information;

[0029] S6: Based on the constructed knowledge graph, connect the graph database with the front-end interface to build a remote auxiliary decision-making system for fault diagnosis during train journey, and perform fault diagnosis based on the remote auxiliary decision-making system for fault diagnosis.

[0030] The corpus of fault data in train traction transmission systems was collected from multiple sources, including fault logs and maintenance manuals of rail trains, academic papers published in professional journals in the field, and relevant professional books.

[0031] After obtaining the faulty corpus, the corpus is cleaned and redundancy is removed.

[0032] The corpus cleaning process mainly involves removing meaningless statements and stop words, and rearranging statements with mixed information. For example, the corpus sentence "When faults such as grounding of the traction motor input connection line, water ingress into the motor electrical connection plug, phase-to-phase resistance imbalance, and poor insulation of the stator winding to ground occur, a grounding fault of the traction converter will occur." becomes "The main causes of traction converter grounding faults include: grounding of the traction motor input connection line, water ingress into the electrical connection plug, phase-to-phase resistance imbalance, and poor insulation of the stator winding to ground." after corpus cleaning. The overall corpus structure becomes a general-to-specific structure.

[0033] Redundancy removal from corpora includes merging synonymous corpora and standardizing and unifying the expression of synonyms, deleting corpora containing the same information, and unifying the expression of multiple words that express the same objective object.

[0034] The aforementioned remote fault diagnosis method for train traction drive systems addresses the problems of existing remote fault diagnosis methods, such as high reliance on expert experience and low utilization of unstructured knowledge, leading to low diagnostic efficiency and time-consuming operation and maintenance decisions. It helps improve the efficiency of on-site fault diagnosis and handling, providing a new and feasible solution for maintaining the safe and stable operation of rail trains.

[0035] Further, S2 includes:

[0036] A candidate vocabulary is constructed for subsequent filtering. An N-Gram model is used to generate the vocabulary. This model divides the corpus into N-byte sequences by performing a sliding window operation on the corpus, with each sequence representing an N-gram combination. After removing N-gram combinations with frequencies below a threshold and common words, the remaining N-gram combinations are the candidate words.

[0037] The initial candidate words generated using the N-Gram model contain meaningless stop words. Subsequently, a predefined part-of-speech (POS) combination template needs to be constructed to filter them. POS typically include five types: noun (n), adjective (a), verb (v), preposition (p), and adverb (ad). Analysis of domain-specific vocabulary reveals that p and ad are uncommon, and their combinations usually end with n or v. Based on this, k predefined POS combinations are defined. After constructing the POS combination template, POS matching is performed on the candidate words. Words that pass the match proceed to the next stage of filtering. This stage calculates the frequency of candidate words in the corpus, i.e., word frequency, and its calculation formula is defined as follows:

[0038]

[0039] Where TF(w) represents the word frequency of word W, C(w) represents the number of times word w appears in the corpus, and N represents the total number of words in the corpus. The result calculated for candidate words is compared with a preset threshold T1. If the result is greater than the threshold, the next step of verification is performed. When using word frequency to initially screen domain vocabulary, there is no fixed standard for defining the threshold T1 for high-frequency words in the domain. The T1 used here is the threshold defined in the TF-H method, which can reduce redundant items while screening high-frequency words.

[0040] After the initial screening, the metrics of Mutual Information (MI) and Information Entropy (IE) are introduced to consider the internal cohesion and boundary clarity of words, respectively, for the final screening.

[0041] The definition of mutual information is:

[0042]

[0043] Where P(w,z) is the probability that word w and word z appear at the same time, and P(w) and P(z) are the probabilities that word w and word z appear separately. A high mutual information value indicates that there is a strong correlation between the two words, and the probability of them forming a new word when combined is high.

[0044] The left and right entropy are defined as follows:

[0045]

[0046] in, Indicate the left / right context of the word w, Given the word w, The probability of its occurrence, E L / R (w) represents the left and right entropy of word w. The larger the left and right entropy, the greater the uncertainty of the surrounding information of the word, the clearer the boundary of the word, and the greater the probability that it is an independent word.

[0047] After calculation, when the mutual information and left and right entropies of a candidate word both exceed a preset threshold T2, it can be finally determined as a true word in the traction drive system fault domain and added to the domain terminology. The value of T2 is determined by the user based on experience or trial and error.

[0048] Subsequently, by integrating all selected candidate words based on their word frequencies, a complete domain lexicon is obtained. The Jieba tool is used for word segmentation. This tool employs prefix-based word graph scanning, which can construct a directed acyclic graph (DAG) of Chinese characters in a sentence. After inputting the self-constructed domain lexicon, this method uses dynamic programming to find the maximum segmentation combination based on word frequency, and ultimately completes the word segmentation task.

[0049] Considering that most of the expert knowledge in the literature comes from the field work environment and has been tested by actual operation processes, a top-down approach is adopted to construct the knowledge graph. First, the ontology is constructed, entities and relations are defined, and then entities and relations are extracted from the corpus according to the defined categories.

[0050] Specifically, the ontology for constructing the fault knowledge graph in S3 includes:

[0051] Top-down ontology construction specifically includes three steps: determining entity types, determining relation types, and determining the head and tail entity types of the relation types.

[0052] Based on the collected literature data and the needs of practical engineering applications, a total of n entities and m relationships are defined.

[0053] Furthermore, a special case exists in the process of constructing knowledge graph ontology: for the same relation, there are two different combinations of head and tail entities, and for the same combination of head and tail entities, there are two different relations. For example, in the relations "fault-fault representation-phenomenon" and "fault-fault cause-phenomenon," although the head and tail entities are the same, the corresponding relations are different. This reflects that phenomena can describe faults from different perspectives; a phenomenon can be both a manifestation of a fault and a cause of a fault.

[0054] Specifically, information extraction is performed on the fault corpus after word segmentation according to a preset method, including extracting information from the corpus after word segmentation according to the proposed dual BiLSTM method. First, entity information is extracted based on the BiLSTM-CRF model, and then relation information is extracted based on the Att-BiLSTM model.

[0055] This method employs a pipelined architecture for extraction, fully leveraging the superior performance of both models on their respective extraction tasks. Furthermore, to compensate for the lack of interaction, the proposed method utilizes the fact that both models contain BiLSTM layers to introduce feature values ​​that can convey hidden information, thereby improving the correlation between sub-tasks in information extraction.

[0056] The entity extraction part uses a BiLSTM-CRF model. This model consists of three parts: a word vector layer, a BiLSTM network layer, and a CRF layer. Compared to traditional LSTM, which only considers historical information up to the current time step, BiLSTM introduces two LSTM layers, forward and backward, thus considering both the current time step and the context information. Conditional Random Fields (CRF) are undirected graph-based models. Due to their ability to achieve global normalization, they can better handle long-range context information, making them particularly suitable for processing sequence data. The text content of manually searched journal articles is a typical example of sequence data, exhibiting a clear chronological order and long-range dependencies. Combining BiLSTM and CRF models allows for a more comprehensive understanding of the input sequence.

[0057] Overall, the training process of BiLSTM can be divided into two stages: forward propagation and backward propagation, which respectively obtain the forward hidden state sequence. and backward hidden state sequence The final output sequence H is obtained by concatenating the two sequences. t Its expression is,

[0058]

[0059] For model training, the most commonly used cross-entropy loss function is employed.

[0060]

[0061] Where T is the text length, x t Let p(x) be the word embedding vector of the t-th word in the text. This vector can be obtained from the Word2Vec model. t q(x) represents the predicted value for the label category. t Then ) is the corresponding true value.

[0062] The relation extraction part uses the Att-BiLSTM model. Att-BiLSTM adds an attention layer between the output layer and the hidden layer to obtain mutual information between words and adjust the output weights of the hidden layer.

[0063] In the attention mechanism layer, the following definition can be made:

[0064]

[0065] Where, α ij Let e ​​be the attention probability of the j-th word to the i-th word. ij The term e represents the influence of the j-th word on the i-th word. ik The term e represents the influence of the k-th word on the i-th word. ij and e ik It can be calculated through a feedforward neural network, and further,

[0066]

[0067] Among them, c i h is the output word feature vector. j This represents the hidden state information at the j-th position of the BiLSTM layer output. Combining this with the softmax function of the output layer, we can obtain...

[0068]

[0069] Where y is the output probability predicted by the model, V is the weight matrix of the model output layer, and c k This is the word feature vector of the output k-th word.

[0070] By combining the two models and introducing the interaction between entity extraction and relation extraction, the proposed information extraction method based on dual BiLSTM has the following structure: Figure 2 As shown,

[0071] In the entity extraction part, the text is labeled using the BIO annotation mode. This method assigns three different labels to words based on their type: B (Begin, representing the starting word of the named entity), I (Inside, representing the subsequent words of the named entity), and O (Outside, representing words that do not belong to any named entity category obtained from ontology construction).

[0072] The interaction in the relation extraction part is achieved by training the BiLSTM during the entity extraction stage to generate the final long-term memory state C. t As a feature value, it is input into the relation extraction model to generate the initial hidden state h in Att-BiLSTM. rel Its definition is as follows:

[0073] h rel =tanh(W r ·C t +b r (9);

[0074] In the formula, W r It is the weight matrix, b r These are bias terms; both are used together to transform C. t To obtain the initial hidden state. rel The word vectors from the text are used together as initial information for the Att-BiLSTM and applied to subsequent relation extraction tasks.

[0075] Finally, the proposed dual BiLSTM method was used to extract entities and relations in sequence to obtain information related to train traction transmission system faults contained in the corpus.

[0076] Specifically, in step five, based on the extracted entity and relation information, the Neo4j graph database is used to construct a knowledge graph of train traction transmission system faults.

[0077] After extracting entity and relationship information, the Cypher language is used to insert this information into the Neo4j graph database, forming a well-structured knowledge graph. Neo4j can intuitively represent the relationships between entities, forming an intuitive graph, and selecting a portion for visualization, such as... Figure 3 As shown.

[0078] Specifically, in one example, the fault remote auxiliary decision-making system of this application includes: a user login interface and a user interaction interface, wherein the user interaction interface is connected to the Neo4j graph database containing fault information;

[0079] The user login interface is used to verify the user's identity and protect the information security of professional knowledge involved in the knowledge graph.

[0080] The user interface is used to ask questions to the decision support system, input fault phenomena and other fault-related questions, display the answers to the questions, and use the intuitive node connection relationships in the Neo4j graph database as auxiliary information.

[0081] In one example, the user login interface is built using the Flask framework in Python and displayed in a web browser. Its main function is to allow users to log in by entering their username and password in the input boxes on the interface. The login logic in the code first looks up the username entered by the user, then uses the bcrypt library to compare the entered password with the stored hash password. If they match, login is granted.

[0082] User Interface: This interface is built using the Flask framework in Python and displayed in a web browser. The left column of the interface includes an introduction to the traction drive system, common system faults, and the decision support system. On the right side, users can submit questions, and the system connects to the Neo4j graph database to return relevant answers. To ensure accurate answers, the system uses Jieba segmentation and fuzzy matching technologies to accurately determine the user's query intent and extract key information. Answers are generated based on predefined answer templates, ensuring clarity, accuracy, and consistent formatting. Furthermore, for more intuitive feedback, the system provides a visual representation of the knowledge graph showing the connections between relevant nodes and edges. Through this interface, users can interact with the knowledge graph to query information related to train traction drive system faults.

[0083] After entering the login interface, users input their username and password in the login box and click the login button to be redirected to the interactive interface. After reading the relevant introduction on the left, users input their query into the box below "Please Enter" on the right side of the decision support system, and then click the decision button. The lower left corner of the interface displays the text answer from the knowledge graph related to the query, while the lower right corner displays the relevant nodes and edge connections. Users can then use the system's returned answer to make real-world fault decisions and perform maintenance, ultimately resolving the fault.

[0084] This application also provides a remote fault diagnosis system for a train traction drive system, including a processor and a memory; the memory is used to store a computer program; the processor is used to execute the program stored in the memory to implement the steps of the method described above. This remote fault diagnosis system for a train traction drive system can implement various embodiments of the above-described remote fault diagnosis method for a train traction drive system and achieve the same beneficial effects, which will not be elaborated upon here.

[0085] The preferred embodiments of the present invention have been described in detail above. It should be understood that those skilled in the art can make numerous modifications and variations based on the concept of the present invention without creative effort. Therefore, all technical solutions that can be obtained by those skilled in the art based on the concept of the present invention through logical analysis, reasoning, or limited experimentation on the basis of existing technology should be within the scope of protection defined by the claims.

Claims

1. A remote fault diagnosis method for train traction transmission systems, characterized in that, include: S1: Collect fault data of the train traction transmission system, and clean and remove redundancy from the fault data; S2: Construct a fault lexicon, and perform word segmentation on the fault corpus based on the fault lexicon; S3: Construct the ontology of the fault knowledge graph, wherein the ontology of the knowledge graph includes the types of fault entities and the types of relationships between entities; S4: Extract information from the faulty corpus after word segmentation according to the preset method to obtain entity information and relation information; S5: Construct a knowledge graph of train traction drive system faults based on the entity information and relationship information; S6: Based on the constructed knowledge graph, connect the graph database with the front-end interface to build a remote auxiliary decision-making system for fault diagnosis during train journey, and perform fault diagnosis based on the remote auxiliary decision-making system for fault diagnosis. S2 includes: A candidate word list with all existing words is constructed for subsequent screening. The N-Gram model is used to perform a sliding window operation of length N on the content of the corpus by bytes, dividing it into several byte fragment sequences of length N. Each sequence is an N-Gram combination. After removing N-Gram combinations with a frequency below the threshold and common words, the remaining N-Gram combinations are candidate words. Construct predefined part-of-speech combination templates to filter candidate words; After constructing the part-of-speech combination template, the candidate words are matched for part-of-speech, and the matched words are further filtered; and the word frequency of the candidate words in the corpus is calculated. The results of candidate word calculation and the preset threshold If the result is greater than the threshold, proceed to the next verification step. After the initial screening, mutual information and left / right entropy metrics are introduced for final screening: When the mutual information and left / right entropy of candidate words both exceed a preset threshold When it is determined that it is a real word in the field of traction transmission system faults, it is added to the field word library, and the selected candidate words are integrated with its word frequency to obtain a complete field word library; The Jieba tool was used for word segmentation. After inputting the self-built domain dictionary, dynamic programming was used to find the maximum segmentation combination based on word frequency to complete the word segmentation task. The step of extracting information from the segmented fault corpus according to a preset method includes: The BiLSTM-CRF model is used in the entity extraction part; In terms of model training, the cross-entropy loss function is used. The relation extraction part uses the Att-BiLSTM model; The two models are combined, and the interaction between entity extraction and relation extraction is introduced. By training the BiLSTM during the entity extraction stage, the final long-term memory state is generated. As a feature value, it is input into the relation extraction model to generate the initial hidden state in Att-BiLSTM. : （9）； In the formula, It is a weight matrix. These are bias terms; both are used together for transformation. To obtain the initial hidden state, The word vectors along with the text are used as the initial information for Att-BiLSTM.

2. The remote fault diagnosis method for train traction transmission system according to claim 1, characterized in that, The cleaning and redundancy removal process for the faulty corpus includes: Remove meaningless statements and stop words, and rearrange statements with mixed information. Synonymous corpora are merged and their unified expressions are standardized. Corpora containing the same information are deleted, and multiple words expressing the same objective object are unified.

3. The remote fault diagnosis method for train traction transmission systems according to claim 1, characterized in that, The ontology for constructing the fault knowledge graph includes: A top-down approach is adopted to construct the knowledge graph. First, the ontology is constructed, entities and relations are defined, and then entities and relations are extracted from the corpus according to the defined categories. Based on the corpus and the needs of actual engineering applications, we define n types of entities and m types of relations.

4. The remote fault diagnosis method for train traction transmission systems according to claim 1, characterized in that, The knowledge graph of train traction drive system faults constructed based on the entity information and relationship information includes: Using the Cypher language, entity and relation information is inserted into the Neo4j graph database to form a well-structured knowledge graph.

5. The remote fault diagnosis method for train traction transmission system according to claim 1, characterized in that, The aforementioned remote auxiliary decision-making system for fault diagnosis during train operation includes: The Flask framework in Python was used to build the user login interface and user interaction interface.

6. A remote fault diagnosis system for train traction drive systems, characterized in that, Including processor and memory; Memory, used to store computer programs; A processor, when executing a program stored in memory, implements the steps of the method described in any one of claims 1-5.

Images