Fault identification method and device for rail vehicle
Through the combination of text features, word vectors and low-dimensional vector sequences, a fault classifier is built, which solves the problems of low efficiency and low accuracy of track vehicle fault recognition, and realizes accurate and efficient identification of track vehicle faults.
Patent Information
- Application Number
- CN202011580553.0
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2020-12-28
- Publication Date
- 2024-12-03
- Estimated Expiration
- 2040-12-28
AI Technical Summary
In the prior art, track vehicle fault recognition efficiency and low accuracy, especially in large-scale fault data, manual matching has problems such as inefficient efficiency, and keyword matching is not effective in retrieving polysynonyms or non-standard words.
Using a fault classifier based on text features, word vectors and low-dimensional vector sequences, the fault phenomenon description statements are preprocessed and trained through the preset BERT, Bi-LSTM and word2vec models to generate a model for fault recognition, and realize the Accurate and efficient identification of rail vehicle failures.
Through the combination of text features, word vectors and low-dimensional vector sequences, a fault classifier can be built, which can accurately and efficiently match historical fault tracking table data, improve the accuracy and efficiency of fault recognition, overcome the low efficiency of manual matching and keywords The problem of low matching accuracy.
Smart Images

Figure CN112612890B_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the field of rail transit technology, and in particular to a fault identification method and device for a rail vehicle. Background Art
[0002] During the daily operation and maintenance of the subway, a series of fault text data tables are generated, usually called fault tracking tables. The fault tracking table is filled in by the subway operation and maintenance personnel, and for each fault event, the main records include: description of the phenomenon corresponding to the fault event, keywords of the phenomenon, and solutions to the fault, etc. Figure 1 shown.
[0003] The fault tracking table has important guiding significance for the subsequent maintenance of subway trains. When subway operation and maintenance personnel encounter a fault on site, they can use the fault tracking table to find faults with similar phenomena in the fault tracking table, thereby locating possible fault modules and analyzing the causes of the faults. In the prior art, fault identification is mainly performed in the following ways:
[0004] The first method is based on the exhaustive search by experienced operation and maintenance personnel. This method requires on-site operation and maintenance personnel to compare samples with similar phenomena in the fault tracking table one by one based on their experience, and then locate the possible fault location according to the corresponding fault phenomenon. At present, the fault tracking table has accumulated data on tens of thousands of faults. As the amount of data increases, manual matching has the problem of low efficiency, and there is a possibility of missed detection.
[0005] The second method is the fault retrieval method based on keyword matching. This method searches for faults with similar phenomena in the fault tracking table based on some keywords of the current fault phenomenon, thereby locating faults with similar phenomena. However, this method is not effective for retrieval of possible polysemous words or non-standard words. For example, searching for the keyword "fault current curve" cannot match the "current fault curve" in the fault tracking table. Summary of the invention
[0006] The rail vehicle fault identification method provided by the present invention is used to overcome the defects in the prior art of low efficiency of manual fault identification and insufficient accuracy of fault identification using keyword matching, and can accurately and efficiently realize fault identification of a target rail vehicle.
[0007] The present invention provides a rail vehicle fault identification method, comprising:
[0008] Inputting the first training sample into a preset fault classifier model to obtain parameters of the preset fault classifier model;
[0009] Determining a fault classifier for fault identification according to the preset fault classifier model and the parameters;
[0010] Inputting the description statement of the fault phenomenon to be identified and the historical fault tracking table data into the fault classifier to perform fault identification on the target rail vehicle;
[0011] The first training sample includes: text features of the fault phenomenon description sentence, word vectors of the fault phenomenon description sentence, and low-dimensional vector sequences of the fault phenomenon description sentence;
[0012] The text features include: TF-IDF features and keyword features.
[0013] According to a rail vehicle fault identification method provided by the present invention, the first training sample is obtained in the following manner:
[0014] Preprocessing the historical fault tracking table data to obtain a fault phenomenon description statement corresponding to the historical fault tracking table data;
[0015] Inputting the vector sequence corresponding to the fault phenomenon description statement into a preset bidirectional long short-term memory Bi-LSTM model to obtain a low-dimensional vector sequence corresponding to the vector sequence;
[0016] splicing the low-dimensional vector sequence, the word vector of the fault phenomenon description sentence, and the text feature of the fault phenomenon description sentence to obtain a first training sample;
[0017] Wherein, the fault phenomenon description statement includes: positive sample data and negative sample data;
[0018] The positive sample data is sample data composed of similar fault phenomenon description statements;
[0019] The negative samples are sample data consisting of sentences describing dissimilar fault phenomena.
[0020] According to a rail vehicle fault identification method provided by the present invention, the vector sequence corresponding to the fault phenomenon description statement is obtained in the following manner:
[0021] Inputting the fault phenomenon description sentence into a preset BERT model to obtain a vector sequence including the fault phenomenon description sentence;
[0022] The preset BERT model is obtained by pre-training with the second training sample;
[0023] The second training sample includes: a first text pair corresponding to the training corpus obtained by performing a preset reprocessing on the preprocessed training corpus;
[0024] The training corpus includes: original corpus of the target rail vehicle and historical fault tracking table data of the target rail vehicle.
[0025] According to a rail vehicle fault identification method provided by the present invention, the word vector of the fault phenomenon description sentence is obtained in the following manner:
[0026] Input the fault phenomenon description sentence into a preset word2vec model to obtain a word vector of the fault phenomenon description sentence;
[0027] Wherein, the preset word2vec model is obtained after being trained in advance by a third training sample;
[0028] The third training sample includes: a second text pair corresponding to the training corpus obtained from the preprocessed training corpus.
[0029] According to a rail vehicle fault identification method provided by the present invention, the preprocessing of the historical fault tracking table data to obtain a fault phenomenon description statement corresponding to the historical fault tracking table data includes:
[0030] Performing word segmentation and character segmentation preprocessing on the historical fault tracking table data, and performing cleaning preprocessing on the historical fault tracking table data after word segmentation and character segmentation preprocessing according to a preset cleaning preprocessing rule to obtain the fault phenomenon description statement;
[0031] Wherein, the preprocessing includes the word segmentation and character segmentation preprocessing and the cleaning preprocessing.
[0032] According to a rail vehicle fault identification method provided by the present invention, the pre-setting and re-processing of the training corpus includes:
[0033] Extracting the training corpus to obtain a first preset number of characters that meet a preset extraction rule;
[0034] Performing a first screening on the first preset number of characters to obtain a second preset number of characters, and replacing the second preset number of characters with a preset symbol;
[0035] The first preset number of characters is screened for a second time to obtain a third preset number of characters, and the third preset number of characters are replaced with any characters.
[0036] According to a rail vehicle fault identification method provided by the present invention, the fault identification of the target rail vehicle includes:
[0037] The fault description statement output by the fault classifier is matched with the fault description statement corresponding to the fault case stored in the historical fault tracking table data to determine at least one fault case corresponding to the historical fault tracking table data.
[0038] The present invention also provides a fault identification device for a rail vehicle, comprising:
[0039] A model parameter determination module, a fault classifier determination module and a fault identification module;
[0040] The model parameter determination module is used to input the first training sample into a preset fault classifier model to obtain parameters of the preset fault classifier model;
[0041] The fault classifier determination module is used to determine a fault classifier for fault identification according to the preset fault classifier model and the parameters;
[0042] The fault identification module is used to input the description statement of the fault phenomenon to be identified and the historical fault tracking table data into the fault classifier to perform fault identification on the target rail vehicle;
[0043] The first training sample includes: text features of the fault phenomenon description sentence, word vectors of the fault phenomenon description sentence, and low-dimensional vector sequences of the fault phenomenon description sentence;
[0044] The text features include: TF-IDF features and keyword features.
[0045] The present invention also provides an electronic device, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein when the processor executes the program, the steps of any of the above-mentioned rail vehicle fault identification methods are implemented.
[0046] The present invention also provides a non-transitory computer-readable storage medium having a computer program stored thereon, and when the computer program is executed by a processor, the steps of any of the above-mentioned methods for identifying faults of rail vehicles are implemented.
[0047] The rail vehicle fault identification method and device provided by the present invention obtain a fault classifier for fault identification based on text features, word vectors and low-dimensional vector sequence training of fault phenomenon description sentences, fully consider the word vectors and sentence vectors of the fault phenomenon description sentences, and accurately and efficiently identify the faults of the target rail vehicle by matching the output results of the fault classifier with the historical fault tracking table data of the target rail vehicle. BRIEF DESCRIPTION OF THE DRAWINGS
[0048] In order to more clearly illustrate the technical solutions in the present invention or the prior art, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.
[0049] Figure 1 It is a fault tracking table of rail vehicles provided by the prior art;
[0050] Figure 2 It is a flow chart of a rail vehicle fault identification method provided by the present invention;
[0051] Figure 3 It is a structural schematic diagram of the Bi-LSTM model based on the residual structure provided by the present invention;
[0052] Figure 4 It is a schematic diagram of the structure of the BERT model provided by the present invention;
[0053] Figure 5 It is a structural schematic diagram of a fault identification device for a rail vehicle provided by the present invention;
[0054] Figure 6 It is a structural schematic diagram of the electronic device provided by the present invention. DETAILED DESCRIPTION
[0055] In order to make the purpose, technical solution and advantages of the present invention clearer, the technical solution of the present invention will be clearly and completely described below in conjunction with the drawings of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.
[0056] Figure 2 is a flow chart of a rail vehicle fault identification method provided by the present invention, such as Figure 2 As shown, the method includes:
[0057] S1. Inputting a first training sample into a preset fault classifier model to obtain parameters of the preset fault classifier model;
[0058] S2. Determine a fault classifier for fault identification according to a preset fault classifier model and parameters;
[0059] S3, inputting the description statement of the fault phenomenon to be identified and the data of the historical fault tracking table into the fault classifier to identify the fault of the target rail vehicle;
[0060] The first training sample includes: text features of the fault phenomenon description sentence, word vectors of the fault phenomenon description sentence, and low-dimensional vector sequences of the fault phenomenon description sentence;
[0061] Text features include: TF-IDF features and keyword features.
[0062] It should be noted that the execution subject of the above method may be a computer device.
[0063] Optionally, when the operation and maintenance personnel find a fault, they will describe the fault phenomenon found in the form of natural language to obtain a fault phenomenon description statement, for example, "the turnout is four-open in the route locked state."
[0064] A word is the smallest language unit that can be used independently. Understanding words is of great significance to understanding language. Therefore, in order to better extract feature information in natural language, the present invention needs to obtain the word vector of the fault phenomenon description sentence.
[0065] Term frequency-inverse document frequency (TF-IDF) vectorizes text features to evaluate the importance of a word to a document set or a document in a corpus. The importance of a word increases in direct proportion to the number of times it appears in a document, but decreases in inverse proportion to the frequency of its appearance in the corpus. By counting TF-IDF features, the keyword features of the fault phenomenon description sentence can be obtained.
[0066] The TF-IDF method is used to extract keywords from the fault phenomenon description sentence. Specifically, the IDF value of each word in the fault phenomenon description sentence is calculated according to formula (1), that is, IDF(x):
[0067]
[0068] Wherein, if each fault phenomenon description sentence is regarded as a text, N represents the total number of texts of word x in the fault phenomenon description sentence, then N(x) represents the total number of texts of the fault phenomenon description sentence;
[0069] According to formula (2), the DF value of each word in the fault description phenomenon sentence is obtained, that is, TF(x):
[0070]
[0071] Wherein, if each fault phenomenon description statement is regarded as a text, n(x) represents the total number of occurrences of word x in the fault phenomenon description statement, and num represents the total number of words in the fault phenomenon description statement;
[0072] After obtaining IDF(x) and DF(x), calculate TF-IDF(x) according to formula (3):
[0073] TF-IDF(x)=IDF(x)*TF(x) (3)
[0074] Then, the TF-IDF value of each word in the fault phenomenon description sentence is sorted, and the word with the largest TF-IDF value in each fault phenomenon description sentence is selected as the keyword of the current fault phenomenon description sentence.
[0075] Optionally, the words or phrases in the fault phenomenon description sentence may be judged to determine whether a certain word or phrase in the fault phenomenon description sentence is a keyword or not, and finally obtain the keyword features of the fault phenomenon description sentence.
[0076] By reducing the dimension of the vector sequence corresponding to the fault phenomenon description sentence, a low-dimensional vector sequence corresponding to the vector sequence of the fault phenomenon description sentence is obtained.
[0077] The obtained text features of the fault phenomenon description sentence, the word vectors of the fault phenomenon description sentence and the low-dimensional vector sequence of the fault phenomenon description sentence are combined into a training sample to obtain a first training sample.
[0078] The first training sample is input into a preset fault classifier model for training, parameters of the preset fault classifier model are determined, and a fault classifier for fault identification is obtained according to the preset fault classifier model and the parameters of the preset fault classifier model.
[0079] The fault phenomenon description sentence to be identified is input into the trained fault classifier for processing, so as to obtain the processing result of the fault description sentence to be identified. If there is one fault phenomenon description sentence to be identified and there are 1000 fault phenomenon description sentences in the historical fault tracking table data, the fault classifier needs to perform 1000 calculations to obtain the similarity probability between the fault phenomenon description sentence to be identified and the fault phenomenon description sentences in the historical fault tracking table data, and sort them according to the similarity probability. The fault phenomenon sentence with a large similarity probability is used as the current fault information closest to the target rail vehicle, so as to realize fault identification of the target rail vehicle.
[0080] The rail vehicle fault identification method provided by the present invention obtains a fault classifier for fault identification based on text features, word vectors and low-dimensional vector sequence training of fault phenomenon description sentences, fully considers the word vectors and sentence vectors of the fault phenomenon description sentences, and matches the output results of the fault classifier with the historical fault tracking table data of the target rail vehicle, thereby accurately and efficiently realizing the fault identification of the target rail vehicle.
[0081] Furthermore, in one embodiment, the first training sample is obtained by:
[0082] Preprocess the historical fault tracking table data to obtain the fault phenomenon description statement corresponding to the historical fault tracking table data;
[0083] Input the vector sequence corresponding to the fault phenomenon description statement into the preset bidirectional long short-term memory Bi-LSTM model to obtain a low-dimensional vector sequence corresponding to the vector sequence;
[0084] The low-dimensional vector sequence, the word vector of the fault phenomenon description sentence, and the text feature of the fault phenomenon description sentence are concatenated to obtain a first training sample;
[0085] The fault phenomenon description statement includes: positive sample data and negative sample data;
[0086] Positive sample data is sample data composed of similar fault phenomenon description statements;
[0087] Negative samples are sample data consisting of sentences with dissimilar fault phenomenon descriptions.
[0088] Optionally, the historical fault tracking table data of the target rail vehicle is preprocessed to obtain the fault phenomenon description statement corresponding to the historical fault tracking table data. The purpose of preprocessing the historical fault tracking table data is, on the one hand, to encode the text data for subsequent processing, and on the other hand, to filter out noise and improve the data quality, thereby improving the effect of subsequent model training.
[0089] The preprocessing of the historical fault tracking table data may specifically include word segmentation, character segmentation preprocessing and cleaning preprocessing.
[0090] According to the preset rules, the fault phenomenon description sentences obtained after preprocessing are divided into positive sample data and negative sample data. Specifically:
[0091] The sample data with similar fault phenomenon description statements is taken as positive sample data, and the sample data with dissimilar fault phenomenon description statements is taken as negative sample data. For example, there are 1000 fault phenomenon description statements, and the following operations are performed on the 1000 fault phenomenon description statements: for each fault phenomenon description statement, search for the remaining 999 fault phenomenon description statements that are similar to the fault phenomenon description statement, and combine the search result with the fault phenomenon description statement to form positive sample data; search for fault phenomenon description statements that are dissimilar to the fault phenomenon description statement, and combine the search result with the fault phenomenon description statement to form negative sample data.
[0092] Optionally, professional operation and maintenance personnel can make a similarity judgment on any two fault phenomenon description statements. Professional operation and maintenance personnel can judge whether any two fault phenomenon description statements are fault phenomenon description statements describing similar fault phenomena based on prior experience. If it is determined that they are similar fault phenomena, the two fault phenomenon description statements are marked to form positive sample data; if it is determined that they are not similar fault phenomena, the two fault phenomenon description statements are marked to form negative sample data.
[0093] There is an existing semantic retrieval method based on neural network search. This method regards the retrieval process as a text semantic matching process. When this method is applied, a large number of corpus files are first trained to obtain the vector representation of each word in the corpus. This process is called the pre-training process. After obtaining the word vector, the word vector and the labeled text to be matched will be sent to the downstream neural network model for training, so as to obtain a downstream model that can distinguish whether two sentences match. This module is called the matching model.
[0094] After obtaining the matching model, the matching model can be used to search the text and obtain semantic matching results. Compared with traditional keyword matching methods and manual search methods, this method can cross the semantic gap and retrieve fault descriptions with similar phenomena from a semantic perspective. However, the information of each word is encoded into a fixed-length vector. In fact, the representation of the word should take into account the context information. The word vector model here cannot generate a word vector representation that takes into account the context.
[0095] The present invention obtains a low-dimensional vector sequence corresponding to the fault phenomenon description statement by inputting the vector sequence corresponding to the acquired fault phenomenon description statement into a preset bidirectional long short-term memory (Bi-directional-Long Short Term Memory Network, Bi-LSTM) model based on residual connection.
[0096] Alternatively, if Figure 3 As shown, O CLS ~O 19 The vector sequence is the vector sequence corresponding to the fault phenomenon description statement. CLS ~O 19 The vector sequence is input into the Bi-LSTM model. CLS ~O 19 Vector sequence input features are extracted. Assume that the input vector size is Where T is the number of time steps (when the input vector sequence is O CLS ~O 19 When T=20), h is the hidden vector dimension. After Bi-LSTM, the output vector dimension is Where m is the dimension of the latent vector output by the LSTM unit. Then, the Conv-1D module is used to reduce the dimension of the sample. After the Conv-1D module, the output vector dimension is This vector is the same as the original input vector Superposition is performed to minimize information loss and reduce the gradient disappearance problem.
[0097] Finally, the output vector sequence is pooled through the GlobalMaxPooling1D and GlobalAveragePooling1D layers to obtain a low-dimensional vector sequence. The first training sample is obtained by combining the input statistical features such as TF-IDF and word vector features. Perform training to obtain the final trained fault classifier.
[0098] The rail vehicle fault identification method provided by the present invention constructs a new LSTM structure based on a residual network, which can effectively complete the task of matching fault phenomenon description sentences. Compared with the related model of word vectors in the prior art, by considering the word vector information of the context in the fault phenomenon description sentence at the word level based on Bi-LSTM, a more accurate text matching effect can be achieved, thereby improving the accuracy of fault identification.
[0099] Furthermore, in one embodiment, the vector sequence corresponding to the fault phenomenon description statement is obtained in the following manner:
[0100] Input the fault phenomenon description sentence into a preset BERT model to obtain a vector sequence including the fault phenomenon description sentence;
[0101] The preset BERT model is obtained after being trained in advance with the second training sample;
[0102] The second training sample includes: a first text pair corresponding to the training corpus obtained by performing a preset reprocessing on the preprocessed training corpus;
[0103] The training corpus includes: the original corpus of the target rail vehicle and the historical fault tracking table data of the target rail vehicle.
[0104] Optionally, the obtained fault phenomenon description sentence is input into a preset BERT (Bidirectional Encoder Representation from Transformers) model to obtain a vector sequence including the fault phenomenon description sentence.
[0105] The preset BERT model is obtained after being pre-trained with the second training sample, and the second training sample can be obtained in the following way:
[0106] The original corpus of the target rail vehicle and the historical fault tracking table data of the target rail vehicle are obtained, and the original corpus and the historical tracking table data are used as training corpus, and preprocessing such as word segmentation, character segmentation and cleaning is adopted to obtain preprocessed training corpus, and the preprocessed training corpus is preprocessed to obtain a first text pair corresponding to the training corpus.
[0107] The preset reprocessing may specifically include preset reprocessing of each character of the word segmentation, character segmentation and cleaned training corpus, for example, replacing the characters in the preprocessed training corpus with special characters set in advance.
[0108] The original corpus of the target rail vehicle may specifically include: externally collected corpus data related to subway operation and maintenance, including various types of technical documents.
[0109] Alternatively, if Figure 4 As shown in the figure, the current corpus in the historical fault tracking data table is the sentence "the turnout is open in the locked state", which is replaced by "the road [M] is open in the locked state" after preset reprocessing. Another corpus is extracted from the original corpus and concatenated with the current corpus. There is a 50% probability that the corpus is the next corpus in the language of the current corpus; there is a 50% probability that it is a randomly extracted corpus unrelated to the current corpus. The two corpuses are separated by the symbol "[SEP]".
[0110] The data after the first two parts of processing (preprocessing and preset reprocessing) is sent to the Transformer model for training (such as Figure 4 The loss function of the training process mainly consists of two parts: the first part requires the model to predict whether the second paragraph is the next paragraph of the first paragraph, which is called NSP Loss; the second part requires the model to predict what the words in the masked position are, which is called Masked Token Loss. The Transformer model trained with the corpus obtained through the above special processing is called the BERT model.
[0111] The rail vehicle fault identification method provided by the present invention aims to solve the problem in the prior art that word vector representations that take the context of the corpus into consideration cannot be generated when model training is performed on the corpus, thereby resulting in insufficient text matching accuracy using the trained model. The BERT model is adopted to generate vector sequence representations of fault phenomenon description sentences, which not only includes vectors of words in the fault phenomenon description sentences, but also includes vector representations of sentences, thereby solving the semantic gap problem in semantic matching and the problem of generating word vectors that take the context into consideration, thereby making the fault classifier built based on the BERT model more accurate when identifying faults of target rail vehicles.
[0112] Furthermore, in one embodiment, the word vector of the fault phenomenon description sentence is obtained in the following manner:
[0113] Input the fault phenomenon description sentence into the preset word2vec model to obtain the word vector of the fault phenomenon description sentence;
[0114] The preset word2vec model is obtained after being trained in advance with the third training sample;
[0115] The third training sample includes: a second text pair corresponding to the training corpus obtained from the preprocessed training corpus.
[0116] Optionally, the training corpus (the original corpus of the target rail vehicle and the historical tracking table data of the target rail vehicle) is preprocessed to obtain a second text pair corresponding to the training corpus, and the obtained second text pair is used as a third training sample, and the preset word2vec model is trained using the third training sample.
[0117] For the third training sample, the third training sample consists of a series of sentences, assuming that represents the i-th sentence, sentence D i It consists of a series of participles, among which D i t Represents the tth participle in the sentence.
[0118] The commonly used model for vectorizing word segmentation is called the bag-of-words model. The principle of the bag-of-words model is to record all possible word segmentations in the training corpus, assuming that their number is V, and then construct a vector B∈R V×1 , each position of the vector B represents a segmentation; for the i-th sentence, its bag-of-words vector B∈R V×1 It is represented as 1 in the position containing the segmentation in the i-th sentence, and 0 in the other positions. The disadvantage of the bag-of-words model is that it only considers the statistical characteristics of words, but not the temporal characteristics of words.
[0119] Therefore, the present invention introduces the word2vec model, which is an extension of the bag-of-words model. The word2vec model is a shallow neural network model that includes an input layer, one or more hidden layers, and an output layer. i When training the word2vec model for the tth word segment of a sentence, the input of the word2vec model is the bag-of-words vectors of all words within the given window range w at position t; the output of the word2vec model is the bag-of-words vector of the tth word.
[0120] After training the word2vec model, for the tth word, the output of the hidden layer obtained by inputting the bag-of-words vector of the tth word into the word2vec model is the word vector of the tth word, and the dimension of the word vector is usually between 100 and 300. The word vector representation obtained by the word2vec model is usually more than 98% lower than the dimension of the bag-of-words vector representation of the bag-of-words model, and in the process of obtaining the representation of the segmented words, the relationship between the segmented words and the segmented words in the context is considered through the window parameter w, and the temporal relationship is naturally taken into account in the model. After the word2vec model is trained, the word vector of each segmented word can be obtained.
[0121] The obtained fault phenomenon description sentence is input into the trained vector model based on the word2vec model. The vector model will obtain the bag-of-words vector of each word in the fault phenomenon description sentence according to the preset bag-of-words model, and input it into the trained Word2vec model to obtain the word vector of each word. Then, by weighted summing up the word vectors of each word, the word vector s of the fault phenomenon description sentence can be further obtained. i , the specific weighted summation formula is as follows:
[0122]
[0123] Among them, the |D i | represents the number of segmented words in the segmented sequence of the fault description sentence to be identified, and the C t Represents the word vector corresponding to the tth word segmentation.
[0124] The rail vehicle fault identification method provided by the present invention builds a vector model based on the Word2vec model, and uses it to vectorize the fault phenomenon description sentence to obtain the word vector after training, so as to locate the fault corresponding to the fault phenomenon description sentence more quickly and accurately.
[0125] Furthermore, in one embodiment, preprocessing the historical fault tracking table data to obtain a fault phenomenon description statement corresponding to the historical fault tracking table data includes:
[0126] Perform word segmentation and character segmentation preprocessing on the historical fault tracking table data, and perform cleaning preprocessing on the historical fault tracking table data after word segmentation and character segmentation preprocessing according to preset cleaning preprocessing rules to obtain fault phenomenon description statements;
[0127] Among them, the preprocessing includes word segmentation and character segmentation preprocessing and cleaning preprocessing.
[0128] Optionally, perform preprocessing on the historical fault tracking table data to obtain fault phenomenon description statements corresponding to the historical fault tracking table data. Specifically:
[0129] Perform word segmentation preprocessing on the text data of the input historical fault tracking table data in units of natural sentences, and perform character segmentation preprocessing in units of characters at the same time; and clean the text data after word segmentation and character segmentation preprocessing according to preset cleaning preprocessing rules. Based on the preset stop word list and special symbol list, filter out meaningless modal particles, characters, and transition and stop words, such as words and symbols like "de", "le", "ma", "?", "|", etc., and obtain fault phenomenon description statements according to the historical fault tracking table data after the above-mentioned word segmentation, character segmentation preprocessing and cleaning preprocessing.
[0130] For the fault identification method of the rail vehicle provided by the present invention, when performing pre-training of the word2vec model, by performing word segmentation, character segmentation and cleaning processing on the input historical fault tracking table data, useful data in the corpus can be retained, noise data can be filtered out, and the quality of the data can be improved.
[0131] Furthermore, in one embodiment, perform preset reprocessing on the training corpus, including:
[0132] Extract the training corpus to obtain the first preset number of characters that meet the preset extraction rules;
[0133] Perform the first screening on the first preset number of characters to obtain the second preset number of characters, and replace the second preset number of characters with preset symbols;
[0134] Perform the second screening on the first preset number of characters to obtain the third preset number of characters, and replace the third preset number of characters with any characters.
[0135] Optionally, perform preset reprocessing on the obtained training corpus, extract all the training corpus to obtain the first preset number of characters that meet the preset extraction rules, where the preset extraction rules can be freely set according to the length of the training corpus text. For example, the preset extraction rule is: extract 15% of the characters of the training corpus.
[0136] The extracted training corpus is screened a second time, a second preset number of characters are screened out from the extracted 15% of the characters, and replaced with a preset symbol such as "[M]", and a third preset number of characters are screened out from the extracted 15% of the characters again, and replaced with any characters to obtain a first text pair corresponding to the training corpus.
[0137] For example, for the sentence "The turnout is open in the state of route locking", 15% of its characters are extracted and processed as follows: 80% of the characters are changed into "[M]" symbols, 10% of the characters are replaced with other arbitrary characters, and 10% of the characters remain the same. The result is "The turnout is open in the state of route locking".
[0138] The rail vehicle fault identification method provided by the present invention obtains a second training sample for training a preset BERT model by replacing part of the data of the training corpus with special symbols, so that the model trained according to the second training sample can learn the words that should be filled in the replaced places according to the given original training corpus, so that the trained BERT model can predict vocabulary based on the relationship between consecutive text fragments and rely on contextual information, and has a certain error correction ability.
[0139] Furthermore, in one embodiment, step S3 may specifically include:
[0140] S31. Match the fault description statement output by the fault classifier with the fault description statement corresponding to the fault case stored in the historical fault tracking table data to determine at least one fault case corresponding to the historical fault tracking table data.
[0141] Optionally, the fault phenomenon description statement to be identified and the historical fault tracking table data are input into a fault classifier to obtain a fault description statement corresponding to the fault phenomenon description statement to be identified, and the fault phenomenon description statement corresponding to the fault phenomenon description statement to be identified output by the fault classifier is matched with the fault phenomenon description statement of each fault case stored in the historical fault tracking table data, and a preset number of fault cases with the best match with the fault phenomenon description statement to be identified are filtered out from the historical fault tracking table data according to the matching results.
[0142] The rail vehicle fault identification method provided by the present invention builds a fault classifier based on a word2vec model, a BERT model and a Bi-LSTM model, and obtains a fault phenomenon description sentence corresponding to the fault phenomenon description sentence to be identified by inputting the fault phenomenon description sentence to be identified into the fault classifier, and matches the fault phenomenon description sentence output by the fault classifier with the fault case in the historical fault tracking table data to obtain the fault case corresponding to the fault phenomenon description sentence to be identified, thereby being able to more accurately locate the fault corresponding to the fault phenomenon description sentence to be identified.
[0143] The rail vehicle fault identification device provided by the present invention is described below. The rail vehicle fault identification device described below and the rail vehicle fault identification method described above can be referred to each other.
[0144] Figure 5 Schematic diagram of the structure of the fault identification device for rail vehicles provided by the present invention, such as Figure 5 As shown, it includes: a model parameter determination module 510, a fault classifier determination module 511 and a fault identification module 512;
[0145] A model parameter determination module 510 is used to input the first training sample into a preset fault classifier model to obtain parameters of the preset fault classifier model;
[0146] A fault classifier determination module 511 is used to determine a fault classifier for fault identification according to a preset fault classifier model and parameters;
[0147] A fault identification module 512 is used to input the description statement of the fault phenomenon to be identified and the data of the historical fault tracking table into the fault classifier to identify the fault of the target rail vehicle;
[0148] The first training sample includes: text features of the fault phenomenon description sentence, word vectors of the fault phenomenon description sentence, and low-dimensional vector sequences of the fault phenomenon description sentence;
[0149] Text features include: TF-IDF features and keyword features.
[0150] The rail vehicle fault identification device provided by the present invention is trained by a model parameter determination module 510 based on the text features, word vectors and low-dimensional vector sequences of the fault phenomenon description sentence, and is combined with a fault classifier determination module 511 to obtain a fault classifier for fault identification. The word vectors and sentence vectors of the fault phenomenon description sentence are fully considered. The output result of the fault classifier is matched with the historical fault tracking table data of the target rail vehicle through the fault identification module 512, thereby accurately and efficiently realizing the fault identification of the target rail vehicle.
[0151] Figure 6 is a schematic diagram of the physical structure of an electronic device provided by the present invention, such as Figure 6 As shown, the electronic device may include: a processor 610, a communication interface 611, a memory 612 and a bus 613, wherein the processor 610, the communication interface 611 and the memory 612 communicate with each other through the bus 613. The processor 610 may call the logic instructions in the memory 612 to execute the following method:
[0152] Inputting the first training sample into a preset fault classifier model to obtain parameters of the preset fault classifier model;
[0153] Determine a fault classifier for fault identification according to a preset fault classifier model and parameters;
[0154] Inputting the description statement of the fault phenomenon to be identified and the data of the historical fault tracking table into the fault classifier to identify the fault of the target rail vehicle;
[0155] The first training sample includes: text features of the fault phenomenon description sentence, word vectors of the fault phenomenon description sentence, and low-dimensional vector sequences of the fault phenomenon description sentence;
[0156] Text features include: TF-IDF features and keyword features.
[0157] In addition, the logic instructions in the above-mentioned memory can be implemented in the form of software functional units and can be stored in a computer-readable storage medium when sold or used as an independent product. Based on such an understanding, the technical solution of the present invention can be essentially or partly embodied in the form of a software product that contributes to the prior art. The computer software product is stored in a storage medium and includes several instructions for a computer device (which can be a personal computer, a server, or a network device, etc.) to perform all or part of the steps of the method described in each embodiment of the present invention. The aforementioned storage medium includes: U disk, mobile hard disk, read-only memory (ROM, Read-only Memory), random access memory (RAM, Random Access Memory), disk or optical disk and other media that can store program codes.
[0158] Furthermore, the present invention discloses a computer program product, the computer program product comprising a computer program stored on a non-transitory computer-readable storage medium, the computer program comprising program instructions, when the program instructions are executed by a computer, the computer can execute the rail vehicle fault identification method provided by the above-mentioned method embodiments, for example, comprising:
[0159] Inputting the first training sample into a preset fault classifier model to obtain parameters of the preset fault classifier model;
[0160] Determine a fault classifier for fault identification according to a preset fault classifier model and parameters;
[0161] Inputting the description statement of the fault phenomenon to be identified and the data of the historical fault tracking table into the fault classifier to identify the fault of the target rail vehicle;
[0162] The first training sample includes: text features of the fault phenomenon description sentence, word vectors of the fault phenomenon description sentence, and low-dimensional vector sequences of the fault phenomenon description sentence;
[0163] Text features include: TF-IDF features and keyword features.
[0164] On the other hand, the present invention further provides a non-transitory computer-readable storage medium having a computer program stored thereon, which is implemented when the computer program is executed by a processor to perform the rail vehicle fault identification method provided in the above embodiments, for example, including:
[0165] Inputting the first training sample into a preset fault classifier model to obtain parameters of the preset fault classifier model;
[0166] Determine a fault classifier for fault identification according to a preset fault classifier model and parameters;
[0167] Inputting the description statement of the fault phenomenon to be identified and the data of the historical fault tracking table into the fault classifier to identify the fault of the target rail vehicle;
[0168] The first training sample includes: text features of the fault phenomenon description sentence, word vectors of the fault phenomenon description sentence, and low-dimensional vector sequences of the fault phenomenon description sentence;
[0169] Text features include: TF-IDF features and keyword features.
[0170] The device embodiments described above are merely illustrative, wherein the units described as separate components may or may not be physically separated, and the components displayed as units may or may not be physical units, that is, they may be located in one place, or they may be distributed on multiple network units. Some or all of the modules may be selected according to actual needs to achieve the purpose of the scheme of this embodiment. Ordinary technicians in this field can understand and implement it without paying creative labor.
[0171] Through the description of the above implementation methods, those skilled in the art can clearly understand that each implementation method can be implemented by means of software plus a necessary general hardware platform, and of course, it can also be implemented by hardware. Based on this understanding, the above technical solution is essentially or the part that contributes to the prior art can be embodied in the form of a software product, and the computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, a disk, an optical disk, etc., including a number of instructions for a computer device (which can be a personal computer, a server, or a network device, etc.) to execute the methods described in each embodiment or some parts of the embodiments.
[0172] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or make equivalent replacements for some of the technical features therein. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of the present invention.
Claims
1. A rail vehicle fault identification method, characterized in that: include: Inputting the first training sample into a preset fault classifier model to obtain parameters of the preset fault classifier model; Determining a fault classifier for fault identification according to the preset fault classifier model and the parameters; Inputting the description statement of the fault phenomenon to be identified and the historical fault tracking table data into the fault classifier to perform fault identification on the target rail vehicle; The first training sample includes: text features of the fault phenomenon description sentence, word vectors of the fault phenomenon description sentence, and a low-dimensional vector sequence of the fault phenomenon description sentence; the low-dimensional vector sequence is obtained by reducing the dimension of the vector sequence corresponding to the fault phenomenon description sentence; The text features include: TF-IDF features and keyword features; The vector sequence corresponding to the fault phenomenon description statement is obtained in the following way: Inputting the fault phenomenon description sentence into a preset BERT model to obtain a vector sequence including the fault phenomenon description sentence; The preset BERT model is obtained after being pre-trained with a second training sample; the second training sample includes: a first text pair corresponding to the training corpus obtained after a preset reprocessing of the pre-processed training corpus; The pre-setting and re-processing of the training corpus includes: Extracting the training corpus to obtain a first preset number of characters that meet a preset extraction rule; Performing a first screening on the first preset number of characters to obtain a second preset number of characters, and replacing the second preset number of characters with a preset symbol; The first preset number of characters is screened for a second time to obtain a third preset number of characters, and the third preset number of characters are replaced with any characters.
2. The rail vehicle fault identification method according to claim 1, characterized in that: The first training sample is obtained by: Preprocessing the historical fault tracking table data to obtain a fault phenomenon description statement corresponding to the historical fault tracking table data; Inputting the vector sequence corresponding to the fault phenomenon description statement into a preset bidirectional long short-term memory Bi-LSTM model to obtain a low-dimensional vector sequence corresponding to the vector sequence; splicing the low-dimensional vector sequence, the word vector of the fault phenomenon description sentence, and the text feature of the fault phenomenon description sentence to obtain a first training sample; Wherein, the fault phenomenon description statement includes: positive sample data and negative sample data; The positive sample data is sample data composed of similar fault phenomenon description statements; The negative sample data is sample data composed of dissimilar fault phenomenon description statements.
3. The rail vehicle fault identification method according to claim 1, characterized in that: The training corpus includes: original corpus of the target rail vehicle and historical fault tracking table data of the target rail vehicle.
4. The rail vehicle fault identification method according to claim 2, characterized in that: The word vector of the fault phenomenon description sentence is obtained in the following way: Input the fault phenomenon description sentence into a preset word2vec model to obtain a word vector of the fault phenomenon description sentence; Wherein, the preset word2vec model is obtained after being trained in advance by a third training sample; The third training sample includes: a second text pair corresponding to the training corpus obtained from the preprocessed training corpus.
5. The rail vehicle fault identification method according to claim 3, characterized in that: The preprocessing of the historical fault tracking table data to obtain a fault phenomenon description statement corresponding to the historical fault tracking table data includes: Performing word segmentation and character segmentation preprocessing on the historical fault tracking table data, and performing cleaning preprocessing on the historical fault tracking table data after word segmentation and character segmentation preprocessing according to a preset cleaning preprocessing rule to obtain the fault phenomenon description statement; Wherein, the preprocessing includes the word segmentation and character segmentation preprocessing and the cleaning preprocessing.
6. The rail vehicle fault identification method according to any one of claims 1 to 5, characterized in that: The performing fault identification on the target rail vehicle comprises: The fault description statement output by the fault classifier is matched with the fault description statement corresponding to the fault case stored in the historical fault tracking table data to determine at least one fault case corresponding to the historical fault tracking table data.
7. A fault identification device for a rail vehicle, characterized in that: include: A model parameter determination module, a fault classifier determination module and a fault identification module; The model parameter determination module is used to input the first training sample into a preset fault classifier model to obtain parameters of the preset fault classifier model; The fault classifier determination module is used to determine a fault classifier for fault identification according to the preset fault classifier model and the parameters; The fault identification module is used to input the description statement of the fault phenomenon to be identified and the historical fault tracking table data into the fault classifier to perform fault identification on the target rail vehicle; The first training sample includes: text features of the fault phenomenon description sentence, word vectors of the fault phenomenon description sentence, and a low-dimensional vector sequence of the fault phenomenon description sentence; the low-dimensional vector sequence is obtained by reducing the dimension of the vector sequence corresponding to the fault phenomenon description sentence; The text features include: TF-IDF features and keyword features; The vector sequence corresponding to the fault phenomenon description statement is obtained in the following way: Inputting the fault phenomenon description sentence into a preset BERT model to obtain a vector sequence including the fault phenomenon description sentence; The preset BERT model is obtained after being pre-trained with a second training sample; the second training sample includes: a first text pair corresponding to the training corpus obtained after a preset reprocessing of the pre-processed training corpus; The pre-setting and re-processing of the training corpus includes: Extracting the training corpus to obtain a first preset number of characters that meet a preset extraction rule; Performing a first screening on the first preset number of characters to obtain a second preset number of characters, and replacing the second preset number of characters with a preset symbol; The first preset number of characters is screened for a second time to obtain a third preset number of characters, and the third preset number of characters are replaced with any characters.
8. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that: When the processor executes the computer program, the steps of the rail vehicle fault identification method according to any one of claims 1 to 6 are implemented.
9. A non-transitory computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the steps of the rail vehicle fault identification method according to any one of claims 1 to 6 are implemented.
Citation Information
Patent Citations
Fault identification method based on fault tracking table
CN110619046A
Power grid fault plan information extraction method and system based on deep learning
CN111428981A
