Entity recognition method in equipment fault domain based on dual neural network and scanning filtering

Abstract

The present application relates to a dual-neural network equipment fault domain entity recognition method based on scanning and filtering. The method comprises: segmenting the fault text according to the equipment fault domain dictionary to obtain segmented text; scanning the segmented text using a sliding window to obtain sequence segments; constructing filtering rules based on the entity characteristics of the equipment fault domain to filter the sequence segments to obtain candidate sequences; constructing an entity recognition model; the entity recognition model includes a Bert model, a BiLSTM network, a TextCNN network, and a softmax decoder; training the entity recognition model using a pre-set specific data training set, optimizing and verifying the entity recognition model based on optimal weights and parameters and a specific data test set, and identifying the candidate sequences based on the optimal entity recognition model to obtain entity recognition results. This method can improve the accuracy of entity recognition.

Description

Technical Field

[0001] The present application relates to the field of data processing technology, and in particular to a method for identifying entities in equipment failure fields using a dual neural network based on scanning and filtering, a computer device, and a storage medium. Background Art

[0002] With the continuous advancement of science and technology, mechanical equipment plays a vital role in various industries. To ensure the safe operation of equipment, health monitoring is essential. Health monitoring encompasses multiple tasks, including equipment fault diagnosis, status detection, and troubleshooting. The monitoring process generates a large amount of information, including failure cases and maintenance records, containing fault information. However, this information is typically recorded in text format, resulting in severe fragmentation and low standardization, making it difficult to effectively utilize. Furthermore, research on extracting textual information about equipment failures is limited. Therefore, when maintenance personnel need to obtain relevant repair information for a specific equipment failure, they can only manually search for fault records, making the information retrieval process time-consuming and labor-intensive. Therefore, how to automatically extract valuable information from existing equipment failure information is an urgent challenge in equipment health monitoring. With the development of deep learning models and text pre-training models, natural language processing technology has made significant progress in recent years. Information extraction, particularly named entity recognition, is a fundamental task for many downstream applications, such as knowledge graph construction and question answering. It can efficiently extract information about various entities in text, converting unstructured data into structured data, and effectively improving knowledge storage and query efficiency. Most current research approaches named entity recognition as a sequence labeling problem. Many researchers have proposed sequence labeling-based models that achieve good results in non-nested entity recognition tasks. However, in the extraction of equipment failure entities, significant entity nesting occurs. For example, in the fault text "The cause of a diesel engine's spontaneous shutdown may be a clogged diesel filter, which requires cleaning or replacing the filter element to troubleshoot," there is entity nesting. For example, the cause of the failure, "diesel filter clogged," also includes the component, "diesel filter." Clearly, in the field of equipment failure, it is necessary to simultaneously recognize both non-nested and nested entities.

[0003] To address the entity nesting problem in named entity recognition (NER), some current methods, following the principles of sequence labeling, propose methods such as composite labeling and hierarchical labeling. However, these methods often suffer from issues such as sparse label distribution, error propagation between layers, and difficulties in parallel training. Other methods consider the entity nesting problem as a hypergraph structure prediction problem. However, hypergraph-based models are often complex and fail to capture the rich boundary information between nested entities. Furthermore, such models do not achieve superior performance in non-nested entity recognition tasks. Furthermore, some approaches have chosen to address the entity nesting problem by directly classifying subsequences of sentences. In theory, span-based methods can perfectly solve NER tasks because they exhaustively enumerate all possible subsequences that could be entities. However, span-based methods generate a large number of negative samples, which not only increases classification time but also easily leads to overfitting and low entity recognition accuracy. Summary of the Invention

[0004] Based on this, it is necessary to provide a method, device, computer equipment and storage medium for entity recognition in the field of equipment failure based on scanning filtering, which can improve the accuracy of entity recognition, in order to address the above technical problems.

[0005] A dual neural network equipment fault domain entity recognition method based on scanning and filtering, the method comprising:

[0006] Obtain the fault text to be identified; segment the fault text according to the equipment fault domain dictionary to obtain the segmented text;

[0007] Use a sliding window to scan the segmented text to obtain sequence fragments;

[0008] According to the entity characteristics of the equipment failure field, filtering rules are constructed to filter sequence fragments to obtain candidate sequences;

[0009] Build an entity recognition model; the entity recognition model includes the Bert model, BiLSTM network, TextCNN network and softmax decoder;

[0010] The entity recognition model is trained using a pre-set specific data training set and the optimal weights and parameters during the model training process are saved. The entity recognition model is optimized and verified based on the optimal weights and parameters and the specific data test set to obtain the optimal entity recognition model.

[0011] The candidate sequence is identified according to the optimal entity recognition model to obtain the entity recognition result.

[0012] In one embodiment, the fault text is segmented according to the equipment fault domain dictionary to obtain the segmented text, including:

[0013] Determine the domain dictionary based on prior knowledge, import the domain dictionary into the spaCy word segmentation tool to segment the fault text and obtain the segmented text.

[0014] In one embodiment, filtering rules are constructed based on entity characteristics of the equipment failure domain to filter sequence segments to obtain candidate sequences, including:

[0015] The sequence segments are filtered according to the word parts of speech, grammatical rules, and semantic similarity in the sequence segments to obtain candidate sequences.

[0016] In one embodiment, identifying the candidate sequence according to the optimal entity recognition model to obtain an entity recognition result includes:

[0017] Input the candidate sequence into the Bert model to obtain the feature vector sequence;

[0018] The feature vector sequence is processed according to the BiLSTM network to obtain the hidden vector;

[0019] Use the TextCNN network to convolve the feature vector sequence to obtain the convolved feature vector;

[0020] Fuse the hidden vector and the convolved feature vector to obtain a fused feature vector;

[0021] The fused feature vector is decoded according to the softmax decoder to obtain the entity recognition result.

[0022] In one embodiment, the feature vector sequence is processed according to the BiLSTM network to obtain the hidden vector, including:

[0023] According to the BiLSTM network, the feature vector sequence is processed and the hidden vector is obtained as follows:

[0024] H out =[h1,h 2， …h t ]

[0025]

[0026] Among them, f t 、i t 、o t represents the forget gate, input gate, and output gate, c t represents the transition vector, C t represents the unit state at time t, h t is the hidden state, W* and b * represents the weight matrix and bias, σ and tanh represent nonlinear activation functions, x t For E bert , represents the feature vector sequence.

[0027] In one embodiment, a TextCNN network is used to convolve a feature vector sequence to obtain a convolved feature vector, including:

[0028] The TextCNN network is used to convolve the feature vector sequence, and the convolved feature vector is obtained as follows:

[0029]

[0030] Among them, E bert is a sequence of feature vectors, W represents the weight matrix of the convolution kernel, h represents the height of the convolution kernel, b represents the bias, and c i Represents the value obtained by the convolution kernel i-th convolution, C represents the feature vector obtained by convolution, C out represents the feature vector after convolution, Represents the feature vector obtained after maximum pooling, and N represents the length of the input sequence.

[0031] In one embodiment, the hidden vector and the convolved feature vector are fused to obtain a fused feature vector, including:

[0032] The hidden vector and the convolved feature vector are fused to obtain the fused feature vector:

[0033] H cat =[H out , C out ].

[0034] A computer device includes a memory and a processor, wherein the memory stores a computer program, and when the processor executes the computer program, the following steps are implemented:

[0035] Obtain the fault text to be identified; segment the fault text according to the equipment fault domain dictionary to obtain the segmented text;

[0036] Use a sliding window to scan the segmented text to obtain sequence fragments;

[0037] According to the entity characteristics of the equipment failure field, filtering rules are constructed to filter sequence fragments to obtain candidate sequences;

[0038] Build an entity recognition model; the entity recognition model includes the Bert model, BiLSTM network, TextCNN network and softmax decoder;

[0039] The entity recognition model is trained using a pre-set specific data training set and the optimal weights and parameters during the model training process are saved. The entity recognition model is optimized and verified based on the optimal weights and parameters and the specific data test set to obtain the optimal entity recognition model.

[0040] The candidate sequence is identified according to the optimal entity recognition model to obtain the entity recognition result.

[0041] A computer-readable storage medium stores a computer program, which, when executed by a processor, implements the following steps:

[0042] Obtain the fault text to be identified; segment the fault text according to the equipment fault domain dictionary to obtain the segmented text;

[0043] Use a sliding window to scan the segmented text to obtain sequence fragments;

[0044] According to the entity characteristics of the equipment failure field, filtering rules are constructed to filter sequence fragments to obtain candidate sequences;

[0045] Build an entity recognition model; the entity recognition model includes the Bert model, BiLSTM network, TextCNN network and softmax decoder;

[0046] The entity recognition model is trained using a pre-set specific data training set and the optimal weights and parameters during the model training process are saved. The entity recognition model is optimized and verified based on the optimal weights and parameters and the specific data test set to obtain the optimal entity recognition model.

[0047] The candidate sequence is identified according to the optimal entity recognition model to obtain the entity recognition result.

[0048] The above-mentioned dual neural network equipment fault field entity recognition method, computer device and storage medium based on scanning and filtering first segment the fault text according to the equipment fault field dictionary, increase the granularity of the fault text with words as the basic unit, reasonably set the sliding window size according to the entity length in different fields, and use the word part of speech, grammatical rules, semantic similarity and other aspects in the sequence fragments to filter the sequence fragments to obtain candidate sequences, extract the sequence fragments through different sampling windows, and then filter the sequence fragments according to the filtering rules to reduce the proportion of negative samples, and then use the Bert model to vector encode the candidate entity sequence obtained after scanning and filtering, so that the subsequent feature extraction module can extract features from the vector sequence to summarize the classification effect of the model. Combining the characteristics of BiLSTM that can extract long-term dependencies between characters in a sequence and TextCNN that can better obtain short text features, a dual neural network fusion method is used to extract sequence features, which improves the feature extraction performance and increases the robustness of the model, thereby improving the accuracy of entity recognition. The softmax decoder is then used to calculate the label probability of the sequence fragment to obtain the entity type of the sequence fragment, solving the problem of the difficulty in simultaneously identifying nested entities and non-nested entities in traditional named entity recognition methods. In addition, through word segmentation, scanning and filtering, the problems of high model complexity, large proportion of negative samples and difficulty in efficient model training in span-based entity recognition methods are solved. BRIEF DESCRIPTION OF THE DRAWINGS

[0049] Figure 1 1. A schematic flow chart of a method for identifying entities in equipment fault domains using a dual neural network based on scanning and filtering in one embodiment;

[0050] Figure 2 Schematic diagram of the overall model framework of a dual neural network equipment fault domain entity recognition method based on scanning and filtering in one embodiment;

[0051] Figure 3 is a schematic diagram of an entity recognition model in one embodiment;

[0052] Figure 4 A schematic diagram of changes in the F1 values of various models during model training in another embodiment;

[0053] Figure 5 FIG. 1 is a diagram showing the internal structure of a computer device in one embodiment. DETAILED DESCRIPTION

[0054] In order to make the purpose, technical solutions and advantages of this application more clear, the following further describes this application in detail with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain this application and are not intended to limit this application.

[0055] In one embodiment, Figure 1 As shown, a dual neural network equipment fault domain entity recognition method based on scanning filtering is provided, comprising the following steps:

[0056] Step 102: Obtain the fault text to be identified; segment the fault text according to the equipment fault domain dictionary to obtain a segmented text.

[0057] Step 104 : Scan the segmented text using a sliding window to obtain sequence segments; construct filtering rules based on entity characteristics in the equipment failure domain to filter the sequence segments and obtain candidate sequences.

[0058] like Figure 2 The figure shows the overall model framework of this application. First, the fault text is segmented according to the equipment fault domain dictionary, and the granularity of the fault text is increased using words as the basic unit. The sliding window size is reasonably set according to the length of entities in different fields. The sequence fragments are filtered using the word parts of speech, grammatical rules, and semantic similarity in the sequence fragments to obtain candidate sequences. Sequence fragments are extracted through different sampling windows and then filtered according to the filtering rules to reduce the proportion of negative samples. In addition, the filtering rules can also be used to formulate filtering rules suitable for specific fields based on different recognition samples and different entity types within the field.

[0059] Step 106, constructing an entity recognition model; the entity recognition model includes a Bert model, a BiLSTM network, a TextCNN network and a softmax decoder; using a pre-set specific data training set to train the entity recognition model and save the optimal weights and parameters during the model training process, the entity recognition model is optimized and verified based on the optimal weights and parameters and the specific data test set to obtain the optimal entity recognition model.

[0060] Building entity recognition models such as Figure 3 As shown, the Bert model is used to vector encode the candidate entity sequence obtained after scanning and filtering, so that the subsequent feature extraction module can extract features from the vector sequence to summarize the classification effect of the model. Combining the characteristics of BiLSTM that can extract long-term dependencies between characters in the sequence and TextCNN that can better obtain short text features, the dual neural network fusion method is used to extract sequence features, which improves the feature extraction performance and increases the robustness of the model, thereby improving the entity recognition accuracy. The softmax decoder is then used to calculate the label probability of the sequence fragment to obtain the entity type of the sequence fragment. Model training corpus is formulated according to specific fields and used to verify the actual performance of the model. The present invention uses the diesel engine failure data in this article to train and verify the model, proving that the method proposed in the present invention has good entity classification performance in the field of equipment failure.

[0061] Step 108: Identify the candidate sequence according to the optimal entity recognition model to obtain an entity recognition result.

[0062] In the above-mentioned dual neural network equipment fault field entity recognition method based on scanning and filtering, the fault text is first segmented according to the equipment fault field dictionary, and the granularity of the fault text is increased with words as the basic unit. The sliding window size is reasonably set according to the entity length in different fields. The word parts of speech, grammatical rules, semantic similarity and other aspects in the sequence fragments are used to filter the sequence fragments to obtain candidate sequences. The sequence fragments are extracted through different sampling windows, and then the sequence fragments are filtered according to the filtering rules to reduce the proportion of negative samples. The Bert model is then used to vector encode the candidate entity sequence obtained after scanning and filtering, so that the subsequent feature extraction module can extract features from the vector sequence to summarize the classification effect of the model. Combining the characteristics of BiLSTM that can extract long-term dependencies between characters in a sequence and TextCNN that can better obtain short text features, a dual neural network fusion method is used to extract sequence features, which improves the feature extraction performance and increases the robustness of the model, thereby improving the accuracy of entity recognition. The softmax decoder is then used to calculate the label probability of the sequence fragment to obtain the entity type of the sequence fragment, solving the problem of the difficulty in simultaneously identifying nested entities and non-nested entities in traditional named entity recognition methods. In addition, through word segmentation, scanning and filtering, the problems of high model complexity, large proportion of negative samples and difficulty in efficient model training in span-based entity recognition methods are solved.

[0063] In one embodiment, the fault text is segmented according to the equipment fault domain dictionary to obtain the segmented text, including:

[0064] Determine the domain dictionary based on prior knowledge, import the domain dictionary into the spaCy word segmentation tool to segment the fault text and obtain the segmented text.

[0065] In one embodiment, filtering rules are constructed based on entity characteristics of the equipment failure domain to filter sequence segments to obtain candidate sequences, including:

[0066] The sequence segments are filtered according to the word parts of speech, grammatical rules, and semantic similarity in the sequence segments to obtain candidate sequences.

[0067] In one embodiment, identifying the candidate sequence according to the optimal entity recognition model to obtain an entity recognition result includes:

[0068] Input the candidate sequence into the Bert model to obtain the feature vector sequence;

[0069] The feature vector sequence is processed according to the BiLSTM network to obtain the hidden vector;

[0070] Use the TextCNN network to convolve the feature vector sequence to obtain the convolved feature vector;

[0071] Fuse the hidden vector and the convolved feature vector to obtain a fused feature vector;

[0072] The fused feature vector is decoded according to the softmax decoder to obtain the entity recognition result.

[0073] In one embodiment, the feature vector sequence is processed according to the BiLSTM network to obtain the hidden vector, including:

[0074] According to the BiLSTM network, the feature vector sequence is processed and the hidden vector is obtained as follows:

[0075] H out =[h1, h2, ... h t ]

[0076]

[0077] Among them, f t 、i t 、o t represents the forget gate, input gate, and output gate, c t represents the transition vector, C t represents the unit state at time t, h t is the hidden state, W * and b * represents the weight matrix and bias, σ and tanh represent nonlinear activation functions, x t For E bert , represents the feature vector sequence.

[0078] In one embodiment, a TextCNN network is used to convolve a feature vector sequence to obtain a convolved feature vector, including:

[0079] The TextCNN network is used to convolve the feature vector sequence, and the convolved feature vector is obtained as follows:

[0080]

[0081] Among them, E bert is a sequence of feature vectors, W represents the weight matrix of the convolution kernel, h represents the height of the convolution kernel, b represents the bias, and c i Represents the value obtained by the convolution kernel i-th convolution, C represents the feature vector obtained by convolution, C out represents the feature vector after convolution, Represents the feature vector obtained after maximum pooling, and N represents the length of the input sequence.

[0082] In one embodiment, the hidden vector and the convolved feature vector are fused to obtain a fused feature vector, including:

[0083] The hidden vector and the convolved feature vector are fused to obtain the fused feature vector:

[0084] H cnt =[H out , C out ].

[0085] In a specific embodiment, the present invention uses the NSP task (NextSentencePrediction) to implement the BERT pre-training process. The NSP task is mainly used to determine whether the two input texts are in a previous and next sentence relationship. Usually, a classification character [CLS] is inserted before the two sentences, and a segmentation character [SEP] is inserted between the two sentences. Since the application inputs a single text sequence into BERT, the segmentation character [SEP] does not need to be inserted during model training, and all semantic vectors output by BERT are selected as the input of the subsequent model.

[0086] Bert can be regarded as a nonlinear mapping function from character sequence to feature vector, and the output of Bert can be expressed as:

[0087] E bert =Bert(X′)

[0088] Where, X′={x′1,x′2,x′3,…,x′ N}, x′ i Represents the i-th character in the input sequence, and N represents the length of the input sequence. bert =[E1,E2,E3,…E N ] as the input of BiLSTM and TextCNN.

[0089] BiLSTM consists of forward LSTM and backward LSTM, which can better obtain sequence context information. The calculation method of LSTM is shown in formula (1). and reverse LSTM output Splicing E bert The output of BiLSTM is Ho ut =[h1,h2,…h t ],

[0090]

[0091] Where, f t 、i t 、o t represents the forget gate, input gate, and output gate, c t represents the transition vector, C t represents the unit state at time t, h t is the hidden state. * and b * Represents the weight matrix and bias, σ and tanh are nonlinear activation functions

[0092] The vector E can be bert Expressed as The calculation process of TextCNN is shown in formula (2), E bert The output of TextCNN is

[0093]

[0094] Where W represents the weight matrix of the convolution kernel, h represents the height of the convolution kernel; b represents the bias; c i Represents the value obtained by the convolution kernel i-th convolution; C represents the feature vector obtained by convolution, and C is obtained by performing maximum pooling on C. q represents the number of convolution kernels.

[0095] The output H of BiLSTM out =[h1,h2,…h t ] and the output of TextCNN Splicing to obtain the fusion feature H cat =[H out ,C out ].

[0096] In one embodiment, segmenting the fault text "The cause of the diesel engine's self-shutdown may be a clogged diesel filter, which needs to be cleaned or replaced to troubleshoot the fault" will result in "diesel engine / self-shutdown / fault / cause / may / be / diesel filter / clogged / needs / cleaning / or / replacement / diesel filter element / troubleshooting".

[0097] The sliding window size is limited based on the number of words contained in the entities within the domain. For example, if the number of words contained in the fault entity in the diesel engine fault text data is less than 6, the maximum window size can be limited to 6 when setting the sliding window. This can effectively reduce the number of sequence fragments. For example, the fault text "Injector carbon deposits cause diesel engine exhaust failure" has 105 subsequences based on words as the basic unit, but after word segmentation, the resulting word segmentation sequence is: "Injector / carbon deposits / cause / diesel engine / exhaust / failure", and the number of subsequences in the word segmentation sequence is 21. This process will effectively solve the problem of high spatiotemporal complexity caused by classifying all subsequences in traditional span-based entity recognition methods.

[0098] The filtering rules are constructed based on the entity characteristics in the diesel engine fault text as shown in Table 1:

[0099] Table 1 Filtering rules

[0100]

[0101] As shown in Table 1, when the scan window size is 1, sequence segments are simply individual words. For faulty entities, non-noun words cannot be faulty entities and can be directly filtered. When the scan window is larger than 1, sequence segments contain words of multiple parts of speech, and the filtering method is shown in the table. While the filtering rules shown in the table clearly cannot fully cover all negative sample data, they can filter out most meaningless sequence segments. Furthermore, rule-based filtering allows for specific domain-specific filtering rules based on different recognition samples and entity types within the domain.

[0102] Construct an entity recognition model; the entity recognition model includes a Bert model, a BiLSTM network, a TextCNN network, and a softmax decoder. The Bert pre-trained model is used to encode the features of the sequence fragments, and BiLSTM and convolutional CNN are used to fuse and extract the sequence features. Finally, the entity labels of the sequence fragments are predicted.

[0103] The entity recognition model was trained using diesel engine fault text data. In the field of entity relationship extraction, the accuracy and recall of different models were compared. When the two indicators are high and low and it is difficult to directly judge the performance, the F1 value can be compared, which is a comprehensive performance of precision and recall. The evaluation index calculation formula is (3) to (5).

[0104]

[0105]

[0106]

[0107] Where TP i is the number of entities correctly identified as type i, FP i is the number of entities that are incorrectly identified as type i, FN i is the number of entities of type i that are mistakenly identified as other entities, and n is the number of entity categories.

[0108] The diesel engine fault text dataset contains 644 text data and 3040 fault entities. The ratio of training set, validation set and test set is approximately 7:1:2.

[0109] Figure 4 The following table shows the F1 performance of different models during training. Among the basic models, BiLSTM performed best overall, reaching convergence after approximately 20 iterations, with a stable curve and a high F1 score. TextCNN performed the worst. The model converged around the 25th iteration, but the subsequent curve showed a slight upward trend. However, its overall performance improvement was not significant, and further training rounds did not significantly improve its performance. RCNN's F1 score rose slowly at the beginning of training, with no significant change in the first few iterations. However, as the number of iterations increased, it achieved performance similar to that of BiLSTM, indicating that it requires more training rounds to achieve good classification results. DPCNN performed poorly on the validation set, with slow convergence and large F1 fluctuations. Due to its complexity, the Transformer model did not converge until approximately 40 iterations, but its F1 score improved significantly with increasing iterations, reaching a final F1 score similar to that of BiLSTM. The performance of the basic model on the validation set shows that the overall performance of the RNN-based model is better than the CNN-based text classification model. However, after introducing the BERT pre-training model, the overall performance of the model is greatly improved, the convergence speed is faster, and its F1 value is higher than all baseline models. The model proposed in this paper is obviously superior to other models in the training process.

[0110] The diesel engine fault text data will be used to test the real performance of the model, and the results are shown in Table 2.

[0111] Table 2 Experimental results

[0112]

[0113] TextCNN performed worst on the test set, with an F1 score of only 79.50% and a recall rate of only 70.07%. BiLSTM, on the other hand, performed best on the test set, with an F1 score of 90.34%, demonstrating the superiority of RNNs in text classification tasks. DPCNN and RCNN achieved similar F1 scores, but RCNN's recall rate was 2.18% higher than DPCNN, indicating that RCNN performed better overall. Based on the validation set results, the proposed model achieved the best results in precision, recall, and F1 score.

[0114] It should be understood that although Figure 1 The steps in the flowchart are shown in sequence as indicated by the arrows, but these steps are not necessarily executed in the order indicated by the arrows. Unless otherwise specified in this document, there is no strict order restriction for the execution of these steps, and these steps can be executed in other orders. In addition, Figure 1 At least part of the steps may include multiple sub-steps or multiple stages. These sub-steps or stages are not necessarily executed at the same time, but can be executed at different times. The execution order of these sub-steps or stages is not necessarily sequential, but can be executed in turn or alternately with other steps or at least part of the sub-steps or stages of other steps.

[0115] In one embodiment, a computer device is provided. The computer device may be a terminal, and its internal structure diagram may be as follows: Figure 5 As shown. The computer device includes a processor, memory, network interface, display screen and input device connected via a system bus. The processor of the computer device is used to provide computing and control capabilities. The memory of the computer device includes a non-volatile storage medium and an internal memory. The non-volatile storage medium stores an operating system and a computer program. The internal memory provides an environment for the operation of the operating system and the computer program in the non-volatile storage medium. The network interface of the computer device is used to communicate with an external terminal via a network connection. When the computer program is executed by the processor, a dual neural network equipment fault field entity recognition method based on scanning filtering is implemented. The display screen of the computer device can be a liquid crystal display screen or an electronic ink display screen, and the input device of the computer device can be a touch layer covering the display screen, or a button, trackball or touchpad provided on the computer device housing, or an external keyboard, touchpad or mouse.

[0116] Those skilled in the art will understand that Figure 5The structure shown in the figure is only a block diagram of a part of the structure related to the solution of the present application, and does not constitute a limitation on the computer device to which the solution of the present application is applied. The specific computer device may include more or fewer components than shown in the figure, or combine certain components, or have a different component arrangement.

[0117] Those skilled in the art will appreciate that all or part of the processes in the above-mentioned embodiments can be implemented by instructing the relevant hardware through a computer program. The computer program can be stored in a non-volatile computer-readable storage medium. When the computer program is executed, it can include the processes of the embodiments of the above-mentioned methods. Among them, any reference to memory, storage, database or other media used in the embodiments provided in this application can include non-volatile and / or volatile memory. Non-volatile memory can include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM) or flash memory. Volatile memory can include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM is available in various forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), double data rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), synchronous link (Synchlink) DRAM (SLDRAM), memory bus (Rambus) direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM (RDRAM).

[0118] The technical features of the above embodiments can be combined arbitrarily. To make the description concise, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

Claims

1. A dual neural network equipment fault domain entity recognition method based on scanning and filtering, characterized in that: The method comprises: Obtaining a fault text to be identified; segmenting the fault text according to a dictionary in the field of equipment faults to obtain a segmented text; Scanning the segmented text using a sliding window to obtain sequence fragments; Constructing filtering rules based on entity characteristics in the equipment failure field to filter the sequence fragments to obtain candidate sequences; Constructing an entity recognition model; the entity recognition model includes a Bert model, a BiLSTM network, a TextCNN network and a softmax decoder; The entity recognition model is trained using a preset specific data training set and the optimal weights and parameters in the model training process are saved. The entity recognition model is optimized and verified based on the optimal weights and parameters and the specific data test set to obtain an optimal entity recognition model. Identify the candidate sequence according to the optimal entity recognition model to obtain an entity recognition result; Identify the candidate sequence according to the optimal entity recognition model to obtain an entity recognition result, include: Inputting the candidate sequence into the Bert model to obtain a feature vector sequence; Processing the feature vector sequence according to the BiLSTM network to obtain a hidden vector; Convolving the feature vector sequence using the TextCNN network to obtain a convolved feature vector; Fusing the hidden vector and the convolved feature vector to obtain a fused feature vector; Decoding the fused feature vector according to the softmax decoder to obtain an entity recognition result; The feature vector sequence is processed according to the BiLSTM network to obtain a hidden vector, including: The feature vector sequence is processed according to the BiLSTM network to obtain the hidden vector: H out =[h1,h2,…h t ] Among them, f t 、i t 、o t represents the forget gate, input gate, and output gate, c t represents the transition vector, C t represents the unit state at time t, h t is the hidden state, W * and b * represents the weight matrix and bias, σ and tanh represent nonlinear activation functions, x t For E bert , represents the feature vector sequence; The feature vector sequence is convolved using the TextCNN network to obtain a convolved feature vector, including: The TextCNN network is used to convolve the feature vector sequence, and the convolved feature vector is obtained as follows: Among them, E bert is a sequence of feature vectors, W represents the weight matrix of the convolution kernel, h represents the height of the convolution kernel, b represents the bias, and c i Represents the value obtained by the convolution kernel i-th convolution, C represents the feature vector obtained by convolution, Represents the feature vector obtained after maximum pooling, C out Represents the feature vector after convolution, q represents the number of convolution kernels, and N represents the length of the input sequence.

2. The method according to claim 1, characterized in that The fault text is segmented according to the equipment fault domain dictionary to obtain a segmented text, including: A domain dictionary is determined based on prior knowledge, and the domain dictionary is imported into the spaCy word segmentation tool to segment the fault text to obtain a segmented text.

3. The method according to claim 1, characterized in that According to the entity characteristics of the equipment failure field, filtering rules are constructed to filter the sequence fragments to obtain candidate sequences, including: The sequence segments are filtered according to the word parts of speech, grammatical rules, and semantic similarity in the sequence segments to obtain candidate sequences.

4. The method according to claim 1, wherein The hidden vector and the convolved feature vector are fused to obtain a fused feature vector, including: The hidden vector and the convolved feature vector are fused to obtain the fused feature vector: H cat =[H out ,C out ]。

Images