CN113901815B - Emergency working condition event detection method based on dam operation log - Google Patents

Abstract

The invention discloses an emergency working condition event detection method for a dam operation log, which comprises the steps of constructing a dam emergency working condition event type set; coding all the word fragments in the dam operation log, and converting the code fragments into embedded vectors corresponding to the word fragments; fusing an embedded vector, a named entity type and a part-of-speech tagging vector corresponding to the word segmentation, and strengthening semantic information of the word segmentation; the sentence-document dual attention fusion context information is used, the sentence-level attention is improved to the important words which possibly trigger events in each sentence, the document-level attention is improved to the important sentences which possibly trigger events in each log document, the local and global semantic information of word segmentation is enhanced, and the problems of word polysense and word and trigger mismatch in traditional Chinese event detection are solved; in order to avoid the problem of unbalanced classification of the positive and negative samples of the classification caused by that each sentence in the common dam log document contains 2 events at most, a training model is adopted to detect the events, and classification of all the documents is realized based on the events contained in each document.

Description

Emergency event detection method based on dam operation log

Technical Field

The present invention relates to an emergency working condition event detection method based on a dam operation log, which performs event detection on the dam operation log in the hydraulic engineering field, specifically performs event detection on various special working condition events and response events experienced by the dam under long-term operation, and belongs to the technical field of natural language processing.

Background Art

The task of event detection is to identify event trigger words from large-scale unstructured natural language texts and correctly classify event types. Trigger words refer to the core words or phrases that can most clearly express the occurrence of events. Event detection is of great significance to event semantic modeling, which facilitates the subsequent structured management and storage of events.

In the field of water conservancy projects, dam facilities provide multiple functions such as flood control, ice prevention, water storage, water supply, and power generation, and are the mainstay of the development of my country's water conservancy. During the long-term operation of several decades, the dam will encounter many natural risk events, such as floods, earthquakes, rainstorms, etc., which may endanger the safety of the dam structure and the safety of life and property of people downstream of the dam. Therefore, after a special event occurs, the dam management personnel will arrange a comprehensive special inspection to maintain the dam structure. In addition, the daily inspection and maintenance of the dam is also an important measure to ensure the safety of the dam body. After various response measures, the inspection personnel will record the cause and inspection results of this inspection event in writing to form a dam operation log file.

By processing the dam operation log, we can analyze the safety status of the dam since its construction and form a dam event knowledge base to improve the intelligent management level of the dam. The emergency condition event detection method for the dam operation log can skip the event trigger to detect all scheduled events in the dam operation log and classify each document into event types, providing a basis for subsequent event extraction, event graph construction and event knowledge base construction tasks.

There are a lot of ambiguities in Chinese texts, and events are generally composed of event triggers and event arguments. Event triggers are mostly verbs, which generally have polysemy and mismatch between triggers and words, which makes event detection methods based on trigger identification prone to classification errors.

Summary of the invention

Purpose of the invention: In view of the problems existing in the prior art and the various natural events and their response measures encountered during the operation of the dam, there is a lack of standardized records of events. The present invention provides an emergency operating condition event detection method based on the dam operation log, which avoids the process of identifying triggers, solves the above problems by simulating triggers in sentences, detects special dam operating condition events from the dam operation log, and classifies the event type to which each document belongs, providing a basis for subsequent event extraction.

Technical solution: A method for detecting emergency conditions based on dam operation logs, comprising the following steps:

(1) Log file preprocessing: First, the dam operation log is sorted and split according to the recording date, each document is numbered, the sentences in each document are sorted, numbered and segmented, each word is marked with entity type and part of speech, and then a set of dam emergency condition event types is constructed; the sorting refers to sorting logs of different dates; the splitting refers to splitting the logs of the same day according to the document content;

(2) Encoded vector embedding: Use the ALBERT preprocessing model to encode all the words in the dam operation log and convert them into embedding vectors corresponding to the words;

(3) BiLSTM feature fusion: Use BiLSTM to fuse the embedding vector, named entity type, and part-of-speech tagging vector corresponding to the word segmentation to enhance the semantic information of the word segmentation;

(4) Dual attention mechanism semantic enhancement: Use sentence-document dual attention to fuse contextual information. Sentence-level attention increases the important words in each sentence that may trigger an event, and document-level attention increases the important sentences in each log document that may trigger an event, solving the problems of polysemy and mismatch between words and triggers in traditional Chinese event detection.

(5) Using the Focal loss function to train the model and realize classification: In order to avoid the imbalance problem of positive and negative samples in the binary classification caused by the fact that each sentence in the ordinary dam log document contains at most two events, the Focal loss function is used to train the model to realize the classification of the events to which all documents belong.

The dam emergency condition event type collection includes typical events such as earthquakes, heavy rains, flood discharges, pre-flood safety inspections, comprehensive special inspections, routine maintenance, and daily inspections.

The named entity types include names, departments, locations, times, dates, measured values, percentages, defect types, etc.; the part-of-speech tagging vectors include nouns, verbs, adjectives, quantifiers, pronouns, etc.

Furthermore, the step (1) includes the following steps:

(1.1) First, divide the dam operation log file into multiple documents according to the log record date, sort and number each document, sort and number the sentences in each document, and use jieba to segment words;

(1.2) Perform entity type tagging and part-of-speech tagging on the word segmentation results. Entity type tagging converts entity type tags into low-dimensional vectors by searching a randomly initialized embedding table. Part-of-speech tagging uses Stanford CoreNLP to mark the part of speech of each word, and then converts the part-of-speech tags into low-dimensional vectors by searching the corresponding embedding table.

(1.3) Predefine dam emergency event types, including earthquakes, heavy rains, flood discharges, pre-flood safety inspections, comprehensive special inspections, routine maintenance, routine inspections and other typical events.

Furthermore, the step (2) includes the following steps:

Use the ALBERT pre-trained model to encode all the word segmentations in (1.1) and convert them into vector representations that can be processed by computers.

Furthermore, the step (3) includes the following steps:

(3.1) Concatenate the embedding vector, entity type vector and part-of-speech tag vector corresponding to each word, where the embedding vector is the vector obtained in step (2), the entity type vector is the mathematical vector corresponding to all word segmentation named entity recognition results such as names, organizations, locations, times, dates, numbers, percentages, etc., and the part-of-speech tag vector is the mathematical vector corresponding to all word segmentation part-of-speech tag results such as nouns, verbs, adjectives, quantifiers, pronouns, etc.;

和

将该两个向量合成为输出向量

(3.2) Use the BiLSTM model to process the concatenated vectors in a single sentence. Each vector is an input, and the bidirectional LSTM unit is used to capture the word context information and output two hidden states respectively.

and

Combine these two vectors into an output vector

Furthermore, the step (4) includes the following steps:

(4.1) In the training set, the emergency condition predefined events contained in each sentence are converted into embedding vectors t ₁ by searching the randomly initialized embedding table, and each document is converted into an embedding vector d using Dov2Vec;

(4.2) For all the words in each sentence, use the local attention mechanism to calculate the weight of each word in the sentence, increase the attention weight of the word that triggers the target event type and simulate the trigger. The calculation formula is as follows:

是局部注意力向量α_s中第k个部分，

是事件类型嵌入向量的转置；所述触发器指代事件触发器，即触发某事件的词，一般为动词；where h _k is the kth part of the output vector h,

is the kth part of the local attention vector _αs ,

is the transpose of the event type embedding vector; the trigger refers to an event trigger, that is, a word that triggers an event, generally a verb;

(4.3) For all the word segments in each sentence, use the global attention mechanism to calculate the weight of the sentence where the word segment is located in its document, obtain the unique meaning of the trigger in this scenario, assist in determining the event type of the sentence, and solve the ambiguity problem of the trigger caused by context information. The calculation formula is as follows:

是全局注意力向量α_d中第k个部分，

是事件类型嵌入向量转置，d^T是文档级嵌入向量转置；where h _k is the kth part of the output vector h,

is the kth part of the global attention vector _αd ,

is the transposed event type embedding vector, d ^T is the transposed document level embedding vector;

(4.4) Weighted fusion of local attention and global attention to improve event detection accuracy. The weight vector of local attention and global attention for events and the weighted fusion formula are calculated as follows:

v _s = α _s · t ₁

v _d = α _d · t ₂

o＝σ(λ·v _s +(1-λ)·v _d )

Among them, the final output value o consists of two parts: _vs and _vd . _vs is generated by the dot product _{of αs} and the event type embedding vector _t1 , which is used to capture local features and simulate hidden event triggers; _vd is generated by the dot product _{of αd} and _t2 , which is used to capture global features and contextual information. σ is the Sigmoid function, and λ∈[0,1] is a hyperparameter that weighs between _vs and _vd .

Furthermore, the step (5) includes the following steps:

The data set is processed in units of sentences, and the training data is composed of <sentence, event type> pairs, which represent whether a given sentence conveys an event of type t. The event type label is 1 or 0. For example, the label of the training pair <inspection of the bank of the reservoir near the dam, the slope of the hub area and the road: no abnormality, daily inspection> is 1, and the label of the training pair <inspection of the bank of the reservoir near the dam, the slope of the hub area and the road: no abnormality, earthquake> is 0. Since the number of events that a single sentence may express is very small compared with the predefined number of events, the imbalance problem of the number of negative samples far exceeding the number of positive samples caused by binary classification recognition is introduced. The model trained by the Focal loss loss function is strengthened to enhance the influence of positive samples and difficult samples on the model. The calculation formula is as follows:

Where x is composed of a sentence and a target event type, y∈{0,1}, o(x ⁽ⁱ⁾ ) is the model prediction value, ||θ|| ² is the sum of the squares of the elements in the model, δ>0 is the weight of the L2 normalization term, β is a parameter for balancing the positive and negative weight ratios of samples, and γ is a parameter for balancing the weight ratios of difficult and easy classification samples. In this experiment, β=0.25 and γ=2 are set.

Finally, the trained model is used to perform event detection on the dam operation log files and classify them based on the event types contained in each document.

An emergency condition event detection system based on dam operation logs performs event detection on dam operation logs in the water conservancy field, including:

Log file preprocessing module: first, sort and split the dam operation log according to the record date, label each document, sort, label and segment the sentences in each document, label each word with entity type and part of speech, and then construct a set of dam emergency condition event types;

Encoded vector embedding module: Use the ALBERT preprocessing model to encode all the words in the dam operation log and convert them into embedding vectors corresponding to the words;

BiLSTM feature fusion module: Use BiLSTM to fuse the embedding vector, named entity type and part-of-speech tagging vector corresponding to the word segmentation to enhance the semantic information of the word segmentation;

Dual attention mechanism semantic enhancement module: uses sentence-document dual attention to fuse contextual information;

A model: Use the focal loss function to train the above model, and use the trained model to classify all document attribution events.

A computer device comprises a memory, a processor and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, the emergency condition event detection method based on the dam operation log as described above is implemented.

A computer-readable storage medium stores a computer program for executing the emergency condition event detection method based on a dam operation log as described above.

Beneficial effects: Compared with the prior art, the emergency condition event detection method based on dam operation log provided by the present invention captures keywords and sentence-level semantic information through local attention, simulates hidden event triggers, realizes event detection without triggers, introduces rich document-level context information through global attention, assists in judging the meaning of words in the real context, skips the trigger recognition link, and directly judges the event type. It avoids the problem of mismatch between Chinese words and triggers and polysemy of a word, and improves the accuracy of event detection.

BRIEF DESCRIPTION OF THE DRAWINGS

FIG1 is a flow chart of model training according to an embodiment of the present invention;

FIG2 is a diagram of a model training framework according to an embodiment of the present invention.

DETAILED DESCRIPTION

The present invention is further explained below in conjunction with specific embodiments. It should be understood that these embodiments are only used to illustrate the present invention and are not used to limit the scope of the present invention. After reading the present invention, various equivalent forms of modifications to the present invention by those skilled in the art all fall within the scope defined by the claims attached to this application.

As shown in Figure 1, the emergency condition event detection method based on the dam operation log mainly includes the following steps:

Step (1) Preprocessing dam operation log files

(1.1) First, the dam operation log file is divided into multiple documents according to the log record date as the training set, and each document is sorted and numbered. The sentences in each document are sorted and numbered, and Jieba is used to segment words. As shown in Figure 2, the sentence "Inspection of the bank near the dam, the slope of the hub area and the road: no abnormality" is first split into "near the dam", "bank", "," "hub area", "slope", "and", "road", "inspection", "situation", ":", "no", "abnormal".

(1.2) Perform entity type tagging and part-of-speech tagging on the word segmentation results. Entity type tagging converts it into a low-dimensional vector by searching a randomly initialized embedding table. Part-of-speech tagging uses Stanford CoreNLP to mark the part of speech of each word, and then converts it into a low-dimensional vector by searching the embedding table.

(1.3) Predefine the types of dam emergency conditions in the dam operation log, including earthquakes, heavy rains, flood discharges, pre-flood safety inspections, comprehensive special inspections, routine maintenance, daily inspections and other typical events.

Step (2) Encode the word segmentation into word vectors

Use the ALBERT pre-trained model to encode all the word segmentations in (1.3) and convert them into vector representations that can be processed by computers.

Step (3) extracts semantic information by concatenating word vectors, named entity types, and part-of-speech tag vectors.

(3.1) Concatenate the embedding vector, entity type vector and part-of-speech tag vector corresponding to each word segmentation, where the embedding vector is the vector obtained in step (2), the entity type vector is the mathematical vector corresponding to all word segmentation named entity recognition results such as name, organization, location, time, date, measurement value, percentage, defect type, etc., and the part-of-speech tag vector is the mathematical vector corresponding to the part-of-speech tag results of all word segmentations such as noun, verb, adjective, quantifier, pronoun, etc.

和

将该两个向量合成为输出向量

(3.2) Use the BiLSTM model to process the concatenated vectors in a single sentence. Each vector is an input, and the bidirectional LSTM unit is used to capture the word context information and output two hidden states respectively.

and

Combine these two vectors into an output vector

Step (4) uses a dual attention mechanism to capture sentence-level context and document-level context, strengthen word vector representation, and simulate triggers.

(4.1) In the training set, the events contained in each sentence are converted into an embedding vector t ₁ by looking up a randomly initialized embedding table, and each document is converted into an embedding vector d using Dov2Vec.

(4.2) For all the words in each sentence, use the local attention mechanism to calculate the weight of each word in the sentence, increase the attention weight of the word that triggers the target event type and simulate the trigger. The calculation formula is as follows:

是局部注意力向量α_s中第k个部分，

是事件类型嵌入向量的转置。如图2中

用于辅助局部注意力机制，对每个分词模拟触发器。where h _k is the kth part of the output vector h,

is the kth part of the local attention vector _αs ,

is the transpose of the event type embedding vector. As shown in Figure 2

Used to assist the local attention mechanism and simulate triggers for each word segmentation.

(4.3) For all the word segments in each sentence, use the global attention mechanism to calculate the weight of the sentence where the word segment is located in this document, obtain the unique meaning of the trigger in this scenario, assist in determining the event type of the sentence, and solve the ambiguity problem of the trigger caused by context information. The calculation formula is as follows:

是全局注意力向量α_d中第k个部分，

是事件类型嵌入向量转置，d^T是文档级嵌入向量转置。如图2中

用于辅助全局注意力，避免局部注意力造成的歧义问题。where h _k is the kth part of the output vector h,

is the kth part of the global attention vector _αd ,

is the transposed event type embedding vector, and ^dT is the transposed document level embedding vector.

Used to assist global attention and avoid ambiguity caused by local attention.

(4.4) Weighted fusion of local attention and global attention improves event detection accuracy. The formula is as follows:

v _s = α _s · t ₁

v _d = α _d · t ₂

o＝σ(λ·v _s +(1-λ)·v _d )

The final output value o consists of two parts: _vs and _vd . _vs is generated by the _dot product _{of αs} and t1, which is used to capture local features and simulate hidden event triggers; _vd is generated by the dot product of _αd and _t2 , which is used to capture global features and contextual information. σ is the Sigmoid function, and λ∈[0,1] is a hyperparameter that weighs between _vs and _vd .

Step (5) uses the Focal loss function to avoid the imbalance problem of positive and negative samples and achieve classification of all documents

The data set is processed in units of sentences, and the training data is composed of <sentence, event type> pairs, which represent whether a given sentence conveys an event of type t. Its label is 1 or 0. For example, the label of the training pair <inspection of the bank of the reservoir near the dam, the slope of the hub area and the road: no abnormality, daily inspection> is 1, and the label of the training pair <inspection of the bank of the reservoir near the dam, the slope of the hub area and the road: no abnormality, earthquake> is 0. Since the number of events that a single sentence may express is very small compared with the predefined number of events, the Focal loss function is introduced to address the imbalance problem that the number of negative samples is much larger than the number of positive samples caused by binary classification recognition, and to enhance the influence of positive samples and difficult samples on the model. The calculation formula is as follows:

Where x is composed of sentences and target event types, y∈{0,1}, o(x ⁽ⁱ⁾ ) is the model prediction value, ||θ|| ² is the sum of squares of each element in the model, δ>0 is the weight of the L2 normalization term, β is the parameter for balancing the positive and negative weight ratios of samples, and γ is the parameter for balancing the weight ratios of difficult and easy classification samples. In this experiment, β=0.25 and γ=2 are set;

Finally, the trained model is used to perform event detection on the dam operation log files and classify the documents based on the event types contained in each document.

An emergency condition event detection system based on dam operation logs performs event detection on dam operation logs in the water conservancy field, including:

Log file preprocessing module: first, sort and split the dam operation log according to the record date, label each document, sort, label and segment the sentences in each document, label each word with entity type and part of speech, and then build a set of dam emergency condition event types;

Encoded vector embedding module: Use the ALBERT preprocessing model to encode all the words in the dam operation log and convert them into embedding vectors corresponding to the words;

BiLSTM feature fusion module: Use BiLSTM to fuse the embedding vector, named entity type and part-of-speech tagging vector corresponding to the word segmentation to enhance the semantic information of the word segmentation;

Dual attention mechanism semantic enhancement module: uses sentence-document dual attention to fuse contextual information;

A model: Use the focal loss function to train the above model, and use the trained model to classify all document attribution events.

Obviously, those skilled in the art should understand that the various steps of the emergency condition event detection method based on the dam operation log or the various modules of the emergency condition event detection system based on the dam operation log of the above-mentioned embodiment of the present invention can be implemented by a general computing device, they can be concentrated on a single computing device, or distributed on a network composed of multiple computing devices, optionally, they can be implemented by a program code executable by the computing device, so that they can be stored in a storage device and executed by the computing device, and in some cases, the steps shown or described can be executed in a different order from that here, or they can be made into individual integrated circuit modules, or multiple modules or steps therein can be made into a single integrated circuit module for implementation. In this way, the embodiments of the present invention are not limited to any specific combination of hardware and software.

Claims (5)

1. The emergency working condition event detection method based on the dam operation log is used for carrying out event detection on the dam operation log in the water conservancy field and is characterized by comprising the following steps:

(1) Preprocessing a log file: firstly, sorting and splitting dam operation logs according to the recording date, marking each document, sorting sentences in each document, marking and word segmentation, marking entity types and parts of speech of each word, and then constructing a dam emergency working condition event type set;

(2) Encoding vector embedding: encoding all the word fragments in the dam operation log by using an ALBERT preprocessing model, and converting the encoded word fragments into embedded vectors corresponding to the words; coding all the segmented words by using an ALBERT pre-training model, and converting the coded words into vector representations which can be processed by a computer;

(3) BiLSTM feature fusion: using BiLSTM to fuse an embedded vector, a named entity type and a part-of-speech labeling vector corresponding to the word segmentation, and reinforcing semantic information of the word segmentation;

(4) Dual attention mechanism semantic enhancement: context information is fused using sentence-document dual attention;

(5) Training a model by using the Focal loss function and realizing classification: the classification of all document attribution events is realized by adopting a Focal loss function training model;

the step (1) comprises the following steps:

firstly, dividing a dam operation log file into a plurality of documents according to log record dates, sequencing and marking each document, sequencing and marking sentences in each document, and using jieba to divide words in word units;

the method comprises the steps of (1.2) marking entity types and parts of speech of a word segmentation result, converting the entity types into low-dimensional vectors by searching an embedded table which is randomly initialized, marking parts of speech of each word by using Stanford CoreNLP, and then converting the parts of speech into the low-dimensional vectors by searching the embedded table;

(1.3) predefining the event types of dam emergency working conditions, including earthquake, heavy rain, flood discharge, pre-flood safety large inspection, comprehensive special inspection, daily maintenance and daily inspection events;

the step (3) comprises the following steps:

(3.1) connecting an embedded vector, an entity type vector and a part-of-speech tagging vector corresponding to each word in series, wherein the embedded vector is the vector obtained in the step (2), the entity type vector is the mathematical vector of the recognition result of the named entity of all the words, and the part-of-speech tagging vector is the mathematical vector of the part-of-speech tagging result of all the words;

(3.2) processing the serial vectors in a single sentence by using the BiLSTM model, wherein each vector is an input, capturing word context information by using the bidirectional LSTM unit, and respectively outputting two hidden states

And->

Synthesizing the two vectors into an output vector +.>

In the step (4): in the training set, the emergency working condition predefined event contained in each sentence is converted into an embedded vector t by searching a randomly initialized embedded table ₁ Converting each document into an embedded vector d by using Dov2 Vec; for all the words in each sentence, calculating the weight of each word in the sentence by using a local attention mechanism, improving the attention weight of the word triggering the target event type and simulating a trigger;

the step (4) comprises the following steps:

(4.1) in the training set, converting the emergency condition predefined event contained in each sentence into an embedded vector t by searching a randomly initialized embedded table ₁ Converting each document into an embedded vector d by using Dov2 Vec;

(4.2) for all the words in each sentence, calculating the weight of each word in the sentence by using a local attention mechanism, improving the attention weight of the word triggering the target event type and simulating the trigger, wherein the calculation formula is as follows:

wherein h is _k Is the kth part of the output vector h,

is the local attention vector alpha _s Part k of (a)>

Is a transpose of the event type embedding vector;

(4.3) for all the words in each sentence, calculating the weight of the sentence in which the word is located in the document by using a global attention mechanism, obtaining the unique meaning of the trigger in the scene, assisting in judging the event type of the sentence, solving the ambiguity problem of the trigger due to the context information, and adopting the following calculation formula:

wherein h is _k Is the kth part of the output vector h,

is the global attention vector alpha _d Part k of (a)>

Is event type embedding vector transpose, d ^T Is a document-level embedded vector transpose;

(4.4) weighting and fusing the local attention and the global attention, and improving the event detection precision, wherein the formula is as follows:

v _s ＝α _s ·t ₁

v _d ＝α _d ·t ₂

o＝σ(λ·v _s +(1-λ)·v _d )

wherein the final output value o is defined by v _s And v _d Two parts are formed; v _s From alpha _s And t ₁ Dot product generation for capturing local features and simulating hidden event triggers; v _d From alpha _d And t ₂ Dot product generation for capturing global features and context information; sigma is a Sigmoid function, λ ε [0,1]Is at v _s And v _d Super parameters for trade-off between.

2. The method for detecting emergency condition events based on dam operation log according to claim 1, wherein in the step (5), the data set is processed in sentence units, training data is formed by < sentence, event type > pairs, and the training data represents whether a given sentence conveys an event of t type, the label is 1 or 0, a Focal loss function is introduced, the influence of positive samples and refractory samples on the model is enhanced, and the calculation formula is as follows:

where x is composed of sentences and target event types, y ε {0,1}, o (x ⁽ⁱ⁾ ) Is a model predictive value of the model, and, |θ| ² Is the sum of squares of all elements in the model, delta is more than 0, the weight of the L2 normalization term, beta is a parameter for balancing the positive and negative weight proportion of the sample, and gamma is a parameter for balancing the difficult-to-classify and easy-to-classify weight proportion of the sample;

and finally, performing event detection on the dam operation log file by using the trained model, and classifying the event types contained in each document.

3. An emergency working condition event detection system based on a dam operation log, which carries out event detection on the dam operation log in the water conservancy field, is characterized by comprising:

the log file preprocessing module: firstly, sorting and splitting dam operation logs according to the recording date, marking each document, sorting sentences in each document, marking and word segmentation, marking entity types and parts of speech of each word, and then constructing a dam emergency working condition event type set;

the code vector embedding module: encoding all the word fragments in the dam operation log by using an ALBERT preprocessing model, and converting the encoded word fragments into embedded vectors corresponding to the words; coding all the segmented words by using an ALBERT pre-training model, and converting the coded words into vector representations which can be processed by a computer;

BiLSTM feature fusion module: using BiLSTM to fuse an embedded vector, a named entity type and a part-of-speech labeling vector corresponding to the word segmentation, and reinforcing semantic information of the word segmentation;

a dual-attention mechanism semantic enhancement module: context information is fused using sentence-document dual attention;

one model: training a model by adopting a Focal loss function, and classifying all document attribution events by using the trained model;

the log file preprocessing module is realized as follows:

firstly, dividing a dam operation log file into a plurality of documents according to log record dates, sequencing and marking each document, sequencing and marking sentences in each document, and using jieba to divide words in word units;

the method comprises the steps of (1.2) marking entity types and parts of speech of a word segmentation result, converting the entity types into low-dimensional vectors by searching an embedded table which is randomly initialized, marking parts of speech of each word by using Stanford CoreNLP, and then converting the parts of speech into the low-dimensional vectors by searching the embedded table;

(1.3) predefining the event types of dam emergency working conditions, including earthquake, heavy rain, flood discharge, pre-flood safety large inspection, comprehensive special inspection, daily maintenance and daily inspection events;

the BiLSTM feature fusion module is realized as follows:

(3.1) connecting an embedded vector, an entity type vector and a part-of-speech tagging vector corresponding to each word in series, wherein the embedded vector is the vector obtained in the step (2), the entity type vector is the mathematical vector of the recognition result of the named entity of all the words, and the part-of-speech tagging vector is the mathematical vector of the part-of-speech tagging result of all the words;

(3.2) processing the serial vectors in a single sentence by using the BiLSTM model, wherein each vector is an input, capturing word context information by using the bidirectional LSTM unit, and respectively outputting two hidden states

And->

Synthesizing the two vectors into an output vector +.>

In the dual-attention mechanism semantic enhancement module: in the training set, the emergency working condition predefined event contained in each sentence is converted into an embedded vector t by searching a randomly initialized embedded table ₁ Converting each document into an embedded vector d by using Dov2 Vec; for all the words in each sentence, calculating the weight of each word in the sentence by using a local attention mechanism, improving the attention weight of the word triggering the target event type and simulating a trigger;

the dual-attention mechanism semantic enhancement module converts emergency working condition predefined events contained in each sentence into an embedded vector t by searching a randomly initialized embedded table ₁ Converting each document into an embedded vector d by using Dov2 Vec; for all the words in each sentence, calculating the weight of each word in the sentence by using a local attention mechanism, improving the attention weight of the word triggering the target event type and simulating a trigger; the method comprises the following steps:

(4.1) in the training set, searching the emergency working condition predefined event contained in each sentenceConversion of a randomly initialized embedding table into an embedding vector t ₁ Converting each document into an embedded vector d by using Dov2 Vec;

(4.2) for all the words in each sentence, calculating the weight of each word in the sentence by using a local attention mechanism, improving the attention weight of the word triggering the target event type and simulating the trigger, wherein the calculation formula is as follows:

wherein h is _k Is the kth part of the output vector h,

is the local attention vector alpha _s Part k of (a)>

Is a transpose of the event type embedding vector;

(4.3) for all the words in each sentence, calculating the weight of the sentence in which the word is located in the document by using a global attention mechanism, obtaining the unique meaning of the trigger in the scene, assisting in judging the event type of the sentence, solving the ambiguity problem of the trigger due to the context information, and adopting the following calculation formula:

wherein h is _k Is the kth part of the output vector h,

is the global attention vector alpha _d Part k of (a)>

Is event type embedding vector transpose, d ^T Is a document-level embedded vector transpose;

(4.4) weighting and fusing the local attention and the global attention, and improving the event detection precision, wherein the formula is as follows:

v _s ＝α _s ·t ₁

v _d ＝α _d ·t ₂

o＝σ(λ·v _s +(1-λ)·v _d )

wherein the final output value o is defined by v _s And v _d Two parts are formed; v _s From alpha _s And t ₁ Dot product generation for capturing local features and simulating hidden event triggers; v _d From alpha _d And t ₂ Dot product generation for capturing global features and context information; sigma is a Sigmoid function, λ ε [0,1]Is at v _s And v _d Super parameters for trade-off between.

4. A computer device, characterized by: the computer device comprising a memory, a processor, and a computer program stored on the memory and executable on the processor, the processor implementing the dam operation log-based emergency condition event detection method according to any one of claims 1-2 when executing the computer program.

5. A computer-readable storage medium, characterized by: the computer readable storage medium stores a computer program for executing the emergency condition event detection method based on the dam operation log according to any one of claims 1 to 2.

Images