A fire plan classification method based on deep learning

Abstract

A fire plan classification method based on deep learning focuses on solving the problem of low query accuracy and lack of intelligent classification capabilities for fire plans. Through this fire plan classification method based on deep learning, effective analysis and management of plan information can be achieved, and the disposal plan can be quickly obtained and released in wartime. It has the characteristics of intuitive and vivid, real-time query, rapid release, dynamic analysis, and auxiliary decision-making, which greatly improves the scientificity and efficiency of fire fighting and rescue operations.

Description

Technical Field

[0001] The invention relates to the field of fire emergency plan management, and specifically refers to a fire emergency plan classification method based on deep learning. Background Art

[0002] At present, with the rapid development of my country's economic construction and the acceleration of urbanization, complex buildings such as high-rise, underground, and large complexes have emerged in large numbers, and the height of buildings has continued to rise, the volume has become increasingly large, and the functions have become increasingly complex. Various new materials, new processes, new products, and new technologies are widely used, which has brought about an increasing number of fire hazards and risks, and the difficulty of fire fighting has increased. The digital plan for fire fighting and rescue refers to a fire fighting and rescue action plan formed by using information technology as a means, relying on the information environment, based on the mastery of existing fire fighting and rescue forces and disposal objects, through the simulation analysis and prediction of fire risks and disaster consequences, and the reasonable evaluation and deployment of fire fighting and rescue resources. It can realize the effective storage and management of combat information in peacetime, and can achieve the rapid acquisition and release of disposal plans in wartime. It has the characteristics of intuitive and vivid, real-time query, rapid release, dynamic analysis, and decision-making assistance, which greatly improves the scientificity and efficiency of fire fighting and rescue operations.

[0003] The current plan management and query have the following problems:

[0004] (1) Low query accuracy

[0005] Fire emergency plans are generally named after units or buildings. The plans contain a large amount of information such as basic unit information, disaster setting, force mobilization, organization and command, social linkage, service support, special reminders, decision support, various pictures, etc. In the query, there will be two major problems: inaccurate plan positioning and incomplete information search. Inaccurate plan positioning means that it is difficult to accurately locate the corresponding plan through the unit name. For example: the plan name is stored as "The Second Affiliated Hospital of China Medical University". When searching for "The Second Affiliated Hospital of China Medical University" or "Shenyang Hospital", it is difficult to find the correct plan. Incomplete information search means that only by finding the corresponding plan can we further find the corresponding basic information, disaster setting, force mobilization, organization and command, social linkage, service support and other information. The amount of information that can be located is small and incomplete.

[0006] (2) Lack of intelligent classification capabilities

[0007] In the process of emergency rescue, emergency plans will provide important help for early emergency response. Recommending relevant fire plans based on on-site conditions is a solution to improve the efficiency of emergency rescue. For example, a fire accident occurred in a plant producing olefin chemical products. There is no relevant fire plan in this area. If there is a similar olefin chemical plant plan for reference, it can also serve as an auxiliary decision support. However, the current emergency plan system does not have the ability to intelligently classify and does not make a large amount of plan information into a highly manageable cluster group. Summary of the invention

[0008] The present invention provides a fire prevention plan classification method based on deep learning, which solves the problems of low query accuracy and lack of intelligent classification capability in the prior art.

[0009] In order to achieve the above purpose, the present invention adopts the following technical solutions:

[0010] A fire plan classification method based on deep learning, characterized in that the steps are:

[0011] 1) Fire plan record segmentation:

[0012] 1.1) Add professional vocabulary to the general Chinese dictionary;

[0013] 1.2) Use the sequential fine-grained word segmentation algorithm combined with the Chinese word library to perform preliminary word segmentation on the plan text;

[0014] 1.3) Using the word segmentation exhaustive method, for sentences with ambiguous word segmentations in the text obtained after the preliminary word segmentation, all possible alternative word segmentation schemes are listed;

[0015] 1.4) For sentences that may be controversial, the scores of various alternatives are calculated through the weighting of key words and the statistical principle of segmentation probability, and the solution with a high score is selected as the word segmentation result;

[0016] 2) Fire plan element keyword extraction:

[0017] 2.1) Using unsupervised text keyword extraction technology based on statistical features and topic models to extract keywords from fire emergency plan text;

[0018] 2.2) Quantify the word segmentation weight and segmentation probability statistics through the fire emergency plan professional vocabulary database to calculate the weight value corresponding to each word;

[0019] 2.3) By ranking the keyword weights, finally obtain the N words with the highest weights as the keywords corresponding to the fire emergency plan, and obtain a vector group consisting of N keywords and their corresponding weights;

[0020] 3) Intelligent classification of fire emergency plans under deep learning:

[0021] 3.1) Use the vector groups obtained in step 2.3) of each pre - plan as the first - level related attributes as input, and obtain vector values through convolution operations;

[0022] 3.2) For a given dataset of X fire protection pre - plan texts, given the number of clusters K (K < X), initialize the category to which each sample belongs, and continuously iterate through the convolutional K - means algorithm to re - divide the categories of the dataset. Finally, divide the X fire protection pre - plan texts into K categories.

[0023] In step 1.3), when there are multiple preliminary word - segmentation results for a sentence, it is considered a text that may have ambiguous word - segmentation.

[0024] In step 1.4), the calculation method is as follows:

[0025] Keyword weight = weight in the keyword library + frequency of occurrence; where the weight in the keyword library is manually assigned, and the frequency of occurrence = the frequency of the keyword in the text / the total number of times all words appear in the text.

[0026] Advantages of the present invention:

[0027] Through the above - mentioned method, the present invention uses new generation information technologies such as big data, deep learning, and artificial intelligence to design a fire protection pre - plan classification method based on deep learning, focusing on solving the problems of low query accuracy of fire protection pre - plans and lack of intelligent classification ability. Through the fire protection pre - plan classification method based on deep learning, effective analysis and management of pre - plan information are realized. During wartime, rapid acquisition and release of disposal plans can be achieved, with characteristics such as intuitive vividness, real - time query, rapid release, dynamic analysis, and auxiliary decision - making, greatly improving the scientific nature and efficiency of fire - fighting and rescue operations. Brief Description of the Drawings

[0028] Figure 1 It is a general design diagram of a fire protection pre - plan classification method based on deep learning. Detailed Embodiments

[0029] Next, the technical solutions in the embodiments of the present invention will be clearly and completely described in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all of the embodiments.

[0030] Example 1: Now input 100 fire plan texts, set the number of categories to 5, that is, divide the 100 fire plan texts into 5 categories. Through text segmentation, keyword vector extraction, and convolution processing, 100 input vectors are obtained as the input of the K-means algorithm. After iterative processing of the K-means algorithm, the 100 fire plan texts are automatically divided into 5 categories. The specific method is:

[0031] 1) Fire plan record segmentation:

[0032] 1.1) Add professional vocabulary to the Chinese general dictionary. For example, add special vocabulary for fire plans such as "building conditions, floor area, building area, fire protection facilities, indoor fire hydrants", add special vocabulary for hazardous chemicals such as "gasoline, diesel, olefins", add professional vocabulary for chemical disposal equipment such as "internal floating roof tank, external floating roof tank, alkylation unit", and improve the accuracy of fire plan word segmentation.

[0033] 1.2) Use a sequential fine-grained word segmentation algorithm combined with a Chinese vocabulary to perform preliminary word segmentation on the plan text.

[0034] 1.3) Use the word segmentation exhaustive method to list all possible alternative word segmentation schemes for sentences in the text obtained after preliminary word segmentation that may have ambiguous word segmentations. When there are multiple preliminary word segmentation results in a sentence, it is considered to be a text that may have ambiguous word segmentations. For example: a fire pool may be divided into two words, "fire" and "pool" or one word, "fire pool"; "People's Bank of China" may be divided into three words, "China", "People" and "Bank" or one word, "People's Bank of China".

[0035] 1.4) For sentences that may be controversial, the scores of various alternatives are calculated through the weighting of key words and the statistical principle of segmentation probability, and the high-scoring plan is selected as the word segmentation result. Specifically, the keyword weight = the weight in the keyword library + the frequency of occurrence. The weight in the keyword library is assigned manually, and the frequency of occurrence = the frequency of the entry (keyword) in the text / the total number of times all words appear in the text. For example, "water pump coupling" is assigned a value of 0.5, and "of" is assigned a value of -0.5. Obviously, in the present invention, compared with the word segmentation of "firefighting" and "water pool", the word segmentation of "firefighting water pool" is more accurate, so when setting the assignment, the set assignment of "firefighting water pool" is greater than the assignment of "firefighting" + "water pool". By setting the assignment, the most accurate word segmentation result can be screened out.

[0036] 2) Fire plan element keyword extraction:

[0037] 2.1) Unsupervised text keyword extraction technology based on statistical features and topic models is used to extract keywords from fire plan text.

[0038] 2.2) Quantify the word segmentation weight and segmentation probability statistics through the fire plan professional vocabulary database to calculate the weight value corresponding to each word.

[0039] 2.3) By ranking the keyword weights, we finally obtain the N words with the highest weights as the keywords of the corresponding fire plan, and obtain a vector group consisting of N keywords and their corresponding weights. Generally, N=100, and the number of keywords included in each vector group can also be set according to the specific situation.

[0040] 3) Intelligent classification of fire emergency plans under deep learning:

[0041] 3.1) Take the vector group obtained in step 2.3) of each plan as the first-level related attribute as input, and obtain the vector value through convolution operation. Specifically, the first M words and their keyword weights can be selected from the keywords as the first-level related attributes as input, and M can be 64, which can also be selected according to the specific situation.

[0042] 3.2) For a given data set of 100 fire plan texts, the number of clusters is given as 5, the category to which each sample belongs is initialized, and the convolutional K-means algorithm is used to continuously iterate and reclassify the data set, and finally the 100 fire plan texts are divided into 5 categories.

Claims

1. A fire plan classification method based on deep learning, characterized in that: The steps are as follows: 1) Fire protection plan record word segmentation: 1.1) Add professional vocabulary to the Chinese general dictionary; 1.2) Use the sequential fine-grained word segmentation algorithm combined with the Chinese word library to perform preliminary word segmentation on the plan text; 1.3) Use the word segmentation exhaustion method to list all possible alternative word segmentation schemes for the sentences that may have ambiguous word segmentation in the text obtained after preliminary word segmentation; when there are multiple preliminary word segmentation results for a sentence, it is considered as the text that may have ambiguous word segmentation; 1.4) For the sentences that may have objections, calculate the scores of each alternative scheme through the keyword weighting and segmentation probability statistics principle, and select the scheme with the highest score as the word segmentation result; keyword weight = weight in the keyword library + occurrence frequency; where the weight in the keyword library is manually assigned, and the occurrence frequency = the number of times the keyword appears in the text / the total number of times all words appear in the text; 2) Fire protection plan element keyword extraction: 2.1) Use the unsupervised text keyword extraction technology of statistical features and topic models to extract keywords from the fire protection plan text; 2.2) Quantify the word segmentation weight and segmentation probability statistics through the fire protection plan professional vocabulary library to calculate the weight value corresponding to each word; 2.3) Through the keyword weight ranking, finally obtain the N words with the highest weights as the keywords corresponding to the fire protection plan, and get a vector group composed of N keywords and their corresponding weights; 3) Intelligent classification of fire protection plans under deep learning: 3.1) Take the vector group obtained in step 2.3) of each plan as the first-level related attribute as the input, and obtain the vector value through convolution operation; 3.2) For the given dataset of X fire protection plan texts, given the number of clusters K, K < X, initialize the category to which each sample belongs, and continuously iterate and re-divide the category of the dataset through the convolutional K-means algorithm, and finally divide the X fire protection plan texts into K categories.

Images