An intelligent management system for digital emergency plans based on artificial intelligence

Abstract

The present application provides an artificial intelligence-based digital emergency plan intelligent management system, which relates to the field of data intelligent analysis. It uses artificial intelligence-based natural language analysis and understanding technology to semantically understand the natural language description of the emergency input by the user and the first digital emergency plan extracted from the digital emergency plan database, so as to intelligently judge whether to return the first digital emergency plan as a recommended emergency plan based on the two-way attention feature interaction representation between the emergency semantic understanding coding feature and the first digital emergency plan semantic coding feature. In this way, by automatically analyzing and understanding the deep semantic information of the emergency and the digital emergency plan related information, it is conducive to quickly finding the most appropriate response plan, and ensuring that the recommended emergency plan is more in line with the actual situation, thereby significantly improving the overall efficiency and reliability of emergency management.

Description

Technical Field

[0001] The present application relates to the field of intelligent data analysis, and more specifically, to an artificial intelligence-based digital emergency plan intelligent management system. Background Art

[0002] In today's society, the occurrence of emergencies is uncertain and complex, covering many fields such as natural disasters, accidents, public health incidents and social security incidents. These emergencies often pose serious threats to the safety of life and property, social stability and economic development. For example, a large earthquake may instantly destroy a large number of buildings and cause numerous casualties; sudden public health events, such as infectious diseases, will spread rapidly and pose a huge challenge to public health.

[0003] Traditional emergency management plans have exposed many problems when facing these emergencies. On the one hand, the previous plan retrieval and matching methods mainly rely on manual screening or simple keyword matching, which is inefficient and inaccurate. When faced with a massive digital plan database and complex and diverse emergency descriptions, it is difficult for humans to quickly and accurately find the most appropriate response plan, resulting in extended response time and missed best processing opportunities. On the other hand, due to the ambiguity and diversity of natural language, simple keyword matching cannot understand the deep semantics of event descriptions and plan content, and is prone to missing key information or misjudgment, reducing the accuracy and applicability of plan selection.

[0004] Therefore, an artificial intelligence-based digital emergency plan intelligent management system is desired. Summary of the invention

[0005] In order to solve the above technical problems, the present application is proposed. The embodiment of the present application provides a digital emergency plan intelligent management system based on artificial intelligence.

[0006] According to one aspect of the present application, there is provided an artificial intelligence-based digital emergency plan intelligent management system, which includes: an emergency description module, used to obtain a natural language description of an emergency input by a user; an emergency semantic understanding module, used to perform semantic understanding on the natural language description of the emergency to obtain a semantic understanding coding feature of the emergency; a first digital emergency plan extraction module, used to extract a first digital emergency plan from a digital emergency plan database; a first digital emergency plan semantic understanding module, used to perform semantic understanding on the first digital emergency plan to obtain a semantic coding feature of the first digital emergency plan; an event-plan bidirectional interaction module, used to perform bidirectional attention feature matching interaction on the semantic coding feature of the first digital emergency plan and the semantic understanding coding feature of the emergency to obtain a bidirectional attention feature matching interaction between the event-digital emergency plan and the semantic understanding coding feature of the emergency To semantic matching interaction features, wherein the event-plan bidirectional interaction module includes: a homography transformation unit, used to perform homography projection transformation on the first digital plan semantic coding features and the emergency semantic understanding coding features to obtain the first digital plan semantic homography projection coding features and the emergency semantic homography projection coding features; a bidirectional semantic matching unit, used to perform bidirectional attention modulation interaction on the first digital plan semantic homography projection coding features and the emergency semantic homography projection coding features to obtain the emergency-digital plan bidirectional semantic matching interaction features; a plan management module, used to obtain an intelligent management result indicating whether to return the first digital plan as a recommended plan based on the emergency-digital plan bidirectional semantic matching interaction features.

[0007] Compared with the prior art, the digital emergency plan intelligent management system based on artificial intelligence provided by the present application uses natural language analysis and understanding technology based on artificial intelligence to semantically understand the natural language description of the emergency input by the user and the first digital emergency plan extracted from the digital emergency plan database, so as to intelligently judge whether to return the first digital emergency plan as a recommended emergency plan based on the two-way attention feature interaction representation between the emergency semantic understanding coding feature and the first digital emergency plan semantic coding feature. In this way, by automatically analyzing and understanding the deep semantic information of the emergency and the digital emergency plan related information, it is conducive to quickly finding the most appropriate response plan and ensuring that the recommended emergency plan is more in line with the actual situation, thereby significantly improving the overall efficiency and reliability of emergency management. BRIEF DESCRIPTION OF THE DRAWINGS

[0008] In order to more clearly illustrate the embodiments of the present application or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present application. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0009] In the attached picture:Figure 1 This is a system block diagram of an artificial intelligence-based digital emergency plan intelligent management system according to an embodiment of the present application.

[0010] Figure 2 This is a data flow diagram of the artificial intelligence-based digital emergency plan intelligent management system according to an embodiment of the present application.

[0011] Figure 3 It is a block diagram of the event-plan bidirectional interaction module in the artificial intelligence-based digital plan intelligent management system according to an embodiment of the present application.

[0012] Figure 4 It is a block diagram of a bidirectional semantic matching unit in an artificial intelligence-based digital emergency plan intelligent management system according to an embodiment of the present application. DETAILED DESCRIPTION

[0013] Below, the exemplary embodiments according to the present application will be described in detail with reference to the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present application, rather than all the embodiments of the present application, and it should be understood that the present application is not limited to the exemplary embodiments described here.

[0014] In modern society, emergencies are often accompanied by uncertainty and complexity, and their impacts are wide-ranging, including but not limited to natural disasters, industrial accidents, public health crises, and social security issues. These unforeseen events pose severe challenges to human life and property, social stability, and economic prosperity. For example, a strong earthquake can destroy infrastructure in an instant and cause heavy casualties; and sudden public health crises, such as epidemic outbreaks, can spread rapidly and pose a serious threat to public health.

[0015] Traditional emergency response plans have exposed significant shortcomings when dealing with such emergencies. On the one hand, the plan search method that relied on manual screening or basic keyword matching in the past has defects in efficiency and accuracy. When faced with a huge digital plan library and complex descriptions of emergencies, this traditional method is difficult to quickly and accurately locate the most appropriate emergency plan, thereby prolonging the response time and missing the best disposal window. On the other hand, due to the ambiguity and ambiguity of natural language itself, simple keyword searches cannot deeply understand the deeper meaning of event descriptions and plan texts, which may lead to the omission or misjudgment of key information, thereby weakening the effectiveness and pertinence of plan selection.

[0016] In response to the above technical problems, the technical concept of the present application is to obtain the natural language description of the emergency event input by the user and extract the first digital plan from the digital plan database, and use the natural language analysis and understanding technology based on artificial intelligence to semantically understand the natural language description of the emergency event and the first digital plan, so as to intelligently judge whether to return the first digital plan as a recommended plan based on the two-way attention feature interaction representation between the semantic understanding coding features of the emergency event and the semantic coding features of the first digital plan. The present application can automatically parse the emergency event description input by the user, and quickly extract the most relevant plans from the massive plan database without manual screening one by one, which greatly shortens the decision-making time and makes the emergency response more timely. At the same time, it can capture the deep semantic information in the event description and the plan text, which goes beyond simple keyword matching, ensuring that the recommended plan is more in line with the actual situation, thereby significantly improving the overall efficiency and reliability of emergency management.

[0017] Figure 1 This is a system block diagram of an artificial intelligence-based digital emergency plan intelligent management system according to an embodiment of the present application. Figure 2 Schematic diagram of data flow of the digital emergency plan intelligent management system based on artificial intelligence according to an embodiment of the present application. Figure 1 and Figure 2 As shown, in the digital plan intelligent management system 100 based on artificial intelligence, it includes: an emergency description module 110, which is used to obtain the natural language description of the emergency input by the user; an emergency semantic understanding module 120, which is used to perform semantic understanding on the natural language description of the emergency to obtain the semantic understanding coding features of the emergency; a first digital plan extraction module 130, which is used to extract the first digital plan from the digital plan database; a first digital plan semantic understanding module 140, which is used to perform semantic understanding on the first digital plan to obtain the semantic coding features of the first digital plan; an event-plan bidirectional interaction module 150, which is used to perform bidirectional attention feature matching interaction on the semantic coding features of the first digital plan and the semantic understanding coding features of the emergency to obtain the bidirectional semantic matching interaction features of the emergency-digital plan; a plan management module 160, which is used to obtain an intelligent management result indicating whether to return the first digital plan as a recommended plan based on the bidirectional semantic matching interaction features of the emergency-digital plan.

[0018] In an embodiment of the present application, the emergency description module 110 is used to obtain a natural language description of an emergency input by a user. It should be understood that the natural language description of the emergency input by the user will include the type information of the emergency, such as a natural disaster, an accident disaster, a public health event, or a social security event, etc., and will also cover the specific details of the event, including the time, location, scale, scope of influence, development trend, etc. of the event. In addition, the user input may also involve factors affected by the event, such as information on personnel, property, environment, infrastructure, etc. At the same time, it may also include factors related to the event, such as the cause of the event or other related circumstances. By performing semantic analysis on the natural language description of the emergency input by the user, it can help the model understand the key feature information of the event, and then ensure that a suitable digital plan is recommended, so that emergency response can be carried out more accurately and efficiently.

[0019] Obtaining the natural language description of the emergency input by the user can provide a data source for the subsequent emergency data analysis. The following is a detailed description of the data acquisition implementation process: As the core hub for emergency response, the emergency command center's information entry terminal undertakes the key information collection task. In the operation room of the command center, the staff quickly took their positions in front of the terminal at the first time of the emergency, and began to enter the event information in detail with their professional qualities and understanding of the on-site situation. For example, in the case of urban waterlogging caused by heavy rain, the staff will type in the text input box: "In the central urban area of ​​XX City, heavy rain began to fall at XX time, and the rainfall reached XX mm in a short time, causing the urban drainage system to be overwhelmed. Serious waterlogging occurred in many low-lying areas, such as XX Square, XX Street and other areas. The depth of waterlogging has exceeded half a meter. A large number of vehicles have been soaked, traffic has been paralyzed, and some residents' homes have been flooded. Elderly and children are trapped, and the water is still rising, posing a great threat to the safety of urban infrastructure and citizens' lives and property." At the same time, in order to improve the efficiency of information entry, especially in urgent situations and when every second counts, voice input devices play an important role. Staff only need to speak the details of the incident clearly and accurately into the microphone, and the device's built-in advanced speech recognition technology will quickly convert speech into text.

[0020] The emergence of mobile emergency applications (APPs) has greatly expanded the convenience and breadth of information acquisition. Both frontline rescuers and ordinary citizens can easily get started. For example, when a forest fire occurs in the wild, after discovering the fire, the forest ranger quickly takes out his mobile phone to open the emergency management APP and clicks to enter the report page. First, quickly select the event type as "forest fire" through the drop-down menu, and then add specific information in detail in the text box: "In the XX area of ​​XX Mountains, an open fire was discovered at about XX time. The fire spread rapidly with the help of wind. The surrounding vegetation is dense, mainly pine trees and shrubs. The terrain is complex, with steep slopes and canyons, which brings great difficulties to fire fighting. At present, some nearby forest rangers have been organized to carry out preliminary firefighting, but the fire has not been effectively controlled. Professional fire brigades and helicopters are urgently needed to support firefighting operations." Moreover, the functional design of the APP is very user-friendly, taking full account of the actual situation on the scene. It not only supports voice input, but also allows users to take photos and upload them. For example, after an earthquake disaster, citizens in a safe area can take photos or short videos of damaged buildings at the earthquake site. These image materials can intuitively reflect the extent of the earthquake damage. The APP system uses advanced image recognition and analysis technology to automatically extract key information from photos and videos, such as the collapse of buildings and the damage to roads, and converts it into natural language descriptions. After organically integrating it with the information manually entered by the user, it is sent to the digital emergency plan intelligent management system in a timely manner. In this way, no matter where the user is, as long as there is a network signal, it can provide the system with valuable emergency information, ensuring that the system can fully and accurately understand the full picture of the incident.

[0021] In short, through the information input terminal of the emergency command center, mobile emergency applications, etc., the natural language description information of the emergency event input by the user can be obtained comprehensively and efficiently.

[0022] In an embodiment of the present application, the emergency semantic understanding module 120 is used to perform semantic understanding on the natural language description of the emergency to obtain the semantic understanding coding features of the emergency. Specifically, in an embodiment of the present application, the emergency semantic understanding module is used to: use a semantic understanding encoder based on a BERT-bidirectional LSTM hybrid model to perform semantic understanding on the natural language description of the emergency to obtain an emergency semantic understanding coding vector as the emergency semantic understanding coding feature. It should be understood that the natural language description of the emergency contains rich semantic information, and there is interrelated context information between each context. Therefore, in order to more clearly understand and analyze the specific semantic meaning contained in the description input by the user, the present application performs semantic understanding on the natural language description of the emergency to take into account the context information in the text, so that the system can better understand the implicit intention and associated information in the description, and obtain the emergency semantic understanding coding vector. In particular, in a specific embodiment of the present application, a semantic understanding encoder based on a BERT-bidirectional LSTM hybrid model is used to perform semantic understanding on the natural language description of the emergency to obtain an emergency semantic understanding coding vector. Those skilled in the art should know that BERT is a pre-trained language model based on Transformer, which is mainly composed of the encoding layer of Transformer, and the encoding layer is based on the self-attention structure. This structure enables BERT to parallelize training, greatly improve training efficiency, and effectively capture the global information of the text sequence. For example, when processing a long document, BERT can quickly grasp the semantics of the entire document as a whole. Although BERT performs well in parallelizing training and capturing global information of text sequences, its encoding layer based on the self-attention structure is difficult to obtain the order and grammatical relationship between words in the text sequence. Although position embedding can mark the position and order of words to a certain extent, the effect is still not ideal. LSTM is a special recurrent neural network (RNN) that can solve the gradient vanishing and gradient exploding problems of RNN when processing long sequences. On this basis, the bidirectional LSTM processes the sequence from front to back and from back to front, so as to better obtain the contextual information of each word, and is particularly good at capturing the order and grammatical relationship between words. Therefore, in this hybrid model, BERT is usually used to perform preliminary processing on the input text to extract the global features and semantic representation of the text. Then, the features output by BERT are used as the input of the bidirectional LSTM, which further processes these features to mine the contextual details and sequential information at the word level, thereby generating a more representative and discriminative emergency semantic understanding encoding vector to better reflect the various semantic information in the description and provide a reliable basis for subsequent intelligent management system decisions.

[0023] In an embodiment of the present application, the first digital plan extraction module 130 is used to extract the first digital plan from the digital plan database. It should be understood that the first digital plan includes comprehensive and specific emergency response measures and guidance to deal with specific types of emergencies. This information generally covers but is not limited to the use of event types, triggering condition levels, action steps, resource requirements, event risk assessments, and subsequent disposal measures. By performing semantic analysis on the first digital plan, the model can be helped to deeply understand the digital plan, and then match it with the natural language description features of the emergency event input by the user. If the match is successful, the first digital plan is returned as a recommended plan, which is conducive to ensuring the accuracy and applicability of the recommended plan, thereby improving the effectiveness of the emergency response.

[0024] In an embodiment of the present application, the first digital plan semantic understanding module 140 is used to perform semantic understanding on the first digital plan to obtain the semantic coding features of the first digital plan. Specifically, in an embodiment of the present application, the first digital plan semantic understanding module is used to: use the semantic understanding encoder based on the BERT-bidirectional LSTM hybrid model to perform semantic understanding on the first digital plan to obtain the semantic coding vector of the first digital plan as the semantic coding feature of the first digital plan. Similarly, considering that the first digital plan is usually stored in the database in the form of text, which also contains deep semantic information and hidden meanings, therefore, in the technical solution of the present application, the first digital plan is semantically understood to deeply analyze the text content of the first digital plan, capture the core information of the plan, response steps, resource allocation and other details, and obtain the semantic coding vector of the first digital plan. In particular, in a specific embodiment of the present application, the semantic understanding encoder based on the BERT-bidirectional LSTM hybrid model is used to perform semantic understanding on the first digital plan to capture the global semantic information of the plan text, and comb the contextual dependencies between words in the text to obtain the semantic coding vector of the first digital plan.

[0025] The following is a detailed description of the specific implementation process of "using the semantic understanding encoder based on the BERT-bidirectional LSTM hybrid model to perform semantic understanding on the first digital plan to obtain the semantic encoding vector of the first digital plan": First, in the data preparation stage, the first digital plan text data extracted from the digital plan database needs to be carefully preprocessed. These digital plans often have different sources and formats. They may be entered in the form of traditional documents in the early stage, or they may be migrated and integrated between different systems. The model first needs to comprehensively sort them out to remove irregular formats, incorrect encodings and other problems. In this process, special characters in the text, such as some non-standard symbols or garbled characters, will be cleaned and corrected. For example, some plans may contain strange characters caused by data transmission errors, which need to be identified and replaced with appropriate content. At the same time, punctuation marks in the text will be properly processed according to their role in semantic expression. For example, some continuous punctuation marks that do not affect semantic understanding, such as multiple consecutive commas or periods, may be streamlined to reduce data redundancy. However, some punctuation marks with important semantic segmentation functions, such as colons and semicolons, will be retained because they play a key role in expressing the steps and classifications of the plan.

[0026] After the initial cleaning, the word segmentation phase begins. This step requires the use of professional natural language processing tools and algorithms to divide the text according to the semantic boundaries of the vocabulary. Taking a fire emergency plan as an example, the sentence "After a fire occurs, the fire alarm system should be activated immediately, personnel should be organized to evacuate to a safe area, and the fire department should be notified quickly to come to rescue" will be accurately divided into vocabulary units such as "fire", "occurrence", "after", "should", "immediately", "activate", "fire alarm system", "organize", "personnel", "evacuate", "to", "safe area", "and", "quickly", "notify", "fire department", "come", and "rescue". Each word becomes the basic unit for subsequent semantic understanding, and their combination and mutual relationship constitute the semantic structure of the entire plan.

[0027] Next, we will enter the processing flow of the BERT model. With its powerful pre-training capabilities and Transformer-based architecture, the BERT model performs in-depth processing on the word sequence after word segmentation. When these words are input into the BERT model, the model will first generate a corresponding word embedding vector for each word. This process is not a simple one-to-one mapping, but fully considers the context of the word in the entire plan text. For example, for the word "fire alarm system", BERT will combine the semantic information of the words "start" and "after the fire" that appear before and after it to generate a unique vector representation rich in contextual semantics. This vector not only contains the semantic characteristics of the "fire alarm system" itself, but also reflects its functional role in the entire fire emergency plan and its relationship with other actions and equipment.

[0028] In the internal multi-layer architecture of the BERT model, the vector representation of words is continuously transferred and updated between different layers. Through the self-attention mechanism, each word can pay attention to the information of other words in the text, thereby capturing long-distance semantic dependencies. For example, when processing the sentence "Organize people to evacuate to a safe area", the vector representation of the word "evacuate" will interact and fuse with the vectors of words such as "organize", "personnel", and "safe area" during the model processing. In this way, the BERT model can fully understand the semantic structure of the entire sentence, grasp the relationship between the action subject, action object, and action target, and then have a deeper understanding of the local semantics of the entire fire emergency plan. This deep capture and fusion of contextual information enables the BERT model to extract rich semantic information when processing complex digital plan texts, laying a solid foundation for subsequent processing.

[0029] The word vector sequence processed by the BERT model will be further input into the bidirectional LSTM layer. With its unique bidirectional processing capability, the bidirectional LSTM scans and analyzes the word vector sequence from two opposite directions. For the forward LSTM, it starts from the beginning of the sequence, processes the word vectors one by one, and passes the previous information to the following word vectors in sequence, while the reverse LSTM starts from the end of the sequence and passes the following information to the previous word vector. For example, for the sentence "When a flood occurs, the first thing to do is to organize the evacuation of people", the forward LSTM will pass the information of "when a flood occurs" to "first organize the evacuation of people", while the reverse LSTM will pass the information of "evacuation of people" in reverse to "when a flood occurs". This can capture the semantic association between words in different positions before and after the text, providing more comprehensive information for semantic understanding.

[0030] During the processing of the bidirectional LSTM, its internal memory units and hidden states are continuously updated according to the input word vector and the state of the previous time step. Taking the phrase "organize personnel evacuation" as an example, when processing the word "evacuation", the bidirectional LSTM will integrate the information of "organization" and "personnel" to update the memory units and hidden states. This update process will take into account the semantic roles of "organization" and "personnel" in the plan and their relationship with the action of "evacuation", so that the semantic representation of the word "evacuation" is richer and more accurate. In this way, the bidirectional LSTM can effectively accumulate and integrate the text sequence information of the entire first digital plan, further improving the depth and accuracy of the understanding of the plan's semantics.

[0031] Finally, after the collaborative processing of BERT and bidirectional LSTM, an encoding vector that can fully represent the semantic information of the first digital plan will be generated. This encoding vector is a highly abstract and generalized version of the entire plan at the semantic level. It combines BERT's fine capture of contextual information and semantic relationships, and bidirectional LSTM's effective processing of sequence information and bidirectional semantic associations. For example, in an earthquake emergency plan, this encoding vector may contain key information such as response measures at different stages after the earthquake, the types and quantities of rescue resources required, and the division of responsibilities of different rescue teams. This information exists in a compact and efficient vector form, providing strong support for the subsequent operations of the digital plan intelligent management system.

[0032] In the embodiment of the present application, the event-plan bidirectional interaction module 150 is used to perform bidirectional attention feature matching interaction on the first digital plan semantic coding feature and the emergency semantic understanding coding feature to obtain the emergency-digital plan bidirectional semantic matching interaction feature. Specifically, Figure 3 FIG. 1 is a block diagram of an event-plan bidirectional interaction module in an artificial intelligence-based digital plan intelligent management system according to an embodiment of the present application. Figure 3 As shown, the event-plan bidirectional interaction module 150 includes: a homography transformation unit 151, which is used to perform a homography projection transformation on the first digital plan semantic coding feature and the emergency event semantic understanding coding feature to obtain the first digital plan semantic homography projection coding feature and the emergency event semantic homography projection coding feature; a bidirectional semantic matching unit 152, which is used to perform a bidirectional attention modulation interaction on the first digital plan semantic homography projection coding feature and the emergency event semantic homography projection coding feature to obtain the emergency event-digital plan bidirectional semantic matching interaction feature.

[0033] It should be understood that the emergency semantic understanding coding features contain key information of emergencies, such as event types, locations, and semantic associations between specific events. The semantic coding features of the first digital plan cover the core elements of the plan, including countermeasures, applicable scenarios, resource requirements, etc. The two are closely related semantically. Based on this, the present application introduces a two-way attention feature matching interaction mechanism to interact the first digital plan semantic coding features and the emergency semantic understanding coding features to obtain emergency-digital plan two-way semantic matching interaction features. In this way, two-way attention can ensure that no important interaction details are missed. For example, in positive attention, the emergency feature can guide the first digital plan feature to focus on the plan content that is closely related to the current event; reverse attention allows the emergency vector to obtain information from the plan vector that may be ignored but is important for handling the event, thereby improving the accuracy of semantic matching.

[0034] Specifically, in the embodiment of the present application, the homography transformation unit is used to: use the first digital plan mapping homography matrix to perform homography projection transformation on the first digital plan semantic coding vector to obtain the first digital plan semantic homography projection coding feature vector as the first digital plan semantic homography projection coding feature, and the process can be expressed as: ;in, is the semantic coding vector of the first digital plan, is the first digital plan mapping homography matrix, is the first digital proposal semantic homography projection encoding feature vector.

[0035] The emergency semantic understanding coding vector is subjected to homography projection transformation using the emergency mapping homography matrix to obtain an emergency semantic homography projection coding feature vector as the emergency semantic homography projection coding feature. The process can be expressed as: ;in, is the semantic understanding encoding vector of the emergency event, is the incident mapping homography matrix, is the semantic homography projection encoding feature vector of the emergency event.

[0036] It should be understood that by using the first digital plan mapping homography matrix and the emergency event mapping homography matrix to perform homography projection transformation on the first digital plan semantic coding vector and the emergency event semantic understanding coding vector respectively, the two original vectors can be mapped to a new feature space. In this new feature space, the relative position relationship between the features in the original vector is maintained, but the features in the original vector can be observed from different perspectives, which is conducive to the model to better understand the key information in the original emergency event and the first digital plan.

[0037] Specifically, Figure 4 FIG. 1 is a block diagram of a bidirectional semantic matching unit in an artificial intelligence-based digital emergency plan intelligent management system according to an embodiment of the present application. Figure 4 As shown, the bidirectional semantic matching unit 152 includes: a bidirectional attention balance field calculation subunit 1521, which is used to calculate the positive and negative first digital plan-emergency semantic bidirectional attention balance field between the first digital plan semantic homography projection coding feature vector and the emergency semantic homography projection coding feature vector; a modulation response subunit 1522, which is used to modulate and respond to the first digital plan semantic homography projection coding feature vector and the emergency semantic homography projection coding feature vector based on the positive and negative first digital plan-emergency semantic bidirectional attention balance field to obtain the emergency-digital plan bidirectional semantic matching interaction coding vector as the emergency-digital plan bidirectional semantic matching interaction feature.

[0038] More specifically, in an embodiment of the present application, the bidirectional attention balance field calculation subunit is used to calculate the forward first digital plan-emergency attention score field of the first digital plan semantic homography projection encoding feature vector relative to the emergency semantic homography projection encoding feature vector. The process can be expressed as: ;in, is the first digital proposal semantic homography projection encoding feature vector, yes The transposed vector of represents matrix multiplication, yes Length, It is the first digital plan for the positive - emergency attention scoring field.

[0039] The reverse first digital plan-emergency attention score field of the semantic homography projection coding feature vector of the emergency event relative to the semantic homography projection coding feature vector of the first digital plan is calculated. The process can be expressed as: ;in, is the semantic homography projection encoding feature vector of the emergency event, yes The transposed vector of represents matrix multiplication, yes Length, It is the reverse first digital plan-emergency attention scoring field.

[0040] The positive and negative first digital plan-emergency semantic bidirectional attention balance field is constructed between the positive first digital plan-emergency attention score field and the reverse first digital plan-emergency attention score field. The process can be expressed as: ;in, It is the first digital plan for the positive direction - the attention score field for emergencies, It is the reverse first digital plan-emergency attention score field, is the feature concatenation operation, is a convolutional code with a convolution kernel of 3×3. It is the first digital plan for positive and negative events - a semantic two-way attention balance field for emergencies.

[0041] Accordingly, considering that in emergency management, the semantic homography projection coding feature vector of the first digital plan contains a large amount of detailed information on the response strategy, resource allocation, action process, etc. of the emergency, while the semantic homography projection coding feature vector of the emergency describes various characteristics of the specific event, such as event type, scale, scope of influence, etc. This information is not only large in quantity, but also complex in structure and diverse in dimensions. In order to sort out the key connection between the two from the massive feature information and find the digital plan content closely related to the current emergency, it is necessary to calculate the first digital plan semantic homography projection coding feature vector relative to the semantic homography projection coding feature vector of the emergency in the present application. The forward first digital plan-emergency attention score field. That is, the calculated forward attention score field can clearly identify the key information part closely related to the semantic homography projection coding feature vector of the first digital plan semantic homography projection coding feature vector. And the attention score field calculation process is like a filter, which can effectively filter out the noise information that is not related to the emergency in the semantic homography projection coding feature vector of the first digital plan semantic homography projection coding feature vector. After removing these noise information, the information processed by the model is more concise and accurate, and can avoid making wrong judgments due to interference from irrelevant information.

[0042] It should be understood that the information contained in the first digital plan and the emergency is rich and complex. The forward first digital plan-emergency attention score field mainly focuses on the association with the emergency from the perspective of the digital plan, but it is not comprehensive to analyze it only from this direction. In order to enable the model to also start from the characteristics of the emergency and examine its influence on the digital plan characteristics, it is necessary to calculate the reverse first digital plan-emergency attention score field of the semantic homography projection coding feature vector of the emergency relative to the first digital plan semantic homography projection coding feature vector in this application. That is, the process of calculating the reverse first digital plan-emergency attention score field provides an opportunity to understand the relationship between the features in the reverse direction, ensuring that important interaction details will not be missed. In fact, the forward first digital plan-emergency attention score field and the reverse first digital plan-emergency attention score field work together to form a more complete and detailed relationship map, which helps to capture the asymmetric relationship in the original first digital plan and the emergency description.

[0043] Next, considering that the positive first digital plan-emergency attention score field mainly reflects the importance distribution of the first digital plan relative to the emergency, the reverse first digital plan-emergency attention score field shows the importance distribution of the emergency relative to the first digital plan from the opposite direction. Both contain key information from different perspectives, and relying only on the attention score field in a single direction cannot fully understand the complex relationship between the digital plan and the emergency. Based on this, it is necessary to construct the positive and negative first digital plan-emergency semantic bidirectional attention balance field between the positive first digital plan-emergency attention score field and the reverse first digital plan-emergency attention score field in this application. By constructing the positive and negative first digital plan-emergency semantic bidirectional attention balance field, the information in the two directions can be effectively integrated to obtain a complete map about the relationship between the two.

[0044] More specifically, in an embodiment of the present application, the modulation response subunit is used to: map the first digital plan semantic homography projection coding feature vector and the emergency semantic homography projection coding feature vector to the positive and negative first digital plan-emergency semantic bidirectional attention balance field respectively to obtain the first digital plan semantic homography projection attention modulation coding vector and the emergency semantic homography projection attention modulation coding vector, and the process can be expressed as: ;in, is the first digital proposal semantic homography projection encoding feature vector, is the semantic homography projection encoding feature vector of the emergency event, represents matrix multiplication, It is the first digital plan for positive and negative events - a semantic two-way attention balance field for emergencies. is the first digital proposal semantic homography projection attention modulation encoding vector, is the attention modulation encoding vector of the semantic homography projection of the sudden event.

[0045] The process of calculating the position point-by-position division between the first digital plan semantic homography projection attention modulation coding vector and the emergency semantic homography projection attention modulation coding vector to obtain the emergency-digital plan bidirectional semantic matching interaction coding vector can be expressed as: ;in, is the first digital proposal semantic homography projection attention modulation encoding vector, is the incident semantic homography projection attention modulation encoding vector, It is the emergency event-digital emergency plan bidirectional semantic matching interactive coding vector.

[0046] Then, the first digital plan semantic homography projection coding feature vector and the emergency semantic homography projection coding feature vector are respectively mapped to the positive and negative first digital plan-emergency semantic bidirectional attention balance field to obtain the first digital plan semantic homography projection attention modulation coding vector and the emergency semantic homography projection attention modulation coding vector. It should be understood that when the two semantic homography projection coding feature vectors are respectively mapped to the positive and negative first digital plan-emergency semantic bidirectional attention balance field constructed above, a reweighting operation is actually performed. The two modulated coding vectors not only contain the information of the original features, but also incorporate the influence of the features from the other party. That is, the features of one party will adjust their expression according to the features of the other party. In this way, the model can understand and represent the complex interaction between features at a higher level of abstraction, ensuring that each vector adjusts itself according to the guidance provided by the balance field, thereby generating two new coding vectors-the first digital plan semantic homography projection attention modulation coding vector and the emergency semantic homography projection attention modulation coding vector.

[0047] Finally, the position point division between the first digital plan semantic homography projection attention modulation coding vector and the emergency semantic homography projection attention modulation coding vector is calculated to obtain the emergency-digital plan bidirectional semantic matching interaction coding vector. It should be understood that the digital plan semantic homography projection attention modulation coding vector and the emergency semantic homography projection attention modulation coding vector have been processed in the previous steps and have integrated rich information, but the relationship between them still needs to be further explored. A simple comparison method is difficult to deeply reveal the intrinsic connection between the two. The position point division operation provides a mechanism for in-depth comparison, which can analyze the relationship between the two vectors at each position of the vector, thereby digging out the subtle proportional relationship between the features. The emergency-digital plan bidirectional semantic matching interaction coding vector generated by the point division operation can integrate the feature information of the two vectors and the proportional relationship between them, which can provide a new perspective for subsequent analysis and decision-making. This representation method can enable the model to more keenly capture the subtle differences and connections between digital plans and emergencies, and help make more accurate judgments in complex emergency scenarios.

[0048] In an embodiment of the present application, the plan management module 160 is used to obtain an intelligent management result for indicating whether to return the first digital plan as a recommended plan based on the emergency event-digital plan bidirectional semantic matching interaction feature. Specifically, in an embodiment of the present application, the plan management module is used to: input the emergency event-digital plan bidirectional semantic matching interaction encoding vector into the classifier-based plan intelligent management module to obtain the intelligent management result, and the intelligent management result is used to indicate whether to return the first digital plan as a recommended plan. More specifically, in an embodiment of the present application, the plan management module is used to: use the fully connected layer of the classifier to fully connect the emergency event-digital plan bidirectional semantic matching interaction coding vector to obtain the emergency event-digital plan bidirectional semantic matching interaction fully connected coding feature vector; input the emergency event-digital plan bidirectional semantic matching interaction fully connected coding feature vector into the Softmax classification function of the classifier to obtain the probability value of the emergency event-digital plan bidirectional semantic matching interaction coding vector belonging to each classification label, and the classification label includes a value for indicating that the first digital plan is returned as a recommended plan and a value for indicating that the first digital plan is not returned as a recommended plan; and the classification label corresponding to the largest of the probability values ​​is determined as the intelligent management result. That is, the bidirectional semantic matching interaction features of the emergency event and the digital plan obtained by bidirectional attention matching interaction using the first digital plan semantic coding vector and the emergency event semantic understanding coding vector are classified and processed, so as to utilize the powerful pattern recognition and classification capabilities of the classifier, and identify and classify the pattern represented by the bidirectional semantic matching interaction coding vector of the emergency event and the digital plan according to the pre-set rules and training data, so as to determine whether to return the first digital plan as a recommended plan. In this way, appropriate plans can be quickly recommended to relevant personnel based on the intelligent management results, and countermeasures can be initiated in a timely manner.

[0049] In particular, when the first digital plan semantic coding vector and the emergency semantic understanding coding vector respectively represent the coded semantic features of the natural language description of the first digital plan and the emergency, when performing an interactive response based on the feature positive and negative attention fields, the insufficient correspondence between the positive and negative attention fields caused by the semantic differences in the source data encoding will cause a sparse associated interactive response of the emergency-digital plan bidirectional semantic matching interactive coding vector, thereby reducing the accuracy of the intelligent management results obtained by inputting the classifier-based plan intelligent management module due to the lack of classification reasoning degree.

[0050] In a preferred example, the emergency event-digital plan bidirectional semantic matching interactive coding vector is input into the plan intelligent management module based on the classifier to obtain the intelligent management result, which includes: based on the feature value of the emergency event-digital plan bidirectional semantic matching interactive coding vector Distance and Distance to get emergency events - digital emergency plan two-way semantic matching interaction - distance matrix and emergency-digital emergency plan bidirectional semantic matching interaction distance matrix .

[0051] The weighted sum of the emergency event-digital plan bidirectional semantic matching interaction distance matrix 1 and the emergency event-digital plan bidirectional semantic matching interaction distance matrix 2 is calculated to obtain the emergency event-digital plan bidirectional semantic matching interaction joint distance matrix. The process can be expressed as: ;in, represents the emergency-digital plan bidirectional semantic matching interaction-distance matrix, represents the emergency-digital emergency plan bidirectional semantic matching interaction distance matrix, It means point multiplication by position. Indicates adding by position point. and represents the weighted hyperparameter, Represents the emergency event-digital emergency plan bidirectional semantic matching interactive joint distance matrix.

[0052] Determine the eigenvalues ​​of the emergency-digital plan bidirectional semantic matching interactive joint distance matrix , and form the emergency-digital plan bidirectional semantic matching interactive joint distance eigenvector .

[0053] The emergency event-digital plan bidirectional semantic matching interaction coding vector as a row vector is matrix-multiplied with the emergency event-digital plan bidirectional semantic matching interaction-distance matrix to obtain the emergency event-digital plan bidirectional semantic matching interaction-distance query vector. The process can be expressed as: ;in, represents the emergency-digital plan bidirectional semantic matching interaction-distance matrix, represents matrix multiplication, represents the emergency event-digital emergency plan bidirectional semantic matching interactive coding vector, and is a row vector, Represents the emergency event-digital emergency plan bidirectional semantic matching interaction-distance query vector.

[0054] The two-distance matrix of the bidirectional semantic matching interaction between the emergency event and the digital plan is multiplied by the self-association matrix of the bidirectional semantic matching interaction coding vector between the emergency event and the digital plan to obtain the two-distance association matrix of the bidirectional semantic matching interaction between the emergency event and the digital plan. The process can be expressed as: ;in, represents the bidirectional semantic matching interaction distance matrix between the emergency event and the digital emergency plan, represents matrix multiplication, represents the emergency event-digital emergency plan bidirectional semantic matching interactive coding vector, and is a row vector, express The transposed vector of It represents the emergency event-digital emergency plan bidirectional semantic matching interaction two-distance association matrix.

[0055] After matrix multiplication of the emergency event-digital plan bidirectional semantic matching interaction distance query vector and the emergency event-digital plan bidirectional semantic matching interaction distance association matrix, further dot multiplication with the emergency event-digital plan bidirectional semantic matching interaction joint distance eigenvector is performed to obtain an optimized emergency event-digital plan bidirectional semantic matching interaction encoding vector. The process can be expressed as: ;in, represents the emergency-digital emergency plan bidirectional semantic matching interaction-distance query vector, represents the emergency-digital plan bidirectional semantic matching interaction distance association matrix, It means point multiplication by position. represents matrix multiplication, represents the emergency-digital plan bidirectional semantic matching interaction joint distance eigenvector, Represents the optimized emergency event-digital plan bidirectional semantic matching interactive coding vector.

[0056] The optimized emergency event-digital plan bidirectional semantic matching interactive coding vector is input into the classifier-based plan intelligent management module to obtain the intelligent management result.

[0057] That is, the first distance matrix and the second distance matrix of the emergency event-digital plan bidirectional semantic matching interaction coding vector are used as the fine-grained metric association cluster representation of the emergency event-digital plan bidirectional semantic matching interaction coding vector, and the dynamic planning of the relationship between association clusters of different association clusters is performed on the emergency event-digital plan bidirectional semantic matching interaction coding vector and the self-association representation of the emergency event-digital plan bidirectional semantic matching interaction coding vector respectively to simulate the sparse activation based on neuron clusters of the association system, and the fine-grained predictable sparsity of the emergency event-digital plan bidirectional semantic matching interaction coding vector is coordinated with the eigenvalue representation of the metric association cluster of the first distance matrix and the second distance matrix of the emergency event-digital plan bidirectional semantic matching interaction coding vector, so as to avoid the lack of classification reasoning degree affected by insufficient association caused by sparsity, and improve the accuracy of the intelligent management result obtained by inputting the emergency event-digital plan bidirectional semantic matching interaction coding vector into the classifier-based plan intelligent management module.

[0058] In summary, the digital emergency plan intelligent management system 100 based on artificial intelligence according to the embodiment of the present application is explained, which uses artificial intelligence-based natural language analysis and understanding technology to semantically understand the natural language description of the emergency input by the user and the first digital emergency plan extracted from the digital emergency plan database, so as to intelligently judge whether to return the first digital emergency plan as a recommended emergency plan based on the two-way attention feature interaction representation between the emergency semantic understanding coding feature and the first digital emergency plan semantic coding feature. In this way, by automatically analyzing and understanding the deep semantic information of the emergency and the digital emergency plan related information, it is helpful to quickly find the most appropriate response plan and ensure that the recommended emergency plan is more in line with the actual situation, thereby significantly improving the overall efficiency and reliability of emergency management.

[0059] As described above, the digital emergency plan intelligent management system 100 based on artificial intelligence according to the embodiment of the present application can be implemented in various terminal devices, such as a server for intelligent management of digital emergency plans based on artificial intelligence. In one example, the digital emergency plan intelligent management system 100 based on artificial intelligence according to the embodiment of the present application can be integrated into the terminal device as a software module and / or a hardware module. For example, the digital emergency plan intelligent management system 100 based on artificial intelligence can be a software module in the operating system of the terminal device, or it can be an application developed for the terminal device; of course, the digital emergency plan intelligent management system 100 based on artificial intelligence can also be one of the many hardware modules of the terminal device.

[0060] Alternatively, in another example, the artificial intelligence-based digital emergency plan intelligent management system 100 and the terminal device may also be separate devices, and the artificial intelligence-based digital emergency plan intelligent management system 100 may be connected to the terminal device via a wired and / or wireless network and transmit interactive information in accordance with an agreed data format.

Claims

1. An artificial intelligence-based digital emergency plan intelligent management system, characterized in that: include: An emergency event description module is used to obtain a natural language description of an emergency event input by a user; An emergency semantic understanding module, used for semantically understanding the natural language description of the emergency to obtain the emergency semantic understanding coding features; A first digital plan extraction module is used to extract a first digital plan from a digital plan database; a first digital plan semantic understanding module is used to perform semantic understanding on the first digital plan to obtain the semantic coding features of the first digital plan; an event-plan bidirectional interaction module is used to perform bidirectional attention feature matching interaction on the semantic coding features of the first digital plan and the semantic understanding coding features of the emergency event to obtain the bidirectional semantic matching interaction features of the emergency event-digital plan, wherein the event-plan bidirectional interaction module includes: a homography transformation unit, which is used to perform homography projection transformation on the semantic coding features of the first digital plan and the semantic understanding coding features of the emergency event to obtain the semantic homography projection coding features of the first digital plan and the semantic homography projection coding features of the emergency event; a bidirectional semantic matching unit, which is used to perform bidirectional attention modulation interaction on the semantic homography projection coding features of the first digital plan and the semantic homography projection coding features of the emergency event to obtain the bidirectional semantic matching interaction features of the emergency event-digital plan; a plan management module, which is used to obtain an intelligent management result indicating whether to return the first digital plan as a recommended plan based on the bidirectional semantic matching interaction features of the emergency event-digital plan.

2. The digital emergency plan intelligent management system based on artificial intelligence according to claim 1 is characterized in that: The emergency semantic understanding module is used to: use a semantic understanding encoder based on a BERT-bidirectional LSTM hybrid model to perform semantic understanding on the natural language description of the emergency to obtain an emergency semantic understanding encoding vector as the emergency semantic understanding encoding feature.

3. The digital emergency plan intelligent management system based on artificial intelligence according to claim 2 is characterized in that: The first digital plan semantic understanding module is used to: use the semantic understanding encoder based on the BERT-bidirectional LSTM hybrid model to perform semantic understanding on the first digital plan to obtain a first digital plan semantic coding vector as the first digital plan semantic coding feature.

4. The digital emergency plan intelligent management system based on artificial intelligence according to claim 3 is characterized in that: The homography transformation unit is used to: use the first digital plan mapping homography matrix to perform a homography projection transformation on the first digital plan semantic coding vector to obtain a first digital plan semantic homography projection coding feature vector as the first digital plan semantic homography projection coding feature; use the emergency event mapping homography matrix to perform a homography projection transformation on the emergency event semantic understanding coding vector to obtain an emergency event semantic homography projection coding feature vector as the emergency event semantic homography projection coding feature.

5. The artificial intelligence-based digital emergency plan intelligent management system according to claim 4 is characterized in that: The bidirectional semantic matching unit includes: a bidirectional attention balance field calculation subunit, which is used to calculate the positive and negative first digital plan-emergency semantic bidirectional attention balance field between the first digital plan semantic homography projection coding feature vector and the emergency semantic homography projection coding feature vector; a modulation response subunit, which is used to modulate and respond to the first digital plan semantic homography projection coding feature vector and the emergency semantic homography projection coding feature vector based on the positive and negative first digital plan-emergency semantic bidirectional attention balance field to obtain an emergency-digital plan bidirectional semantic matching interaction coding vector as the emergency-digital plan bidirectional semantic matching interaction feature.

6. The digital emergency plan intelligent management system based on artificial intelligence according to claim 5 is characterized in that: The bidirectional attention balance field calculation subunit is used to: calculate the forward first digital plan-emergency attention score field of the first digital plan semantic homography projection coding feature vector relative to the emergency semantic homography projection coding feature vector; calculate the reverse first digital plan-emergency attention score field of the emergency semantic homography projection coding feature vector relative to the first digital plan semantic homography projection coding feature vector; construct the forward and reverse first digital plan-emergency semantic bidirectional attention balance field between the forward first digital plan-emergency attention score field and the reverse first digital plan-emergency attention score field.

7. The artificial intelligence-based digital emergency plan intelligent management system according to claim 6 is characterized in that: The modulation response subunit is used to: map the first digital plan semantic homography projection coding feature vector and the emergency semantic homography projection coding feature vector to the positive and negative first digital plan-emergency semantic bidirectional attention balance field to obtain the first digital plan semantic homography projection attention modulation coding vector and the emergency semantic homography projection attention modulation coding vector; Calculate the position point-by-position division between the first digital plan semantic homography projection attention modulation coding vector and the emergency event semantic homography projection attention modulation coding vector to obtain the emergency event-digital plan bidirectional semantic matching interaction coding vector.

8. The artificial intelligence-based digital emergency plan intelligent management system according to claim 7 is characterized in that: The plan management module is used to: input the emergency event-digital plan bidirectional semantic matching interactive coding vector into the classifier-based plan intelligent management module to obtain the intelligent management result, and the intelligent management result is used to indicate whether to return the first digital plan as a recommended plan.

9. The artificial intelligence-based digital emergency plan intelligent management system according to claim 8 is characterized in that: The plan management module is used to: use the fully connected layer of the classifier to fully connect the emergency event-digital plan bidirectional semantic matching interaction coding vector to obtain the emergency event-digital plan bidirectional semantic matching interaction fully connected coding feature vector; input the emergency event-digital plan bidirectional semantic matching interaction fully connected coding feature vector into the Softmax classification function of the classifier to obtain the probability value of the emergency event-digital plan bidirectional semantic matching interaction coding vector belonging to each classification label, and the classification label includes a value for indicating that the first digital plan is returned as a recommended plan and a value for indicating that the first digital plan is not returned as a recommended plan; and the classification label corresponding to the largest of the probability values ​​is determined as the intelligent management result.

Images