CN116757460A - Emergency command scheduling platform construction method and system based on deep learning - Google Patents

Abstract

The invention discloses a method and a system for constructing an emergency command and dispatch platform based on deep learning, wherein the method comprises the following steps: classifying, clustering and abnormality detecting event atom library data, monitoring emergency event nodes by using a knowledge reasoning technology based on graph generation antagonism network GraphGAN, generating emergency triplet information, and constructing an event knowledge graph; the method comprises the steps of using a self-encoder MGA based on a multi-mode diagram, converting a knowledge graph into a multi-mode representation form by encoding and decoding different types of nodes and edges, enhancing the expression capability and reasoning capability of a model, capturing the interdependence relationship of the nodes, and judging an emergency event; the method comprises the steps of introducing multitask learning, adding a bert as a pre-training model method, performing user personalized event emergency plan recommendation, event link scene recommendation and link classification tasks, and performing scoring calculation by distributing different task weights to obtain the event emergency plan recommendation with interpretability.

Description

Emergency command scheduling platform construction method and system based on deep learning

Technical Field

The invention relates to the technical field of emergency management command scheduling, in particular to an emergency command scheduling platform construction method and system based on deep learning.

Background

The emergency command and dispatch platform is used as a command and dispatch platform in sudden or emergency scenes, has important emergency management and emergency response functions, can help a command decision maker to quickly make a correct decision, coordinates emergency resources and rescue actions, reduces disaster loss and casualties to the greatest extent, and thoroughly realizes autonomous and controllable platform and faces a certain challenge due to the complexity of the emergency command and dispatch platform.

At present, in the scene of a new generation of emergency command and dispatch platform for chaotic engineering, the emergency command and dispatch platform in the prior art has the characteristics of low system reliability, poor emergency event processing capability and the like. Therefore, deep learning is applied to the construction of the emergency command and dispatch platform, user-personalized event emergency plan recommendation, event link scene recommendation and event classification tasks are carried out, the emergency event processing capacity and efficiency of the emergency command and dispatch platform are improved, the robustness of a transaction system is improved, and the method has important practical application significance.

Disclosure of Invention

The invention aims to solve the problems that: the emergency command dispatching platform construction method and system based on deep learning are used for improving the emergency event processing capacity and efficiency of the emergency command dispatching platform.

In order to achieve the above purpose, the present invention adopts the following technical scheme:

the emergency command scheduling platform construction method based on deep learning comprises the following steps:

step 1, establishing event atom library data, monitoring emergency nodes by using a knowledge reasoning technology based on graph generation antagonism network GraphGAN, generating emergency triplet information, and constructing an event knowledge graph;

step 2, using a self-encoder MGA based on a multi-mode diagram to encode and decode different types of nodes and edges, converting a knowledge graph into a multi-mode representation form, capturing the inter-dependent relationship of the nodes, and judging an emergency event;

and 3, introducing a multi-task learning model MTL-Att, adding an attention mechanism and combining a bert pre-training model to construct the MTL-Att-bert model, and performing user personalized event emergency plan recommendation, event link scene recommendation and event classification tasks by distributing different task weight scoring calculations.

Specifically, in step 1, knowledge reasoning based on graph generation of an countermeasure network GraphGAN is used to obtain emergency time nodes to expand a fault knowledge graph, the countermeasure network GraphGAN comprises a formative model and a discriminant model, specifically,

generating a model for fitting or estimating the probability of real connection distribution, selecting the most likely nodes from the node set VThe connected nodes, expressed as: />G is a generator;

a discriminant model for discriminating the node v _C And calculates the output node and />The possibility of an edge exists between the two, expressed as: />D is a discriminator;

minmax objective function of generator G and arbiter D, formula (1) below:

（1）

wherein ,is a discriminator for discriminating node pairs +.>Is a connectivity of (2);

outputting a scalar representing the node +.> and />Probability of edge existing between them;

for all nodes->Is a vector of (2); />As a cost function;

data true distribution +_>Representing the generated sample, ++>For +.>Calculating expected values in a true distribution;

to be a generator model +.>Is the input condition of the generator, < >>Is a parameter of the generator and,to>Sample generated by generator G>C is the node index;

the parameters of the generator and the arbiter are updated by continuous alternating training, and the arbiter D passes through the parameters from the generator and the arbiterTraining is performed on the positive samples from generator G, which is updated according to a gradient strategy, guided by the arbiter D.

Further, the arbiter D is a sigmoid function, as shown in the following formula (2):

（2）

wherein ,is node-> and />K-dimensional vector representation in the arbiter, thus +.>Can be regarded as all +.>Is a collection of (3);

for inner product operation and conversion by means of a sigmoid function,/for the inner product operation>Taking the negative number for the inner product and performing the exponential operation, ++>For the discriminator +/for a given real sample>And prediction sample->Is a predicted result of (a);

for a given node pairUpdating the corresponding node expression vector by gradient descent method>The following formula (3):

(3)

wherein ,for->Gradient of->Representing the variable->Is true of distribution of->Representing the generation of samples from a generator>，/>Representing updating of the discriminator parameter with gradient descent +.>。

Further, the generator G optimizes updating by a gradient descent method, and the gradient of the generator is as follows formula (4):

(4)

wherein the gradient isIs weighted +.>Gradient of->Is a desired sum of (2);

generate sample->Under the condition of the generator G, utilizing the expected value of the discriminator;

generate sample->Generating samples from generator G condition>Is a desired value of (2);

the implementation of generator G is defined by a softmax function, as follows equation (5):

(5)

wherein ,, />∈R ^k is node-> and />K-dimensional vector representation, θ, in generator _G All->Is a collection of (3);

for the inner product and doing the exponential operation, +.>Generator G according to the specified conditions->Generate sample->。

Then, based on the above formula (5), a Graph Softmax is proposed, and the estimated connection distribution is calculated by embedding network structure information through the SoftmaxThe method specifically comprises the following steps: BFS wide search is performed on the original network omega, and the original network omega is unfolded to v _c Tree T being the root node _c ，N _C (v) Representing node->Is to assume a given node +.>Neighbor node +.>E Nc (v), definition ∈Nc (v)>Equivalent to->Is defined as the following equation (6):

(6)

wherein each nodeAre all ∈10 by the root node>Starting fromOnly one path arrives, the unique path being defined as: />, wherein />，/>；

In the condition->Lower and sample->A set of related conditions;

graph Softmax definitionThe following formula (7):

(7)

wherein ,for +.>Predicting the next sample->Probability of (2);

for +.>Predicting the next sample->Is a probability of (2).

Further, in step 2, based on the multi-modal graph, the self-encoder MGA performs decoding operations by using the encoder for different types of nodes and edges and the decoder, specifically as follows:

the multi-mode self-encoder is a feedforward neural network and comprises an input layer, a first hiding layer, a second hiding layer, a third hiding layer and an output layer, wherein the input layer is formed by combining image feature vectorsAnd text feature vector +.>The input is obtained in two modalities, the input layer is followed by a first hidden layer, for each modality the mapping input is denoted +.>Andthe method comprises the steps of carrying out a first treatment on the surface of the Will-> and />Connected, joint mapping to a second hidden layer, denoted joint multi-modal entity embedding；

The decoder stage and the encoder have a completely symmetrical structure, mapping the embedded as input to a third hidden layer, denoted as and />And corresponding-> and />Sharing the same dimension; the decoder will and />Mapping to the output layer-> and />The method comprises the steps of carrying out a first treatment on the surface of the The output layer and the input layer of each mode have the same dimension;

the overall architecture of the multi-mode self-encoder is defined as:

(8)

(9)

(10)

(11)

(12)

(13)

(14)

where f is the sigmoid activation function, and />Representing the mapping weight matrix of the modal image and text from layer j to layer (j + 1) respectively,j= 1,2,3,4；/> and />Bias items of image and text in the j-th hidden layer, respectively,>is embedded->The symbol # -represents a join operation;

the multi-mode self-encoder trains with the aim of minimizing reconstruction errorsIs the sum of the dissimilarities between the two modal input and output layers, equation (15) below:

(15)

then, the knowledge graph is converted into a multi-modal representation using a transform model that maps entities and relationships in a low-dimensional continuous vector space, wherein the embedding of the head entity h, the tail entity t, and the relationship r satisfies the following triplet formula (16):

（16）

for triples in the knowledge-graph, the TransE model attempts to minimize the distance between h+r and t, while for triples outside the knowledge-graph, attempts to maximize the distance between them

Minimizing losses during TransE trainingThe following formula (17):

(17)

where γ is the spacing hyper-parameter, max (0, x) is intended to obtain the positive part of x, d is the dissimilarity function,a symbolic representation representing the head, relation and tail entities of the triplet at the time of prediction, S representing the original dataset consisting of triples present in the knowledge-graph and S' representing the corrupted dataset consisting of triples not present in the knowledge-graph; s' is considered a negative sample in training, constructed by replacing the head or tail entity of each triplet as shown in equation (18) below:

(18)

wherein ,Eis an entity set;

constructing a TransAE model, combining the two models of the formula (17) and the formula (18), and simultaneously learning multi-mode knowledge and structural knowledge, wherein each entity in a knowledge graph corresponds to a plurality of pictures and a sentence description, extracting visual and text feature vectors of the knowledge graph, inputting the feature vectors into a multi-mode self-encoder to obtain joint embedding as entity representation, and randomly initializing relation embedding at the beginning of training, wherein the entity and the relation representation are used for training the model; structural lossfor (h, r, t) is expressed as the following formula (19):

(19)

wherein and />Representing the embedded representation of h and t, respectively, < >> and ；/>Respectively representing embedded representation forms of h 'and t';

a regularizer Ω (θ) is added for the parameter set θ, weighted α, training the model by minimizing the total loss L, as shown in equation (20):

（20）

parameters in self-encoder and relational embedding,/>,/>,/>Randomly initialized at the beginning of training, weight parameters β and α are chosen to balance the loss +.>、/>And regular term->Is of the size and importance of (a).

Further, in step 3, a multi-task learning model MTL-Att is introduced, an attention mechanism is added to combine with a bert pre-training model, specifically, in the multi-task learning model, features of each task are extracted, common features are learned, a shared bert model is used as a feature extractor to form an MTL-Att-bert model, in the training process, the MTL-Att-bert model learns a weight vector for each task through the attention mechanism to weight output of each task, and the importance of each task is adjusted through adjusting the value of the weight vector, so that an event emergency plan recommendation result is obtained and is used for user-personalized event emergency plan recommendation, scene recommendation of related event links and link classification tasks.

The invention also provides a system for constructing the emergency command and dispatch platform based on deep learning, which is constructed by using the method for constructing the emergency command and dispatch platform based on any one of the above steps, and comprises the following modules:

the event atom library acquisition module is used for acquiring input data and constructing a data atom library;

the event knowledge graph construction module; the method comprises the steps of generating an countermeasure network GraphGAN knowledge reasoning technology based on a graph, monitoring emergency nodes, generating emergency triplet information and constructing an event knowledge graph;

the multi-modal map self-encoder module: the method comprises the steps of performing encoding and decoding operations on nodes and edges of different types, converting a knowledge graph into a multi-mode representation form, capturing the mutual dependence of the nodes, and performing emergency judgment;

the MTL-Att-bert model building module is used for extracting the characteristics of each task by combining a bert pre-training model through using an attention mechanism;

and the event emergency plan recommending module is used for carrying out user personalized event emergency plan recommending, related event link scene recommending and link classifying tasks.

Compared with the prior art, the technical scheme provided by the invention has the following technical effects:

1: the invention generates enough emergency triplet information based on a knowledge reasoning technology of a graph generation countermeasure network (GraphGAN) to construct an event knowledge graph.

2: the model of the invention uses a self-encoder based on a multi-modal diagram and combines a multi-task learning strategy, thereby improving the robustness and efficiency of the system when an emergency occurs in the scene of an emergency command and dispatch platform of chaotic engineering and making up the defect of the prior art for a new generation of emergency command and dispatch service.

Drawings

FIG. 1 is a flow chart of the emergency command and dispatch platform construction method based on deep learning.

Detailed Description

In order to make the objects, technical solutions and advantages of the present invention more apparent, the technical solutions of the application will be further elaborated in conjunction with the accompanying drawings, and the described embodiments are only a part of the embodiments to which the present invention relates. All non-innovative embodiments in this example by others skilled in the art are intended to be within the scope of the invention.

The invention discloses a method for constructing an emergency command and dispatch platform based on deep learning, which is shown in fig. 1 and comprises the following steps:

step 1, establishing event atom library data, monitoring emergency nodes by using a knowledge reasoning technology based on graph generation antagonism network GraphGAN, generating emergency triplet information, and constructing an event knowledge graph;

step 2, using a self-encoder MGA based on a multi-mode diagram to encode and decode different types of nodes and edges, converting a knowledge graph into a multi-mode representation form, capturing the inter-dependent relationship of the nodes, and judging an emergency event;

and 3, introducing a multi-task learning model MTL-Att, adding an attention mechanism and combining a bert pre-training model to construct the MTL-Att-bert model, and performing user personalized event emergency plan recommendation, event link scene recommendation and event classification tasks by distributing different task weight scoring calculations.

Firstly, in step 1, a knowledge reasoning technology based on graph generation of an countermeasure network GraphGAN is used to obtain emergency time nodes to expand a fault knowledge graph, wherein the countermeasure network GraphGAN comprises a formative model and a discriminant model, and in particular,

generating a model for fitting or estimating the probability of real connection distribution, and selecting the most likely node from the node set VThe connected nodes, expressed as: />G is a generator;

a discriminant model for discriminating the node v _C And calculates the output node and />There is a possibility of edges between, expressed as: />D is a discriminator;

the minmax objective function of generator G and arbiter D is as follows:

；

wherein the gradient isIs weighted +.>Gradient of->Is a desired sum of (2);

the implementation of generator G is defined by a softmax function, the following formula:

wherein ,, />∈R ^k is node-> and />K-dimensional vector representation, θ, in generator _G All->Is a collection of (3);

for the inner product and doing the exponential operation, +.>Generator G according to the specified conditions->Generate sample->。

The parameters of the generator and the arbiter are updated by continuous alternating training, and the arbiter D passes through the parameters from the generator and the arbiterTraining is performed on the positive samples from generator G, which is updated according to a gradient strategy, guided by the arbiter D.

In order to optimize the technical solution, in a specific embodiment of the present invention, the generator G is updated according to the gradient strategy under the guidance of the arbiter D.

Specifically, the arbiter is optimized as follows:

the arbiter D is a sigmoid function, as follows:

；

wherein ,is node-> and />K-dimensional vector representation in the arbiter, thus +.>Can be regarded as all +.>Is a collection of (3);

it should be noted that the arbiter may also use other methods, such as SDNE model.

For a given node pairUpdating the corresponding node expression vector by gradient descent method>The following formula:

further, the generator is optimized as follows:

for a generator, its objective function is to minimize the minimum function, so the update is optimized by gradient descent, and the gradient of the generator is calculated as follows:

；

it should be noted that the gradientIs weighted +.>Gradient of->That is to say if a generating node is identified as a negative sample node, then the probabilityThe weight corresponding to the gradient of the generation node is very large, so that the whole gradient is large.

The implementation of the production model is defined by a softmax function, as follows:

；

wherein ,, />∈R ^k is node-> and />K-dimensional vector representation, θ, in generator _G All->Is a collection of (3);

for the inner product and doing the exponential operation, +.>Generator G according to the specified conditions->Generate sample->。

Based on such a setting, the estimated connection distribution can be calculatedThen, a sample set +.>Finally, the method of SGD is used for getting more +.>。

Further, the estimated connection distribution in the generated model is realized through softmax, however, the traditional softmax model is generally used for calculating softmax values of all nodes in the whole graph, which is time-consuming; at the same time, the topology of the network itself would imply rich information, while the softmax function ignores these entirely, since it is "fair" looking at each node, just the normalization task is completed.

It is common practice to have a hierarchical softmax and a negative sampling approach to mitigate computational overhead, but neither consider the structural information of the graph. The algorithms like deep, node2vec have acquired the structure information of the graph already at the time of (bias) random walk, so that the following softmax can be accelerated by only negative sampling; while GraphGAN does not actually take the network structure into consideration at present, the network embedding obtained by learning is meaningless.

GraphGAN is intended to de-embed network structure information by Softmax, so Softmax is modified here and graphsoftmax is proposed.

Graph Softmax is still used to calculate the estimated connection distributionThe conditions that it needs to satisfy are:

(1) Normalized normalization: it is desirable to satisfy an efficient probability distribution,

(2) Graph-structure-aware: with network structure information, a simple idea is that for two vertices in the graph, their probability of connectivity should decrease as their shortest distance increases.

(3)Computationally efficient：Should cover only a few nodes in the graph, such as some and nodes +.>Closer nodes.

To solve for such graph softmax, a BFS wide search of the original network ω is developed to v _c Tree T being the root node _c ，N _C (v) Representing nodesIs to assume a given node +.>Neighbor node +.>E Nc (v), definition ∈Nc (v)>Equivalent to->The associated probability of (a) is as follows:

；

wherein each nodeAre all ∈10 by the root node>The only path to start arrives, the unique path is defined as: />, wherein />，/>；

Graph Softmax definitionThe following formula:

;

wherein ,for +.>Predicting the next sample->Probability of (2);

for +.>Predicting the next sample->Is a probability of (2).

What the generative model has last to do is sampling, in particular,

a simple method calculates all graph softmax valuesThen, based on the probability value, the tape is carried outRandom sampling of weights.

Another online sampling strategy is the slave treeRoot node of->Starting random walk, v is the node selected as the sample if the next step of the node v to which the current walk is to walk back to the parent node.

In a specific embodiment of the invention, a multi-modal graph-based self-encoder (Multimodal Graph Autoencoder, MGA) is used, and knowledge maps are converted into multi-modal representation forms through encoding and decoding operations on different types of nodes and edges, so that the expressive capacity and the reasoning capacity of a model are enhanced, the dependency relationship of the mutual nodes is captured, and emergency judgment is carried out.

First, the multi-mode self-encoder is a feedforward neural network, which comprises an input layer, a first hidden layer, a second hidden layer, a third hidden layer and an output layer, wherein the input layer is derived from image feature vectorsAnd text feature vector +.>The input is obtained in two modalities, the input layer is followed by a first hidden layer, for each modality the mapping input is denoted +.> and />The method comprises the steps of carrying out a first treatment on the surface of the Will-> and />Connected, joint mapping to the second hidden layer, expressed as joint multi-modal entity embedding +.>；

The decoder stage and the encoder have a completely symmetrical structure, mapping the embedded as input to a third hidden layer, denoted as and />And corresponding-> and />Sharing the same dimension; the decoder will and />Mapping to the output layer-> and />。

The output layer and the input layer of each modality have the same dimensions, the output layer also being called reconstruction layer, the purpose of which is to reconstruct the input feature vectors.

The overall architecture of the proposed multi-modal self-encoder is defined as:

；

；

；

；

；

；

；

where f is the sigmoid activation function, and />Representing the mapping weight matrix of the modal image and text from layer j to layer (j + 1) respectively,j= 1,2,3,4；/> and />Bias items of image and text in the j-th hidden layer, respectively,>is embedded->The symbol # -indicates the join operation.

Finally, the entire multi-mode self-encoder is trained with the goal of minimizing reconstruction errorsIs the sum of dissimilarities between the two modal input and output layers, as follows:

；

in another embodiment of the present invention, a knowledge graph is converted into a multi-modal representation using a transform model that maps entities and relationships in a low-dimensional continuous vector space, where the embedding of the head entity h, the tail entity t, and the relationship r satisfies the following formula:

；

for triples in the knowledge-graph, the TransE attempts to minimize the distance between h+r and t, while for triples outside the knowledge-graph, it attempts to maximize the distance between them.

In training the TransE, we minimize the loss:

；

Where γ is the spacing hyper-parameter, max (0, x) is intended to obtain the positive part of x, and d is the dissimilarity function, which may be the L1 norm or the L2 norm. S represents the original dataset consisting of triples present in the knowledge-graph, and S' represents the corrupted dataset consisting of triples not present in the knowledge-graph. S' is considered a negative sample in training, and is constructed by replacing the head or tail entity of each triplet as follows:

the TransAE model is a new knowledge graph representation learning method, and a multi-mode self-encoder is combined with the TransE model, wherein the TransE is a simple and effective knowledge graph representation learning method.

In the TransAE model, we combine the two models, while learning multi-modal knowledge and structural knowledge. Each entity in the knowledge graph corresponds to a plurality of pictures and a sentence description, firstly, visual and text feature vectors of the entity are extracted, and then the feature vectors are input into a multi-mode self-encoder to be jointly embedded to be used as entity representation. The relationship embedding is randomly initialized at the beginning of training. These entities and relationships represent models used to train us.

Structural lossfor (h, r, t) is represented by the formula:

；

wherein and />Representing the embedded representation of h and t, respectively. For better generalization we also add a regularizer Ω (θ) for the parameter set θ, whose weight is α.

Thus, we can train our model by minimizing the total loss L, as follows:

；

parameters in self-encoder and relational embedding,/>,/>,/>Randomly initialized at the beginning of training, weight parameters β and α are chosen to balance the loss +.>、/>And regular term->Is of the size and importance of (a).

Finally, in step 3, in order to improve the effect and efficiency of event emergency plan recommendation, multitask learning is introduced based on step 2, and a method of pre-training is performed by using bert, user-personalized event emergency plan recommendation, scene recommendation of related event links and link classification tasks are performed, and scoring calculation is performed by distributing weights of different tasks, so as to obtain the event emergency plan recommendation with interpretability.

Specifically, aiming at the defects of low utilization rate, poor generalization capability, high computational complexity and poor interpretability of the single-task learning data, in the embodiment, a multi-task learning strategy is used, and a plurality of tasks share a model, so that the data can be better utilized, the generalization performance is improved, and the computational complexity is reduced.

In the task, the multi-task learning method can improve the event emergency plan recommending effect and efficiency, and perform user personalized event emergency plan recommending, scene recommending of related event links and link classifying tasks, and scoring calculation is performed by distributing weights of different tasks, so that fault emergency plan recommending with interpretability is obtained. The knowledge entity may also provide varying degrees of assistance. And some information can be transferred between different tasks, and labels between different tasks are transferred through a knowledge entity, so that promotion effect on different tasks is achieved.

The present embodiment then selects a multitasking learning model with attention mechanism (Multi-Task Learning with Attention, MTL-Att). The model extracts features of each task by using an attention mechanism in combination with a BERT pre-training model, thereby learning common features, and uses a shared BERT model as a feature extractor in multi-task learning. In the training process, the model respectively adds a specific output layer for each task by learning the relation among the tasks, and dynamically adjusts the weight of the attention mechanism, thereby improving the performance of each task.

The MTL-Att-bert model can learn the importance of how to assign three different tasks, user-personalized event contingency plan recommendations, scene recommendations for related event links, and link classification, through an attention mechanism. Specifically, the model may learn a weight vector for each task and use the weight vectors to weight the output of each task to arrive at a final prediction result. The model can learn the three tasks simultaneously in the training process, and the goal of multi-task learning is realized. Finally, the importance of each task is better adjusted by the model through adjusting the value of the weight vector, so that a better event emergency plan recommendation result is obtained.

In summary, the idea of the invention is: firstly, aiming at the problems of a knowledge graph entity and a relationship list, a text classification algorithm based on deep learning is used for mining more emergency nodes, constructing more relevant triplet information and generating an emergency knowledge graph. Then, using a multi-modal graph-based self-encoder, knowledge maps can be converted into multi-modal representations by encoding and decoding different types of nodes and edges, thereby enhancing the expressive and inferential capabilities of the model. And capturing the dependency relationship of the mutual nodes, and judging the emergency event. Finally, in order to improve the event emergency plan recommending effect and efficiency, a method of introducing multitask learning and adding a bert as a pre-training model is adopted to conduct user personalized event emergency plan recommending, scene recommending of related event links and link classifying tasks, and scoring calculation is conducted by distributing weights of different tasks, so that the event emergency plan recommending with interpretability is obtained.

The foregoing description of the preferred embodiments of the present invention is provided for the purpose of limiting the invention, and is intended to be illustrative of the present invention as defined by the appended claims. Any modification, equivalent replacement or improvement made by those skilled in the art without departing from the principles of the present invention should be included in the scope of the present invention.

Claims (9)

1. The emergency command scheduling platform construction method based on deep learning is characterized by comprising the following steps of:

step 1, establishing event atom library data, monitoring emergency nodes by using a knowledge reasoning technology based on graph generation antagonism network GraphGAN, generating emergency triplet information, and constructing an event knowledge graph;

step 2, using a self-encoder MGA based on a multi-mode diagram to encode and decode different types of nodes and edges, converting a knowledge graph into a multi-mode representation form, capturing the inter-dependent relationship of the nodes, and judging an emergency event;

and 3, introducing a multi-task learning model MTL-Att, adding an attention mechanism and combining a bert pre-training model to construct the MTL-Att-bert model, and performing user personalized event emergency plan recommendation, event link scene recommendation and event classification tasks by distributing different task weight scoring calculations.

2. The emergency command and dispatch platform construction method based on deep learning according to claim 1, wherein in step 1, knowledge reasoning based on graph generation of an countermeasure network graph gan is used to obtain emergency time nodes to expand a fault knowledge graph, the countermeasure network graph gan comprises a forming model and a discriminant model, and specifically,

generating a model for fitting or estimating the real connection distribution probability, and selecting the most likely node from the node set VThe connected nodes, expressed as: />G is a generator;

a discriminant model for discriminating nodesAnd calculates the output node +.> and />The possibility of an edge exists between the two, expressed as: />D is a discriminator;

minmax objective function of generator G and arbiter D, formula (1) below:

（1）

wherein ,is a discriminator for discriminating node pairs +.>Is a connectivity of (2);

outputting a scalar representing the node +.> and />Probability of edge existing between them;

for all nodes->Is a vector of (2); />As a cost function;

data true distribution +_>Representing the generated sample, ++>For +.>Calculating expected values in a true distribution;

to be a generator model +.>Is the input condition of the generator, < >>Is a parameter of the generator, +.>To>Sample generated by generator G>C is the node index;

the parameters of the generator and the arbiter are updated by continuous alternating training, and the arbiter D passes through the parameters from the generator and the arbiterTraining is performed on the positive samples from generator G, which is updated according to the gradient strategy, guided by the arbiter D.

3. The emergency command and dispatch platform construction method based on deep learning of claim 2, wherein the discriminator D is a sigmoid function, and the following formula (2) is:

（2）

wherein ,is node-> and />K-dimensional vector representation in the arbiter, thus +.>Is all thatIs a collection of (3);

for inner product operation and conversion by means of a sigmoid function,/for the inner product operation>Taking the negative number for the inner product and performing the exponential operation, ++>For the discriminator +/for a given real sample>And prediction sample->Is a predicted result of (a);

for a given node pairUpdating the corresponding node expression vector by gradient descent method>The following formula (3)

(3)

wherein ,for->Gradient of->Representing the variable->Is true of distribution of->Representing the generation of samples from a generator>，/>Representing updating of the discriminator parameter with gradient descent +.>。

4. The emergency command and dispatch platform construction method based on deep learning according to claim 3, wherein the generator G optimizes updating by a gradient descent method, and the gradient of the generator is represented by the following formula (4):

(4)

wherein the gradient isIs weighted +.>Gradient of->Is a desired sum of (2);

generate sample->Under the condition of the generator G, utilizing the expected value of the discriminator;

generate sample->Generating samples from generator G condition>Is a desired value of (2);

the implementation of generator G is defined by a softmax function, as follows equation (5):

(5)

wherein ,, />∈R ^k is node-> and />K-dimensional vector representation, θ, in generator _G All->Is a collection of (3);

for the inner product and doing the exponential operation, +.>Generator G according to the specified conditions->Generate sample->。

5. The emergency command and dispatch platform construction method based on deep learning of claim 4, wherein,

based on the above formula (5), graph Softmax is proposed, and the estimated connection distribution is calculated by embedding network structure information into SoftmaxThe method specifically comprises the following steps: BFS wide search is performed on the original network omega, and the original network omega is unfolded to v _c Tree T being the root node _c ，N _C (v) Representing node->Is to assume a given node +.>Neighbor node +.>∈Nc(v) Definition ofEquivalent to->Is +.>The following formula (6):

(6)

wherein each nodeAre all ∈10 by the root node>The only path to start arrives, the unique path is defined as: />, wherein mFor the number of path samples, +.>，/>，In the condition->Lower and sample->A set of related conditions;

graph Softmax definitionThe following formula (7):

(7)

wherein ,for samples under given conditions->Predicting the next sample->Probability of (2);

for samples under given conditions->Predicting the next sample->Is a probability of (2).

6. The emergency command and dispatch platform construction method based on deep learning as claimed in claim 5, wherein in step 2, the multi-modal graph-based self-encoder MGA performs the decoding operation by using the encoder for different types of nodes and edges and the decoder, specifically as follows:

the multi-mode self-encoder is a feedforward neural network and comprises an input layer, a first hiding layer, a second hiding layer, a third hiding layer and an output layer, wherein the input layer is formed by combining image feature vectorsAnd text feature vector +.>The input is obtained in two modalities, the input layer is followed by a first hidden layer, for each modality the mapping input is denoted +.> and />The method comprises the steps of carrying out a first treatment on the surface of the Will-> and />Connected, joint mapping to the second hidden layer, expressed as joint multi-modal entity embedding +.>；

The decoder stage and the encoder have a completely symmetrical structure, mapping the embedded as input to a third hidden layer, denoted as and />And corresponding-> and />Sharing the same dimension; the decoder will and />Mapping to the output layer-> and />The method comprises the steps of carrying out a first treatment on the surface of the The output layer and the input layer of each mode have the same dimension;

the overall architecture of the multi-mode self-encoder is defined as:

(8)

(9)

(10)

(11)

(12)

(13)

(14)

where f is the sigmoid activation function, and />Representing the mapping weight matrix of the modal image and text from layer j to layer (j + 1) respectively,j = 1,2,3,4；/> and />The bias terms of the image and text in the j-th hidden layer,is embedded->The symbol # -represents a join operation;

the multi-mode self-encoder trains with the aim of minimizing reconstruction errorsIs the sum of the dissimilarities between the two modal input and output layers, equation (15) below:

(15)。

7. the emergency command and dispatch platform construction method based on deep learning of claim 6, wherein in step 2, a knowledge graph is converted into a multi-modal representation form by using a transition model, the transition model maps entities and relations in a low-dimensional continuous vector space, and the embedding of a head entity h, a tail entity t and a relation r satisfies the following triplet formula (16):

（16）

for triples in the knowledge-graph, the TransE model attempts to minimize the distance between h+r and t, while for triples outside the knowledge-graph, it attempts to maximize the distance between them;

minimizing losses during TransE trainingThe following formula (17):

(17)

where γ is the spacing hyper-parameter, max (0, x) is intended to obtain the positive part of x, d is the dissimilarity function,a symbolic representation representing the head, relation and tail entities of the triplet at the time of prediction, S representing the original dataset consisting of triples present in the knowledge-graph and S' representing the corrupted dataset consisting of triples not present in the knowledge-graph; s' is considered a negative sample in training, constructed by replacing the head or tail entity of each triplet as shown in equation (18) below:

(18)

wherein ,Eis an entity set;

constructing a TransAE model, combining the two models of the formula (17) and the formula (18), and simultaneously learning multi-mode knowledge and structural knowledge, wherein each entity in a knowledge graph corresponds to a plurality of pictures and a sentence description, extracting visual and text feature vectors of the knowledge graph, inputting the feature vectors into a multi-mode self-encoder to obtain joint embedding as entity representation, and randomly initializing relation embedding at the beginning of training, wherein the entity and the relation representation are used for training the model;

structural lossfor (h, r, t) is expressed as the following formula (19):

(19)

wherein and />Representing the embedded representation of h and t, respectively, < >> and ；/>Respectively representing embedded representation forms of h 'and t';

a regularizer Ω (θ) is added for the parameter set θ, weighted α, training the model by minimizing the total loss L, as shown in equation (20):

（20）

parameters in self-encoder and relational embedding,/>,/>,/>Randomly initialized at the beginning of training, weight parameters β and α are chosen to balance the loss +.>、/>And regular term->Is of the size and importance of (a).

8. The emergency command scheduling platform construction method based on deep learning according to claim 7, wherein in step 3, a multi-task learning model MTL-Att is introduced, an attention mechanism is added to combine with a bert pre-training model, specifically, in the multi-task learning model, features of each task are extracted, common features are learned, a shared bert model is used as a feature extractor to form an MTL-Att-bert model, in the training process, the MTL-Att-bert model learns a weight vector for each task through the attention mechanism to weight the output of each task, and the importance of each task is adjusted by adjusting the value of the weight vector to obtain an event emergency plan recommendation result, which is used for user-personalized event emergency plan recommendation, related event link scene recommendation and link classification tasks.

9. A deep learning-based emergency command and dispatch platform construction system, which is characterized by being constructed by using the emergency command and dispatch platform construction method based on any one of claims 1-8, and comprising the following modules:

the event atom library acquisition module is used for acquiring input data and constructing a data atom library;

the event knowledge graph construction module; the method comprises the steps of generating an countermeasure network GraphGAN knowledge reasoning technology based on a graph, monitoring emergency nodes, generating emergency triplet information and constructing an event knowledge graph;

the multi-modal map self-encoder module: the method comprises the steps of performing encoding and decoding operations on nodes and edges of different types, converting a knowledge graph into a multi-mode representation form, capturing the mutual dependence of the nodes, and performing emergency judgment;

the MTL-Att-bert model building module is used for extracting the characteristics of each task by combining a bert pre-training model through using an attention mechanism;

and the event emergency plan recommending module is used for carrying out user personalized event emergency plan recommending, related event link scene recommending and link classifying tasks.