Emergency event identification method based on neural network model and computer device

Abstract

The application provides an emergency event identification method based on a neural network model and a computer device, obtains a multi-source perception data set covering a target area, performs feature coding processing on the multi-source perception data set to obtain a multi-dimensional feature representation set of a perception data unit, calls a pre-trained emergency event identification neural network model to perform event identification processing on the multi-dimensional feature representation set to generate an event identification result of the perception data unit, determines an emergency event type existing in the target area and spatio-temporal feature information of the emergency event in the perception data unit according to the event identification result, generates an emergency warning instruction containing an event positioning mark based on the emergency event type and the spatio-temporal feature information, and sends the emergency warning instruction to a target emergency response terminal to start a response operation. The application improves the comprehensiveness and accuracy of emergency event identification, enhances the conversion efficiency of identification results to actual responses, and effectively supports timely warning of emergency events.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence, and more specifically, to an emergency event identification method and computer device based on a neural network model. Background Technology

[0002] With the development of information technology and sensing devices, emergency event identification technology has been widely applied in fields such as public safety and disaster early warning. This technology analyzes relevant data to identify potential or ongoing emergency events, supporting timely warnings and responses. Currently, it typically involves acquiring single-type monitoring data, extracting single-dimensional features, and then using traditional classification models to analyze these features and generate event identification results. However, this identification method, relying solely on single-source data and single-dimensional features, cannot simultaneously cover information on changes in the physical state of the event and information on human feedback intentions, resulting in an incomplete characterization of the event's essence. Furthermore, the generated identification results often only contain probabilistic judgments of the event type, lacking specific spatiotemporal location information, making it difficult to directly translate into actionable early warning instructions, thus affecting the timeliness and accuracy of emergency response. Summary of the Invention

[0003] This invention provides an emergency event identification method and computer device based on a neural network model.

[0004] In a first aspect, embodiments of the present invention provide an emergency event identification method based on a neural network model. The method includes: acquiring a multi-source sensing data set covering a target area, the multi-source sensing data set containing multiple sensing data units with timestamps, each sensing data unit consisting of environmental sensing data and interactive feedback data; performing feature encoding processing on the multi-source sensing data set to obtain a multi-dimensional feature representation set of the sensing data units, the multi-dimensional feature representation set containing temporal correlation features and semantic correlation features; calling a pre-trained emergency event identification neural network model to perform event identification processing on the multi-dimensional feature representation set to generate an event identification result for the sensing data units, the event identification result containing a predicted probability distribution of different emergency event categories; determining the type of emergency event existing in the target area and the spatiotemporal feature information of the emergency event in the sensing data units based on the event identification result; generating an emergency warning instruction containing an event location identifier based on the emergency event type and the spatiotemporal feature information, and sending the emergency warning instruction to a target emergency response terminal to initiate a response operation.

[0005] Secondly, embodiments of the present invention provide a computer device, including: a memory storing a computer program; and a processor for loading the computer program to implement the emergency event identification method based on a neural network model as described above.

[0006] The emergency event identification method based on a neural network model provided by this invention acquires a multi-source sensing data set containing environmental sensing data and interactive feedback data. This simultaneously covers information on the physical state changes of emergency events and information on human intervention intentions, avoiding the problem of incomplete coverage of event information by single-source data. By performing feature encoding processing on the multi-source sensing data, temporal correlation features reflecting the evolution of the state over continuous time and semantic correlation features reflecting the contextual intentions of interactive content are extracted, forming a multi-perspective characterization of the physical development process of the event and the intention of human intervention. This allows subsequent model inference to be based on both types of key attribute information simultaneously. A pre-trained emergency event identification neural network model is invoked to jointly process the two-dimensional features. The model's feature fusion capabilities enable complementary reasoning between two types of features, enhancing the depth of understanding of the essence of events. Based on the identification results, the model determines the type of emergency event and its spatiotemporal features, transforming the probability distribution output by the model into event type identifiers and spatiotemporal location descriptions in specific scenarios, ensuring the practicality of the identification results. The model generates and sends emergency warning commands containing event location identifiers, converting the identification results into actionable response information. This allows emergency response terminals to quickly initiate corresponding operations based on event type, spatiotemporal features, and location identifiers, improving both the comprehensiveness and accuracy of emergency event identification and the efficiency of converting identification results into actual responses, thus effectively supporting timely warnings and accurate responses to emergency events. Attached Figure Description

[0007] Figure 1 This is a flowchart of an emergency event identification method based on a neural network model provided in an embodiment of the present invention.

[0008] Figure 2 This is a schematic diagram of the composition of a computer device provided in an embodiment of the present invention. Detailed Implementation

[0009] Please see Figure 1 , Figure 1 The flowchart illustrates an emergency event identification method based on a neural network model, provided in an embodiment of the present invention. This method, which can be executed by a computer device, includes the following steps:

[0010] Step S100: Obtain a multi-source sensing data set covering the target area. The multi-source sensing data set contains multiple sensing data units with timestamps. Each sensing data unit consists of environmental sensing data and interactive feedback data.

[0011] The target area refers to the geographical region requiring emergency event monitoring and identification, such as a city block, an industrial park, or a nature reserve. A multi-source sensing dataset is a collection of data gathered from various sources, including but not limited to sensor networks, surveillance cameras, social media, and emergency alarm systems. A timestamp is added to each sensing data unit to record the specific time of data collection for subsequent time-series analysis. A sensing data unit is the basic data unit in the multi-source sensing dataset, containing both environmental sensing data and interactive feedback data. Environmental sensing data, collected from various environmental sensors, describes the environmental conditions of the target area, such as temperature, humidity, air pressure, air quality, and noise levels. Interactive feedback data is obtained through user interaction, communication devices, or other feedback channels, such as user comments on social media, alarm information from emergency alarm systems, and personnel behavior information from surveillance cameras.

[0012] In practical applications, various types of sensor networks can be deployed for environmental sensing data. For example, in a city block, temperature sensors can be installed to monitor ambient temperature in real time, and air quality sensors can be installed to detect the concentration of pollutants in the air. These sensors collect data at regular time intervals and add a timestamp to each data point. For interactive feedback data, it can be obtained from data sources such as social media platforms and emergency alarm systems through data interfaces. For example, by collaborating with social media platforms, posts and comments related to emergency events can be obtained through API interfaces, while alarm information can be obtained from emergency alarm systems, and these data can also be timestamped. These sensing data units from different data sources are integrated to form a multi-source sensing data set covering the target area.

[0013] Step S200: Perform feature encoding processing on the multi-source sensing data set to obtain a multi-dimensional feature representation set of the sensing data unit. The multi-dimensional feature representation set includes temporal correlation features and semantic correlation features.

[0014] Feature encoding is the process of transforming raw multi-source sensory data sets into more representative and analyzable feature representations. A multidimensional feature representation set is a collection of features containing multiple dimensions, which can more comprehensively describe the characteristics of sensory data units. Temporal correlation features reflect the relationships and changes of sensory data units over time, such as data change trends between adjacent time stamps and periodic change patterns. Semantic correlation features embody the semantic information and contextual relationships of interactive content within sensory data units, such as the topic, intent, and sentiment in user comments.

[0015] In one implementation, step S200 may specifically include the following steps S210 to S250:

[0016] Step S210: Perform data cleaning on the multi-source sensing data set to eliminate noise interference information in the sensing data units and obtain a denoised sensing data set.

[0017] Data cleaning is a crucial preprocessing step for raw data, aiming to remove noise, errors, missing values, and other interfering information to improve data quality and usability. Noise interference can originate from sensor measurement errors, data transmission interference, human input errors, and more. A denoised sensing dataset is a clean and reliable dataset obtained after data cleaning.

[0018] In practice, missing values ​​can be filled using interpolation. For example, if temperature data is missing at a certain time stamp in environmental sensing data, linear interpolation can be used to estimate the missing value based on the temperature values ​​of adjacent time stamps. Noisy data can be removed using filtering algorithms. For example, median filtering can be used to process salt-and-pepper noise in image data collected by surveillance cameras. The specific implementation of median filtering is that for each pixel in the image, the median value of its neighboring pixels is taken as the new value for that pixel, thus effectively removing noise points. Through these data cleaning methods, noise interference information in the sensing data units can be eliminated, resulting in a denoised sensing data set.

[0019] Step S220: Perform time-series correlation extraction processing on the environmental sensing data in the denoised sensing data set to generate state evolution features of sensing data units under continuous time stamps as time-series correlation features. The state evolution features include a description of the change trend between adjacent time stamps.

[0020] Temporal correlation extraction is the process of mining temporal correlations and changes from environmental sensing data in a denoised sensing dataset. State evolution features describe the state changes of sensing data units at consecutive time points, reflecting the dynamic changes in the environmental state. The description of trends between adjacent time points specifically illustrates the direction and extent of change in the environmental sensing data between two adjacent time points, such as whether the temperature is increasing or decreasing, and by what magnitude.

[0021] In one implementation, step S220 may specifically include the following steps S221 to S225:

[0022] Step S221: Extract the environmental sensing data change area between adjacent sensing data units and analyze the difference characteristics of attribute parameters in the change area.

[0023] Adjacent sensing data units refer to two sensing data units that are adjacent in time series. Environmental sensing data change areas refer to the regions where environmental sensing data changes between adjacent sensing data units. Attribute parameter difference characteristics refer to the differences between different attribute parameters within the change area, such as the temperature difference within a temperature change area, or the humidity difference within a humidity change area.

[0024] In practice, extracting areas of environmental sensing data change between adjacent sensing data units can be achieved by comparing environmental sensing data from adjacent time stamps. For example, for temperature data, the temperature difference between adjacent time stamps can be calculated, and areas with non-zero differences can be marked as temperature change areas. Similarly, for humidity data, the humidity difference between adjacent time stamps can be calculated to mark humidity change areas. Then, the differences in attribute parameters of these change areas can be analyzed. For example, for temperature change areas, the average temperature difference, maximum temperature difference, and minimum temperature difference within the area can be calculated to more comprehensively describe the degree and characteristics of temperature changes.

[0025] Step S222: Determine the direction and degree of state change of the sensing data unit between adjacent timestamps based on the attribute parameter difference characteristics.

[0026] The direction of state change refers to whether the state of the sensing data unit evolves in a certain direction between adjacent time stamps, such as whether the temperature rises or falls, or whether the air pressure increases or decreases. The description of the degree of change specifically describes the magnitude of the state change, such as how many degrees the temperature has increased, or how many Pascals the air pressure has increased.

[0027] Determining the direction and degree of state changes based on differences in attribute parameters can be achieved through simple comparisons and calculations. For example, for temperature data, a positive temperature difference between adjacent time stamps indicates a temperature increase, with the absolute value of the difference representing the magnitude of the increase; conversely, a negative difference indicates a temperature decrease, with the absolute value representing the magnitude of the decrease. Similarly, for air pressure data, the direction of pressure change is determined by the sign of the pressure difference, and the degree of change is determined by the absolute value of the difference. In this way, the state changes of the sensing data unit between adjacent time stamps can be accurately described.

[0028] Step S223: Construct a state evolution analysis model for the sensing data unit based on the description of the direction and degree of state change.

[0029] The state evolution analysis model is a mathematical model used to describe the state evolution of a sensing data unit at consecutive time points. Based on the direction and degree of state change, it can predict the state changes of the sensing data unit at future time points.

[0030] Various methods can be employed when constructing a state evolution analysis model. For example, the Autoregressive Integral Moving Average (ARIMA) model from time series analysis can be used. For instance, a stationarity test is performed on the data describing the direction and degree of state change. If the data is non-stationary, differencing is performed to transform it into a stationary series. Then, based on the graphical characteristics of the autocorrelation function (ACF) and partial autocorrelation function (PACF), the order (p, d, q) of the ARIMA model is determined. Here, p represents the order of the autoregressive term, d represents the order of differencing, and q represents the order of the moving average term. Finally, methods such as least squares are used to estimate the model parameters, yielding the final state evolution analysis model.

[0031] Step S224: Calculate the consistency description of the state change trend of the sensing data unit under continuous time stamps through the state evolution analysis model.

[0032] The description of consistent state change trends refers to whether the state change trends of a sensed data unit remain consistent across consecutive time stamps. For example, if the temperature consistently shows an upward trend across multiple adjacent time stamps, it indicates that the temperature state change trend is consistent; if the temperature sometimes rises and sometimes falls, it indicates that the state change trend is inconsistent.

[0033] Specifically, a state evolution analysis model is used to predict state changes at consecutive timestamps, obtaining the predicted direction and degree of state change for each timestamp. Then, the predicted directions of state change for adjacent timestamps are compared. If the predicted directions of state change for adjacent timestamps are the same, the state change trend is considered consistent between these two timestamps; otherwise, the state change trends are considered inconsistent. Finally, the number of time intervals with consistent state change trends among consecutive timestamps is counted, and their proportion of the total number of time intervals is calculated to describe the degree of consistency in the state change trends.

[0034] Step S225: Perform periodic pattern recognition processing on the state evolution process of the sensing data unit, and extract the state repetition features of the sensing data unit within a preset time window as state evolution features.

[0035] Periodic pattern recognition processing is the process of identifying periodic patterns in the state evolution of sensing data units. A preset time window refers to a pre-defined time range, such as a day, a week, or a month. State repetition characteristics refer to the recurring state characteristics of sensing data units within the preset time window; for example, the temperature value is roughly the same at the same time every day, or the air pressure change pattern is similar at the same time of week. Specifically, the Fourier transform method can be used. The basic principle of the Fourier transform is to convert a signal in the time domain into a signal in the frequency domain, and to identify periodic patterns by analyzing the signal characteristics in the frequency domain. Specifically, the state evolution data of the sensing data units is first divided according to the preset time window. Then, a Fourier transform is performed on the data within each time window to obtain its spectrum. The peaks in the spectrum represent the main frequency components of the signal. By analyzing the position and amplitude of these peaks, periodic patterns in the data can be identified. For example, if there is a significant peak in the spectrum, corresponding to a frequency of once a day, it indicates that the data has a daily periodic pattern. Finally, the state repetition characteristics corresponding to these periodic patterns are extracted as state evolution characteristics.

[0036] Step S230: Perform semantic association extraction processing on the interactive feedback data in the denoised perception dataset to generate semantic association features of the interactive content in the perception data unit. The semantic association features contain a contextual description of the interactive intent.

[0037] Semantic association extraction is the process of extracting semantic information and contextual relationships from interactive feedback data in a denoised perceptual dataset. Interactive content refers to the specific information contained in the interactive feedback data, such as user comments and alarm information. Semantic association features are features that reflect the semantic information and contextual relationships of interactive content, such as the topic, intent, and sentiment of the interactive content. Contextual descriptions of interactive intent further illustrate the association and meaning of the interactive intent within the context, such as the association between a topic in a user comment and other topics, or the association between an event in an alarm message and relevant background information.

[0038] In one implementation, step S230 may specifically include the following steps S231 to S235:

[0039] Step S231: Perform word segmentation on the text content in the interactive feedback data to obtain multiple interactive word units.

[0040] Word segmentation is the process of dividing the text content in interactive feedback data into independent word units according to certain rules. An interactive word unit is a single word or phrase obtained after word segmentation. Various word segmentation algorithms can be used to segment the text content in interactive feedback data. For example, for Chinese text, the Jieba word segmentation algorithm can be used. The Jieba word segmentation algorithm is a rule-based algorithm that combines dictionary matching and machine learning methods, and it can handle the word segmentation problem of Chinese text well. The specific implementation process is as follows: First, the Jieba word segmentation algorithm loads a pre-built dictionary containing a large number of Chinese words. Then, for the input text content, the algorithm matches the longest word in the dictionary. If the match is successful, the word is output as an interactive word unit; if the match fails, the text is segmented, and matching attempts continue. In this way, the text content is divided into multiple interactive word units. For English text, the word segmentation tools in the NLTK library can be used, which provide simple and effective English word segmentation functions.

[0041] Step S232: Call the pre-trained semantic encoding model to perform context encoding processing on multiple interactive word units to generate semantic vector representations of interactive feedback data.

[0042] Pre-trained semantic encoding models are models pre-trained on large-scale corpora, enabling them to learn the semantic information and contextual relationships of words and sentences. Contextual encoding is the process of encoding multiple interactive word units in conjunction with their contextual information to generate vector representations that better reflect their semantic meaning. Semantic vector representation converts interactive feedback data into a representation in vector space, allowing for more convenient semantic analysis and processing through vector operations and comparisons.

[0043] The BERT model can be used to perform contextual encoding on multiple interactive word units by calling a pre-trained semantic encoding model. The BERT model is a pre-trained language model based on the Transformer architecture. It learns contextual information about words and sentences by using bidirectional encoder representations from transformers. Specifically, multiple interactive word units are input into the BERT model. The BERT model encodes the input words, generating a contextual embedding vector for each word. These contextual embedding vectors are then aggregated, for example, using average pooling or max pooling, to obtain a semantic vector representation of the entire interactive feedback data. In this way, the interactive feedback data is transformed into a more semantically informative vector representation.

[0044] Step S233: Analyze the semantic association strength between different interactive word units in the semantic vector representation, and generate an association graph containing semantic association weights.

[0045] Semantic association strength refers to the degree of semantic connection between different interactive word units. For example, the semantic association strength between "fire" and "smoke" is relatively strong, while the semantic association strength between "fire" and "apple" is relatively weak. An association graph is a graphical structure used to represent the semantic association relationships between different interactive word units, where nodes represent interactive word units, edges represent semantic association relationships, and the weight of the edges represents the semantic association strength.

[0046] The strength of semantic associations between different interactive word units in a semantic vector representation can be analyzed using cosine similarity. Cosine similarity measures the similarity between two vectors by calculating the cosine of the angle between them. A cosine value closer to 1 indicates greater similarity and stronger semantic association; a cosine value closer to 0 indicates less similarity and weaker semantic association. Specifically, the vector for each interactive word unit is first extracted from the semantic vector representation. Then, the cosine similarity between any two interactive word unit vectors is calculated. The cosine similarity value is used as a measure of semantic association strength. Finally, an association graph is generated based on these semantic association strength values. For example, a graph database (such as Neo4j) can be used to store and manage the association graph, with interactive word units as nodes and semantic association strength as edge weights, constructing an association graph that includes semantic association weights.

[0047] Step S234: Extract the semantic features of the core interactive intent from the interaction feedback data based on the relationship graph as semantic association features.

[0048] Core interaction intent refers to the most important intent and information in the interaction feedback data, such as the main topic in user comments or the core event in alarm information. Semantic features are features that can reflect the semantic information of core interaction intent.

[0049] Extracting semantic features of core interactive intent from a relational graph can be implemented as follows: Analyze the relational graph to identify nodes with high connectivity; these nodes typically represent related words of the core interactive intent. For example, in a relational graph of interactive feedback data about fire alarms, the node "fire" might have high connectivity because it has strong semantic associations with nodes such as "smoke," "flame," and "alarm." Then, extract the semantic vector representations of these core nodes and use them as semantic features of the core interactive intent. Finally, use these semantic features as semantic association features. In this way, the semantic information of the core interactive intent in the interactive feedback data can be extracted more accurately.

[0050] Step S235: Perform multi-round context enhancement processing on semantic association features to strengthen the contextual coherence description of interaction intent.

[0051] Multi-round context enhancement processing involves processing semantically related features multiple times to further enhance their contextual coherence and semantic information. Contextual coherence description of interaction intent refers to the coherent and consistent description of the interaction intent within its context, such as whether topic transitions in user comments are natural, or whether event descriptions in alarm messages are complete.

[0052] Multi-round context enhancement of semantic association features can be achieved using methods such as Recurrent Neural Networks (RNNs) or Long Short-Term Memory Networks (LSTMs). Taking LSTM as an example, the semantic association features are first input into the LSTM network. The LSTM network updates the hidden state at the current time step based on the input semantic association features and the hidden state at the previous time step. Then, the updated hidden state is used as the new semantic association feature for the next round of processing. After multiple rounds of processing, the contextual coherence of the semantic association features is enhanced. Finally, the semantic association features after multiple rounds of context enhancement are output, strengthening the contextual coherence description of the interaction intent.

[0053] Step S240: Input the temporal correlation features and semantic correlation features into the feature integration module for correlation analysis processing to obtain an integrated feature set with a unified representation.

[0054] The feature integration module is used to integrate and analyze temporal and semantic features. Association analysis is the process of analyzing the relationships between temporal and semantic features. Unified representation refers to converting different types of features into a unified representation that facilitates subsequent processing and analysis. The integrated feature set is the set of features containing the unified representation obtained after processing by the feature integration module.

[0055] In one implementation, step S240 may specifically include the following steps S241 to S245:

[0056] Step S241: Perform feature mutual information calculation on temporal correlation features and semantic correlation features to generate a mutual information matrix that reflects the degree of information sharing between the two types of features.

[0057] Feature mutual information calculation is the process of calculating the mutual information between temporally related features and semantically related features. Mutual information is an indicator that measures the degree of information sharing between two random variables; it represents the extent to which information about one variable can be predicted by the other. The mutual information matrix is ​​a matrix in which the elements represent the mutual information values ​​between temporally related features and semantically related features.

[0058] The following general formula can be used to calculate the mutual information of temporal and semantic features: In this equation, X represents a temporal correlation feature, where x is its specific value; Y represents a semantic correlation feature, where y is its specific value; p(x,y) is the joint probability distribution of X and Y; and p(x) and p(y) are the marginal probability distributions of X and Y, respectively. Specifically, the joint probability distribution and marginal probability distributions of the temporal correlation feature and the semantic correlation feature are calculated. These probability distributions are estimated by analyzing the frequency of different feature values ​​in the statistical data. Then, the mutual information value between any two features is calculated according to the above formula. Finally, these mutual information values ​​are combined into a mutual information matrix. For example, assuming there are n temporal correlation features and m semantic correlation features, the mutual information matrix is ​​an n×m matrix, where each element represents the mutual information value between a temporal correlation feature and a semantic correlation feature.

[0059] Step S242: Based on the mutual information values ​​of each feature pair in the mutual information matrix, filter out key feature pairs whose mutual information values ​​exceed a preset threshold.

[0060] Key feature pairs refer to feature pairs whose mutual information value exceeds a preset threshold in temporal and semantic related features. These feature pairs have a strong degree of information sharing. The preset threshold is a pre-set mutual information value standard used to filter key feature pairs.

[0061] Filtering key feature pairs based on the mutual information values ​​of each feature pair in the mutual information matrix can be implemented as follows: First, determine a preset threshold. The preset threshold can be adjusted according to the specific application scenario and data characteristics. For example, in some cases, the preset threshold can be set to the average value of all elements in the mutual information matrix, or a suitable value can be determined based on experience and experimental results. Then, iterate through each element in the mutual information matrix, marking feature pairs whose mutual information values ​​exceed the preset threshold as key feature pairs. Finally, extract these key feature pairs for subsequent processing. By filtering key feature pairs, the number of features can be reduced, improving the efficiency and accuracy of subsequent processing.

[0062] Step S243: Perform feature space alignment processing on key feature pairs, and map temporally related features and semantically related features to the same feature space dimension through linear transformation.

[0063] Feature space alignment is the process of mapping features from different feature spaces to the same feature space dimension. Linear transformation is a method of transforming vectors in one vector space to another through linear combination. The same feature space dimension refers to transforming temporal and semantic features to the same dimensional space for subsequent fusion and analysis.

[0064] Principal Component Analysis (PCA) can be used to align key feature pairs in their feature space. The basic principle of PCA is to find the principal components of the data and project it into a low-dimensional feature space while preserving the main information. Specifically, the temporal and semantic features of the key feature pairs are first combined into a matrix. This matrix is ​​then standardized to have a mean of 0 and a variance of 1. Next, the covariance matrix is ​​calculated. By solving for the eigenvalues ​​and eigenvectors of the covariance matrix, the principal components of the data are obtained. The top k principal components with the largest eigenvalues ​​are selected as the main features (e.g., the top k principal components whose cumulative variance contribution exceeds a preset threshold (e.g., 95%)), and the original data is projected onto these principal components to obtain the dimensionality-reduced feature vectors. Finally, both the temporal and semantic features are projected into this low-dimensional feature space, achieving feature space alignment. This can be understood as performing standardization, such as Z-score, on the temporal and semantic features before PCA.

[0065] Step S244: Based on the spatial alignment results of key feature pairs, use feature concatenation operation to perform dimensional fusion processing on temporal correlation features and semantic correlation features to generate an initial integrated feature set with a unified representation.

[0066] Feature concatenation is the operation of concatenating temporal and semantic features, after feature space alignment, in terms of dimensionality. Dimensional fusion is the process of fusing features from different dimensions into a higher-dimensional feature representation. The initial integrated feature set is a set of features with a unified representation obtained after feature concatenation. Based on the spatial alignment results of key feature pairs, dimensional fusion of temporal and semantic features using feature concatenation can be achieved through simple vector concatenation.

[0067] Step S245: Perform feature redundancy detection on the initial integrated feature set, remove redundant feature dimensions that repeatedly express the same information, and generate the integrated feature set.

[0068] Feature redundancy detection is the process of detecting whether there are feature dimensions that repeatedly express the same information in the initial integrated feature set. Redundant feature dimensions refer to those feature dimensions that do not contribute additionally to subsequent analysis and processing; they may be due to duplication during data acquisition or redundancy during feature extraction. The integrated feature set is the feature set after feature redundancy detection, which removes redundant feature dimensions.

[0069] Redundancy detection of the initial integrated feature set can be performed using correlation analysis. Correlation analysis measures the similarity between features by calculating the correlation coefficient. Specifically, the correlation coefficient between any two feature dimensions in the initial integrated feature set is first calculated. The Pearson correlation coefficient can be used to measure linear correlation. Then, a correlation threshold is set, and feature dimensions with correlation coefficients exceeding the threshold are marked as redundant feature dimensions. Finally, these redundant feature dimensions are removed to generate the integrated feature set. By removing redundant feature dimensions, the number of features can be reduced, computational complexity can be lowered, and the generalization ability of the model can be improved.

[0070] Step S250: Based on the correlation strength evaluation results of features of different dimensions in the integrated feature set, dynamically adjust the fusion emphasis of temporal correlation features and semantic correlation features in feature encoding processing to generate a multidimensional feature representation set.

[0071] Association strength assessment results are the evaluation and analysis of the association strength between features of different dimensions in the integrated feature set, such as the degree of association between features obtained through methods like mutual information calculation and correlation analysis. Dynamic adjustment refers to changing the fusion ratio and method of temporal and semantic association features in feature encoding processing in real time based on the association strength assessment results. Fusion emphasis direction refers to whether the feature encoding process places more emphasis on fusing temporal or semantic association features, or a balanced fusion of both. The multidimensional feature representation set is a set containing features of multiple dimensions obtained after dynamically adjusting the fusion emphasis direction.

[0072] Based on the association strength evaluation results of features of different dimensions in the integrated feature set, dynamically adjusting the fusion emphasis of temporal and semantic association features in feature encoding processing can be implemented as follows: Determine the association relationships between features of different dimensions based on the association strength evaluation results. For example, if a temporal association feature has a high association strength with multiple semantic association features, it indicates that the temporal association feature plays an important role in the fusion. Then, adjust the weights of temporal and semantic association features in feature encoding processing according to these association relationships. For example, if it is found that semantic association features have a greater impact on the final result in certain situations, appropriately increase the weight of semantic association features and decrease the weight of temporal association features. Finally, based on the adjusted weights, fuse the temporal and semantic association features to generate a multidimensional feature representation set. By dynamically adjusting the fusion emphasis, the multidimensional feature representation set can more accurately reflect the characteristics of the perceptual data unit.

[0073] Step S300: Call the pre-trained emergency event recognition neural network model to perform event recognition processing on the multi-dimensional feature representation set, and generate the event recognition result of the perception data unit. The event recognition result includes the predicted probability distribution of different emergency event categories.

[0074] A pre-trained emergency event recognition neural network model is a neural network model pre-trained on large-scale emergency event data. It learns the characteristics and patterns of emergency events and is used to recognize events from new sensor data. Event recognition processing involves inputting a multi-dimensional feature representation set into the emergency event recognition neural network model. Through the model's calculation and inference, it determines the types of emergency events that may exist in the sensor data unit. The event recognition result is the model's output regarding the identification results of emergency event types in the sensor data unit, including the predicted probability distribution for different emergency event categories. The predicted probability distribution refers to the probability that the model predicts for each emergency event category to occur.

[0075] In one implementation, step S300 may specifically include the following steps S310 to S350:

[0076] Step S310: Input the multidimensional feature representation set into the feature processing layer of the emergency event recognition neural network model, perform hierarchical feature fusion processing, and generate a fused feature vector with multi-granularity representation.

[0077] The feature processing layer is used to process and fuse the input set of multidimensional feature representations. Hierarchical feature fusion processing fuses features from different dimensions in the multidimensional feature representation set at multiple levels to extract richer feature information. Multi-granularity representation refers to the ability of the fused feature vector to represent feature information at different granularity levels, such as different levels from local details to overall features. The fused feature vector is a comprehensive feature vector obtained after hierarchical feature fusion processing, containing the main information from the multidimensional feature representation set.

[0078] To perform hierarchical feature fusion processing on the feature processing layer of an emergency event recognition neural network model, a convolutional neural network (CNN) structure can be used. The basic principle of a CNN is to extract and fuse features from input data through a combination of convolutional layers, pooling layers, and fully connected layers. Specifically, the multidimensional feature representation set is first input into the convolutional layer of the CNN. The convolutional layer extracts local features by sliding convolutional kernels across the input data. Different sized convolutional kernels can extract feature information of different granularities; for example, small convolutional kernels can extract local detailed features, while large convolutional kernels can extract overall features. Then, the output of the convolutional layer is downsampled by the pooling layer to reduce the dimensionality of the features while retaining the main feature information. Methods such as max pooling or average pooling can be used. Finally, the output of the pooling layer is combined and fused through a fully connected layer to generate a fused feature vector with multi-granularity representations. For example, for a multidimensional feature representation set containing multiple temporal and semantic features, hierarchical feature fusion processing by a CNN can yield a fused feature vector that comprehensively reflects these feature information.

[0079] Step S320: The context modeling layer of the emergency event recognition neural network model is used to perform context dependency modeling on the fused feature vector to generate context-enhanced features in the time and semantic dimensions.

[0080] The context modeling layer is used to model and analyze the contextual dependencies of the fused feature vector. Contextual dependencies refer to the temporal and semantic associations and dependencies between different elements in the fused feature vector, such as the association between the states of an emergency event at different points in time, and the association between different semantic information. The context-enhanced features in the temporal and semantic dimensions are features obtained after contextual dependency modeling, and they can more comprehensively reflect the temporal and semantic contextual information of the emergency event.

[0081] In one implementation, step S320 may specifically include the following steps S321 to S325:

[0082] Step S321: Input the fused feature vector into the time dependency sub-layer of the context modeling layer to perform time series dependency analysis and generate time-dimensional context dependency features.

[0083] The time dependency sublayer is a sublayer within the context modeling layer, specifically designed to handle the dependencies between different points in time series data. Time series dependency analysis involves analyzing time-related information in the fused feature vector to identify associations and dependencies between different points in time. The time-dimensional context dependency features are obtained after time series dependency analysis and reflect the development and changes of an emergency event over time.

[0084] Long Short-Term Memory (LSTM) networks can be used to analyze the sequential dependencies of time series data by inputting the fused feature vectors into the time dependency sublayer of the context modeling layer. LSTM effectively handles long-term dependencies in time series data by controlling the flow and retention of information through gating mechanisms. Specifically, the fused feature vectors are first input into the LSTM network in chronological order. The LSTM network updates the hidden state at the current time step based on the input at the current time step and the hidden state at the previous time step. The hidden state contains historical and contextual information of the time series. Then, the hidden state output by the LSTM network is used as the context dependency feature of the time dimension. For example, for emergency event data containing multiple timestamps, the LSTM network can process the data to obtain the context dependency feature of the time dimension for each timestamp. These features can reflect the development trend and change patterns of the emergency event over time.

[0085] Step S322: Input the fused feature vector into the semantic dependency sub-layer of the context modeling layer, perform semantic content contextual relationship analysis, and generate semantic dimension contextual dependency features.

[0086] The semantic dependency sublayer is another sublayer in the context modeling layer, used to process the contextual relationships of semantic content in the fused feature vector. The contextual relationship analysis of semantic content involves analyzing semantically relevant information in the fused feature vector to identify the associations and dependencies between different semantic elements. The contextual dependency features of the semantic dimension are features obtained after the contextual relationship analysis of semantic content; they reflect the semantic coherence and relevance of the emergency event.

[0087] To analyze the semantic content's contextual relationships, the multi-head attention mechanism in the Transformer architecture can be used to input the fused feature vector into the semantic dependency sublayer of the context modeling layer. The basic principle of multi-head attention is to simultaneously focus on different parts of the input data by multiple attention heads, calculating the importance weight of each part, thereby modeling the contextual relationships of semantic content. Specifically, the fused feature vector is first input into the multi-head attention mechanism. The mechanism calculates the query vector, key vector, and value vector for each position. Then, it calculates the attention score by taking the dot product of the query vector and key vector, and performs a weighted summation of the value vectors to obtain the output for each position. The outputs of multiple attention heads are concatenated and linearly transformed to obtain the final semantic-dimensional contextual dependency features. For example, for a fused feature vector containing multiple semantically related features, processing with the multi-head attention mechanism can yield semantic-dimensional contextual dependency features that reflect the contextual relationships of semantic content.

[0088] Step S323: Perform cross-dimensional correlation analysis on the context-dependent features of the time dimension and the context-dependent features of the semantic dimension to generate cross-dimensional context-dependent features.

[0089] Cross-dimensional correlation analysis analyzes context-dependent features in both the temporal and semantic dimensions to identify their relationships and interactions. Cross-dimensional contextual correlation features, obtained after cross-dimensional correlation analysis, comprehensively reflect the contextual information between the temporal and semantic dimensions.

[0090] Cross-dimensional correlation analysis of temporal and semantic context-dependent features can be performed using a combination of fully connected layers and attention mechanisms. Specifically, the temporal and semantic context-dependent features are first concatenated. Then, the concatenated vector is input into a fully connected layer for linear transformation, yielding a new feature vector. Next, an attention mechanism is used to process this new feature vector, calculating the importance weight for each dimension. The weighted summation through the attention mechanism yields the cross-dimensional context-dependent features. For example, for a vector containing temporal and semantic context-dependent features, processing with a fully connected layer and attention mechanism can produce cross-dimensional context-dependent features that reflect the relationship between the two dimensions.

[0091] Step S324: Based on the cross-dimensional context association features, perform weighted fusion processing on the context dependency features of the time dimension and the context dependency features of the semantic dimension to generate fused context dependency features.

[0092] Weighted fusion is a process of combining temporal and semantic context-dependent features based on cross-dimensional contextual association features. The fused context-dependent feature is a comprehensive feature obtained after weighted fusion, which includes temporal and semantic contextual information as well as the relationships between them.

[0093] Weighted fusion of temporal and semantic context-dependent features based on cross-dimensional contextual features can be implemented as follows: Calculate the weights of the temporal and semantic context-dependent features based on the cross-dimensional contextual features. The softmax function can be used to convert the cross-dimensional contextual features into weight values, ensuring the sum of the weights is 1. Then, multiply the temporal and semantic context-dependent features by their respective weights. Finally, sum the weighted features to generate the fused context-dependent features. For example, if the cross-dimensional contextual features indicate that the temporal context-dependent features are more important in the current context, then give the temporal context-dependent features a higher weight, thus highlighting the temporal information in the fusion.

[0094] Step S325: Perform dimensional alignment processing on the fused context-dependent features to generate context-enhanced features in the temporal and semantic dimensions.

[0095] Dimension alignment involves adjusting the dimensions of context-dependent features to ensure consistent representation and size across both the temporal and semantic dimensions. Context-enhanced features in the temporal and semantic dimensions are obtained after dimensional alignment; they more accurately reflect the temporal and semantic contextual information of emergency events and maintain dimensional consistency, facilitating subsequent processing and analysis.

[0096] Dimension alignment of fused context-dependent features can be achieved using linear transformation methods. Specifically, the target dimensions for both the temporal and semantic dimensions are first determined. These target dimensions can be set based on the model design and subsequent processing requirements. Then, a linear transformation matrix is ​​used to project the fused context-dependent features onto the target dimensions. This linear transformation matrix can be obtained through training, ensuring that the projected features retain as much information as possible from the original features. Finally, context-enhanced features with consistent dimensions in both the temporal and semantic dimensions are obtained.

[0097] Step S330: The probability prediction layer based on the emergency event recognition neural network model performs event probability prediction processing on the context enhancement features to generate a probability distribution set containing the predicted probability values ​​corresponding to each emergency event category.

[0098] The probability prediction layer is a layer in the emergency event recognition neural network model used to predict the probability of events based on context-enhanced features. The event probability prediction process calculates the probability of each emergency event category based on the context-enhanced features. The probability distribution set is the set of predicted probability values ​​corresponding to each emergency event category output by the model.

[0099] The probabilistic prediction layer of an emergency event recognition neural network model can use the softmax function to process context-enhanced features for event probability prediction. The basic principle of the softmax function is to transform the input vector into a probability distribution such that the sum of the probabilities of all elements is 1. Specifically, the context-enhanced features are first input into the fully connected layer of the probabilistic prediction layer for a linear transformation, resulting in an output vector. Then, this output vector is input into the softmax function to calculate the probability value corresponding to each emergency event category.

[0100] Step S340: Perform category confidence calibration on the probability distribution set, adjust the confidence weights of the predicted probability values ​​of different emergency event categories, and generate the calibrated probability distribution set.

[0101] Category confidence calibration is the process of adjusting and calibrating the confidence levels of predicted probability values ​​for different emergency event categories within a probability distribution set. Confidence weights refer to the degree of trust in the predicted probability value for each emergency event category. By adjusting the confidence weights, the predicted probability values ​​can more accurately reflect the actual situation. The calibrated probability distribution set is the set obtained after category confidence calibration, and the predicted probability values ​​within it are more reliable and trustworthy.

[0102] Class confidence calibration of a probability distribution set can be performed using the Platt scaling method, which calibrates the probability distribution set through a logistic regression model. Specifically, a calibration dataset is first prepared, containing samples of known emergency event categories and their corresponding predicted probabilities. Then, a logistic regression model is trained on the calibration dataset to learn the relationship between the predicted probabilities and the true categories. The output of the logistic regression model is a calibrated probability value. Finally, each predicted probability value in the probability distribution set is input into the trained logistic regression model to obtain the calibrated probability value, generating the calibrated probability distribution set.

[0103] Step S350: Generate event identification results for the sensing data unit based on the calibrated probability distribution set. The event identification results include the predicted probability distribution of each emergency event category.

[0104] The event identification results generated from the calibrated probability distribution set for the sensing data unit can be directly used as the event identification results. This is because the calibrated probability distribution set has undergone confidence level calibration, more accurately reflecting the probability of each emergency event category occurring. For example, if the predicted probability of a fire event is 0.7, the predicted probability of an earthquake event is 0.2, and the predicted probability of other events is 0.1 in the calibrated probability distribution set, then the event identification result is the predicted probability distribution containing these probability values, indicating that in the current sensing data unit, the fire event has the highest probability of occurring, followed by the earthquake event, and other events have a lower probability of occurring.

[0105] Step S400: Determine the type of emergency event existing in the target area and the spatiotemporal characteristics of the emergency event in the sensing data unit based on the event identification results.

[0106] The event identification result is obtained through an emergency event identification neural network model, containing the predicted probability distribution of each emergency event category. Emergency event type refers to different types of emergency events that may exist in the target area, such as fire, earthquake, and flood. Spatiotemporal feature information refers to the temporal and spatial characteristics of the emergency event within the sensing data unit, such as the time of occurrence, duration, area of ​​occurrence, and direction of spread of the emergency event.

[0107] In one implementation, step S400 may specifically include the following steps S410 to S450:

[0108] Step S410: Analyze the predicted probability distribution in the event recognition result, and detect the probability peak points that exceed the preset threshold and their corresponding sensing data unit timestamp information.

[0109] Analyzing the predicted probability distribution involves interpreting and analyzing the set of probability distributions in the event identification results to identify key information. A pre-set confidence threshold is a predetermined probability value standard used to filter out emergency event predictions with high credibility. Probability peak points are the points in the predicted probability distribution where the probability value reaches its peak; these points typically indicate a higher probability of the emergency event occurring. Sensing data unit timestamp information refers to the timestamp of the sensing data unit corresponding to the probability peak point, used to determine the time of the emergency event.

[0110] Analyzing the predicted probability distribution in the event recognition results and detecting probability peaks exceeding a preset confidence threshold and their corresponding sensing data unit timestamps can be implemented as follows: Set a preset confidence threshold, then iterate through the predicted probability distribution in the event recognition results to find points whose probability values ​​exceed the threshold. For each point exceeding the threshold, check if it is a probability peak, i.e., its probability value is greater than that of its neighboring points. If it is a probability peak, record the corresponding sensing data unit timestamp.

[0111] Step S420: Based on the timestamp information corresponding to the probability peak point, and combined with the continuous change trend of the predicted probability of the preceding and following sensing data units, determine the start and end time points of the emergency event, and generate time boundary description information.

[0112] The timestamp information corresponding to the probability peak point refers to the timestamp of the sensing data unit corresponding to the probability peak point detected in step S410. The continuous change trend of the predicted probability of the sensing data units before and after the probability peak point refers to the change in the predicted probability value of the sensing data units before and after the probability peak point, such as whether it gradually increases or gradually decreases. The start and end times of the emergency event refer to the times when the emergency event begins and ends. The time boundary description information is a detailed description of the start and end times of the emergency event.

[0113] Based on the timestamp information corresponding to the probability peak point, and combined with the continuous trend of predicted probability changes in the preceding and following sensing data units, the start and end times of the emergency event are determined, and time boundary description information is generated. For example, this can be implemented as follows: Starting from the probability peak point, trace back the predicted probability values ​​of the sensing data units to find the point where the predicted probability value begins to rise significantly, and take this as the start time point of the emergency event. A significant rise can be determined by setting an increase threshold, for example, the predicted probability value rising by more than 0.2 within several consecutive timestamps. Then, starting from the probability peak point, trace back the predicted probability values ​​of the sensing data units to find the point where the predicted probability value begins to fall significantly, and take this as the end time point of the emergency event. A significant fall can also be determined by setting a fall threshold, for example, the predicted probability value falling by more than 0.2 within several consecutive timestamps. Finally, record the start and end times to generate time boundary description information.

[0114] Step S430: Extract the environmental sensing data of the sensing data unit corresponding to the probability peak point, and determine the initial occurrence area identifier and diffusion direction description of the emergency event in the sensing data unit.

[0115] The environmental sensing data of the sensing data unit corresponding to the probability peak point refers to the environmental sensing data contained in the sensing data unit corresponding to the probability peak point detected in step S410, such as temperature, humidity, and air pressure. The initial occurrence area identifier refers to the identifier of the area where the emergency event begins, such as a street block or building number. The diffusion direction description refers to the direction in which the emergency event spreads from the initial occurrence area to other areas, such as east, west, or southeast.

[0116] Extracting environmental sensing data from the sensing data unit corresponding to the probability peak point to determine the initial occurrence area and spread direction of the emergency event within that sensing data unit can be implemented as follows: Extract the environmental sensing data from the sensing data unit corresponding to the probability peak point. The corresponding sensing data unit can be located from a multi-source sensing data set based on timestamp information, and then the environmental sensing data within it can be extracted. Next, analyze the abnormal change areas in the environmental sensing data to determine the initial occurrence area of ​​the emergency event. For example, if the temperature data in a certain area shows an abnormal increase, and other environmental sensing data in that area also shows corresponding changes, then that area is marked as the initial occurrence area, and its identification information is obtained. Finally, by comparing environmental sensing data from adjacent timestamps, analyze the expansion of the abnormal change areas to determine the spread direction of the emergency event.

[0117] Step S440: Construct a trajectory model of the emergency event in the continuous sensing data unit based on the initial occurrence area identification, diffusion direction description, and time boundary description information.

[0118] The development trajectory model is a model used to describe the development and evolution of an emergency event within a continuous sensing data unit. It can predict the future development trend and location of the emergency event. Initial occurrence area identification, diffusion direction description, and time boundary description information are the basic data for constructing the development trajectory model, providing the starting location, development direction, and time range of the emergency event, respectively.

[0119] In one implementation, step S440 may specifically include the following steps S441 to S445:

[0120] Step S441: Extract the historical development status sequence of emergency events from the continuous sensing data unit. The historical development status sequence includes historical timestamps, historical area identifiers, and descriptions of historical diffusion directions.

[0121] Continuous sensing data units refer to multiple sensing data units that are consecutive in time, recording the status information of an emergency event at different points in time. The historical development state sequence is a sequence of information extracted from the continuous sensing data units regarding the development state of the emergency event over past time, including historical timestamps, historical area identifiers, and historical diffusion direction descriptions. Historical timestamps are the time markers for the emergency event at different points in the past; historical area identifiers are the identifiers of the areas affected by the emergency event at different points in the past; and historical diffusion direction descriptions are the diffusion directions of the emergency event at different points in the past.

[0122] Extracting the historical development sequence of emergency events from continuous sensing data units can be implemented as follows: First, determine the time range of the continuous sensing data unit based on time boundary description information. Then, within this time range, extract relevant information for each sensing data unit from the multi-source sensing data set. For each sensing data unit, extract its timestamp as a historical timestamp, determine the area affected by the emergency event as a historical area identifier based on environmental sensing data and event identification results, and determine the historical diffusion direction description by comparing the area changes of adjacent timestamps. Finally, assemble this information into a historical development sequence in chronological order.

[0123] Step S442: Calculate the state transition probabilities of adjacent timestamps in the historical development state sequence, and construct a Markov state transition matrix that reflects the evolution law of emergency event state.

[0124] State transition probability refers to the probability that an emergency event will transition from one state to another at adjacent time stamps. A Markov state transition matrix is ​​a matrix whose elements represent the transition probabilities between different states of an emergency event, reflecting the patterns of state evolution.

[0125] To construct a Markov state transition matrix reflecting the evolution of emergency event states, the state transition probabilities of adjacent timepoints in a historical development state sequence are statistically analyzed. This can be implemented as follows: Define the state space of the emergency event, for example, by combining different regional identifiers and diffusion directions as different states. Then, traverse the historical development state sequence, counting the number of state transitions between adjacent timepoints. For example, the number of times state A (street A, eastward diffusion) transitions to state B (street B, eastward diffusion). Next, calculate the probability of each state transition, i.e., the number of transitions divided by the total number of occurrences of that state. Finally, these state transition probabilities are assembled into a Markov state transition matrix. For example, if state A appears 10 times in the historical development state sequence, and 8 of those transitions lead to state B, then the probability of transitioning from state A to state B is 0.8. Filling the matrix with the transition probabilities between all states yields the Markov state transition matrix.

[0126] Where possible, a verification process can also be included, which involves calculating the autocorrelation coefficient of the state sequence. If the correlation coefficient with a lag of 2 or higher is significantly non-zero (e.g., |ρ|>0.3), then a higher-order Markov model or RNN is used instead.

[0127] Step S443: Input the initial occurrence region identifier, diffusion direction description and time boundary description information as initial state parameters into the Markov state transition matrix.

[0128] The initial occurrence area identifier, diffusion direction description, and time boundary description information are the starting information of the emergency event determined in steps S430 and S420. The initial state parameters refer to a set of parameters combining these starting information, used as input to the Markov state transition matrix to begin predicting the development state of the emergency event.

[0129] Using the initial occurrence area identifier, diffusion direction description, and time boundary description as initial state parameters, and inputting them into the Markov state transition matrix, can be implemented as follows: The initial occurrence area identifier and diffusion direction description are combined into an initial state. For example, if the initial occurrence area identifier is block A and the diffusion direction is eastward, then the initial state is (block A, eastward diffusion). Then, based on the state space of the Markov state transition matrix, the matrix row corresponding to the initial state is found. Finally, the initial state is used as the input to the matrix to begin calculating the state transition. For example, in the Markov state transition matrix, the row corresponding to (block A, eastward diffusion) is found, and the state at the next timestamp is predicted based on the transition probability of that row.

[0130] Step S444: Calculate the possible development states of the emergency event in the subsequent sensing data unit by using the state transition probability, and generate a state prediction set containing multiple path development possibilities.

[0131] The state transition probability is the probability of an emergency event transitioning between different states, recorded in the Markov state transition matrix. Subsequent sensing data units refer to sensing data units after the end of the time boundary description information. Possible development states refer to the possible states of the emergency event in subsequent sensing data units, such as different combinations of area identifiers and diffusion directions. The state prediction set is a collection containing the multiple possible development paths of the emergency event in subsequent sensing data units; it shows the possible development paths of the emergency event and their corresponding probabilities.

[0132] Calculating the possible development states of an emergency event in subsequent sensing data units using state transition probabilities, and generating a state prediction set containing multiple development possibilities, can be implemented as follows: Starting from the initial state, based on the state transition probabilities of the Markov state transition matrix, calculate the possible states and their probabilities for the next time stamp. For example, if the initial state is (street A, eastward diffusion), the probability of transitioning from this state to (street B, eastward diffusion) in the Markov state transition matrix is ​​0.8, and the probability of transitioning to (street C, southward diffusion) is 0.2. Then, the possible states for the next time stamp are (street B, eastward diffusion) and (street C, southward diffusion), with corresponding probabilities of 0.8 and 0.2, respectively. Then, for each possible state, continue to calculate the possible states and their probabilities for the next time stamp based on the state transition probabilities, and so on, until a predetermined number of time steps are reached. Finally, all possible development states and their corresponding probabilities are combined into a state prediction set.

[0133] Step S445: Perform confidence screening on the state prediction set, retain the development state paths whose predicted probabilities exceed a preset threshold, and generate a development trajectory model.

[0134] Confidence filtering involves selecting development path predictions from the state prediction set, retaining only those paths with higher prediction probabilities to improve the reliability and accuracy of the development trajectory model. A preset threshold is a pre-defined probability standard used to filter development path predictions. The development trajectory model, obtained after confidence filtering, contains the main development trajectories of emergency events within the continuous sensing data unit.

[0135] The confidence level of the state prediction set is filtered to retain development state paths whose predicted probabilities exceed a preset threshold, generating a development trajectory model. This can be implemented as follows: A preset threshold, such as 0.5, is set. Then, each development state path in the state prediction set is traversed, and its predicted probability is checked against the preset threshold. Paths with predicted probabilities exceeding the threshold are retained, while those with predicted probabilities below the threshold are discarded. Finally, the retained development state paths are organized according to chronological order and logical relationships to generate a development trajectory model. For example, in a state prediction set with three development state paths, two paths have predicted probabilities of 0.6 and 0.8, exceeding the preset threshold of 0.5, while the other path has a predicted probability of 0.3, below the preset threshold. The first two paths are retained, and they are connected chronologically to generate a development trajectory model for the emergency event. This model can reliably demonstrate the development trend of the emergency event within the continuous sensing data unit.

[0136] Step S450: Generate spatiotemporal feature information containing regional coverage and development trend predictions through the development trajectory model.

[0137] Regional coverage refers to the size of the geographical area affected by an emergency at different points in time, while development trend prediction is a forecast of the future direction and speed of the emergency. Spatiotemporal characteristic information integrates the temporal and spatial characteristics of the emergency, providing a more comprehensive description of its development.

[0138] In one implementation, step S450 may specifically include the following steps S451 to S456:

[0139] Step S451: Perform density clustering analysis on the development trajectory identifier sequence output by the development trajectory model to identify the boundaries of dense regions where trajectory points are distributed, and calculate the coverage parameters based on the center identifiers of the dense region boundaries to generate a regional coverage description with the center identifiers and coverage parameters as elements.

[0140] The development trajectory identifier sequence is a series of identifiers output by the development trajectory model, representing the location and status of an emergency event at different points in time. Density clustering analysis is a data clustering method that divides data points into different clusters based on their density. A dense region of trajectory point distribution refers to an area in the development trajectory identifier sequence where trajectory points are relatively concentrated. The boundary of a dense region refers to the edge limit of these dense regions. The center identifier is the identifier of the center location of a dense region, and the coverage parameter is a parameter used to describe the size of the dense region, such as radius or area. Density clustering analysis of the development trajectory identifier sequence output by the development trajectory model can be performed using the DBSCAN (Density-Based Spatial Clustering of Applications with Noise) algorithm. The basic idea of ​​the DBSCAN algorithm is to divide data points into core points, boundary points, and noise points by defining a neighborhood radius and a minimum number of points. A core point is a data point whose neighborhood contains at least the minimum number of points; a boundary point is a data point whose neighborhood contains fewer points than the minimum number of points but intersects with the neighborhood of a core point; and a noise point is a data point that is neither a core point nor a boundary point. Specifically, first, a neighborhood radius and a minimum number of points are set. Then, each data point in the development trajectory identifier sequence is traversed, and the number of data points in its neighborhood is calculated. If the number of data points in a data point's neighborhood is greater than or equal to the minimum number of points, it is marked as a core point; if the number of data points in a data point's neighborhood is less than the minimum number of points, but intersects with the neighborhood of a core point, it is marked as a boundary point; otherwise, it is marked as a noise point. Next, all core points and their neighboring boundary points are merged into a cluster, thus obtaining different dense regions. For each dense region, its boundary and center identifier are calculated. The center identifier can be obtained by calculating the average coordinates of all data points within the dense region. Finally, coverage parameters are calculated based on the boundaries of the dense regions; for example, the radius of the circumcircle of the dense region can be used as the coverage parameter. The center identifier and coverage parameters are combined to generate a description of the region's coverage area. For example, for a sequence of fire event development trajectories, the DBSCAN algorithm identifies several dense areas, calculates the center coordinates and circumscribed circle radius of each dense area, and generates a description of the area coverage such as "Center coordinates: (x1, y1), coverage radius: r1; Center coordinates: (x2, y2), coverage radius: r2".

[0141] Step S452: Extract the set of directional markers for the development trajectory marker sequence under continuous timestamps, calculate the angular change rate and marker movement gradient of the directional markers of adjacent timestamps based on the time interval of each unit, and construct a development trend direction stability evaluation function based on time units.

[0142] The direction identifier set refers to the identifier of the emergency event's development direction corresponding to each timestamp in the development trajectory identifier sequence. The angle change rate is the ratio of the angle change between direction identifiers of adjacent timestamps to the time interval, reflecting the speed of change in the emergency event's development direction. The identifier movement gradient is the ratio of the displacement between development trajectory identifiers of adjacent timestamps to the time interval, reflecting the speed of the emergency event's development. The development trend direction stability assessment function is a function used to evaluate the stability of the emergency event's development direction; it uses time units as a benchmark and comprehensively considers both the angle change rate and the identifier movement gradient.

[0143] Extracting the set of directional markers for a development trajectory marker sequence at consecutive timestamps can be determined by analyzing the positional changes of adjacent timestamps within the sequence. For example, for a development trajectory marker sequence of an emergency event, each marker contains the positional coordinates of the emergency event at a specific timestamp. By calculating the vector between the positional coordinates of adjacent timestamps, the direction of movement of the emergency event within each time interval is obtained, and these directions are used as directional markers. Calculating the rate of change of the angle of the directional markers of adjacent timestamps based on the time interval of each unit can be implemented, for example, by calculating the angle between the directional markers of adjacent timestamps. The cosine of the angle between two vectors can be calculated using the dot product formula, and then the angle can be obtained using the inverse cosine function. Next, the angle is divided by the time interval to obtain the rate of change of the angle. For the marker movement gradient, the displacement distance between the development trajectory markers of adjacent timestamps is calculated, and the displacement distance is divided by the time interval to obtain the marker movement gradient.

[0144] Constructing a stability evaluation function for the development trend direction based on time units can comprehensively consider the rate of change of angle and the gradient of the marker movement. For example, after normalizing them separately to eliminate dimensions, a weighted sum can be performed.

[0145] Step S453: Perform sliding window weighted averaging on the direction identifier set based on the direction stability evaluation function to generate a trend prediction direction sequence after direction smoothing, and calculate the cumulative slope change of the trend prediction direction sequence within a preset time window.

[0146] The sliding window weighted average reduces data fluctuations and makes the direction indicators smoother by sliding a fixed-size window across the set of direction indicators and applying a weighted average to the data within the window (which can be understood as normalizing to eliminate dimensional differences before weighting). The trend prediction direction sequence is the direction sequence obtained after processing with the sliding window weighted average, and can more accurately reflect the development trend of the emergency event. The cumulative slope change refers to the sum of the slope changes of the trend prediction direction sequence within a preset time window, which can reflect the degree of change in the development trend of the emergency event.

[0147] The sliding window weighted average processing of the direction identifier set based on the direction stability evaluation function can be performed as follows: First, determine the size w of the sliding window. Then, starting from the first element of the direction identifier set, select windows of length w sequentially. For each direction identifier within the window, calculate the weight of each identifier according to the direction stability evaluation function. The weight can be adjusted according to the value of the direction stability evaluation function; for example, the smaller the value of the direction stability evaluation function, the more stable the direction, and the corresponding weight can be set larger. Next, perform a weighted average of the direction identifiers within the window to obtain the smoothed direction identifier at the center position of the window. Combine the smoothed direction identifiers at the center positions of all windows to generate the direction-smoothed trend prediction direction sequence.

[0148] Calculating the cumulative slope change of a trend prediction direction sequence within a preset time window can be implemented as follows: First, determine the size of the preset time window, T. Then, select a subsequence of length T within the trend prediction direction sequence. For each adjacent data point in the subsequence, calculate its slope change. The slope change can be obtained by calculating the ratio of the direction difference between adjacent data points to the time interval. Finally, sum all the slope changes within the subsequence to obtain the cumulative slope change. For example, in a trend prediction direction sequence for a flood event, with a preset time window of 10 time units, select a subsequence of length 10, calculate the directional slope change of adjacent data points within this subsequence, and sum these slope changes to obtain the cumulative slope change. This cumulative amount reflects the degree of change in the trend direction of the flood within these 10 time units.

[0149] Step S454: Construct a dynamic region expansion model based on the coverage parameters and the cumulative slope change of the trend prediction direction sequence to predict the expansion direction and expansion rate of the region coverage in subsequent sensing data units.

[0150] The dynamic area expansion model is used to describe the dynamic changes in the coverage area of ​​an emergency event over time. It comprehensively considers the cumulative slope changes of the coverage parameter and the trend prediction direction sequence. The expansion direction refers to the direction in which the coverage area of ​​the emergency event expands in subsequent sensing data units, and the expansion rate description refers to the speed at which the coverage area of ​​the emergency event expands, such as the increase in coverage area per unit time.

[0151] The following method can be used to construct a dynamic regional expansion model based on the coverage parameter and the cumulative slope change of the trend prediction direction sequence. First, analyze the relationship between the coverage parameter and the cumulative slope change of the trend prediction direction sequence. Generally, the larger the cumulative slope change, the more drastic the change in the development trend of the emergency event, and the more unstable the expansion direction of the regional coverage area may be; the larger the coverage parameter, the larger the current impact range of the emergency event, and the expansion rate may also be affected. This relationship can be described by establishing a regression model. For example, assume a linear relationship between the expansion rate v and the coverage parameter r and the cumulative slope change s: v = a × r + b × s + c, where a, b, and c are coefficients to be determined (it can be understood that before adding, r and s are normalized, for example, to [0,1] to eliminate the problem of dimensional differences). This regression model can be fitted using historical data to determine the values ​​of the coefficients. The expansion direction can be determined based on the trend prediction direction sequence. The average direction of the trend prediction direction sequence can be used as the basis for the expansion direction, while considering the influence of the cumulative slope change on the expansion direction. If the cumulative slope change is large, a certain random offset can be made based on the average direction to reflect the uncertainty of the expansion direction. Alternatively, an XGBoost regression model can be trained using historical data, with inputs r and s to predict v.

[0152] The expansion direction and rate of the predicted area coverage in subsequent sensing data units can be calculated by inputting the current coverage parameters and the cumulative slope change of the trend prediction direction sequence into a dynamic area expansion model. For example, for a fire event, given the current coverage radius r and the cumulative slope change s, these can be substituted into the regression model v = a × r + b × s + c to calculate the expansion rate v. Simultaneously, the expansion direction is determined based on the trend prediction direction sequence, generating an expansion direction and rate description such as "Expansion direction: 30° east of south, expansion rate: 2 square kilometers per hour".

[0153] Step S455: Describe the expansion direction and expansion rate into the trajectory prediction function, generate a probability distribution map containing the trajectory path within the subsequent time window, and extract the largest connected region in the probability distribution map with a confidence level exceeding a preset threshold as the boundary of the predicted development path.

[0154] The trajectory prediction function is a function that predicts the future trajectory path of an emergency event based on its expansion direction and rate. The probability distribution map, output by the trajectory prediction function, represents the possible trajectory paths of the emergency event within subsequent time windows and their corresponding probability distributions. The confidence level refers to the probability value corresponding to each region in the probability distribution map, indicating the likelihood that that region represents the actual trajectory path of the emergency event. The preset threshold is a pre-defined probability value standard used to filter out regions with high confidence levels. The largest connected region identifier is the identifier of the largest interconnected region in the probability distribution map whose confidence level exceeds the preset threshold. The predicted development path boundary refers to the boundary of this largest connected region, representing the range of the most likely development path of the emergency event within subsequent time windows.

[0155] The Monte Carlo simulation method can be used to describe the expansion direction and rate as inputs to the trajectory prediction function. The basic idea of ​​Monte Carlo simulation is to simulate the development of an emergency event through random sampling. Specifically, the possible movement range of the emergency event at each time step is first determined based on the expansion direction and rate. For example, if the expansion direction is 30° east of south and the expansion rate is 2 square kilometers per hour, then the emergency event may move a certain distance in that direction within each hourly time step. Then, random sampling is performed within the possible movement range to generate multiple possible trajectory paths. For each trajectory path, its probability of occurrence is calculated. The probability values ​​can be adjusted based on the uncertainties in the expansion rate and direction. All possible trajectory paths and their probabilities are combined to generate a probability distribution map.

[0156] Extracting the identifier of the largest connected region with a confidence level exceeding a preset threshold from the probability distribution map can be implemented as follows: First, set a preset threshold. Then, traverse each region in the probability distribution map, marking regions with confidence levels exceeding the preset threshold as valid regions. Next, use a connected component analysis algorithm, such as Depth-First Search (DFS) or Breadth-First Search (BFS), to find all interconnected valid regions. Finally, select the region with the largest area from these connected regions and extract its identifier as the boundary of the predicted development path. For example, for a probability distribution map of an earthquake event, with a preset threshold of 0.6, by traversing and analyzing the connected component, find the largest connected region with a confidence level exceeding 0.6, whose boundary identifier is "boundary coordinates: [(x1,y1),(x2,y2),…,(xn,yn)]". This boundary identifier represents the range of the most likely development path of the earthquake within the subsequent time window.

[0157] Step S456: Based on the description of regional coverage and the predicted development path boundary, generate spatiotemporal feature information containing spatial coverage identifier sequences and trend direction change vectors. The spatial coverage identifier sequences are used to mark the real-time coverage area of ​​the emergency event in the sensing data unit, and the trend direction change vectors are used to describe the predicted offset angle and velocity changes of the emergency event's development direction.

[0158] The regional coverage description, generated in step S451, includes the center identifier and coverage parameters of the emergency event at different time points. The predicted development path boundary, extracted in step S455, represents the range of the most likely development path of the emergency event within subsequent time windows. The spatial coverage identifier sequence is a series of identifiers used to mark the real-time coverage area of ​​the emergency event in the sensing data unit, and can be generated based on the regional coverage description and the predicted development path boundary. The trend direction change vector is a vector whose direction represents the predicted offset angle of the emergency event's development direction, and whose magnitude represents the change in the emergency event's development speed.

[0159] Generating a spatial coverage identifier sequence based on regional coverage description and predicted development path boundaries can be implemented as follows: First, determine the coverage area of ​​the emergency event at the current time point based on the regional coverage description. Then, combine this with the predicted development path boundaries to predict the coverage area of ​​the emergency event at subsequent time points. The position and size of the coverage area can be updated progressively based on the expansion direction and expansion rate descriptions. The coverage area identifiers at each time point are combined to generate a spatial coverage identifier sequence. For example, for a typhoon event, the current center position and coverage radius of the typhoon are obtained based on the regional coverage description. Combined with the predicted development path boundaries, the movement path and coverage changes of the typhoon in the next few hours are predicted, generating a spatial coverage identifier sequence such as "Time: t1, Center coordinates: (x1, y1), Coverage radius: r1; Time: t2, Center coordinates: (x2, y2), Coverage radius: r2;...". Generating a trend direction change vector can be achieved by analyzing the trend prediction direction sequence and the expansion rate description. First, calculate the direction difference between the trend prediction direction sequence at the current time point and subsequent time points to obtain the predicted offset angle. Then, calculate the change in the development speed of the emergency event based on the expansion rate description. The predicted offset angle and velocity changes are combined into a vector to serve as the trend direction change vector. For example, for a forest fire event, if the trend prediction direction sequence shows that the fire's development direction will shift from due east to 20° east of south over a future period, and the spread rate increases from 1 square kilometer per hour to 1.5 square kilometers per hour, then the trend direction change vector can be represented as "direction offset: 20°, velocity change: 0.5 square kilometers per hour". Combining the spatial cover identifier sequence and the trend direction change vector generates spatiotemporal feature information containing both the spatial cover identifier sequence and the trend direction change vector. This information can comprehensively describe the temporal and spatial development of the emergency event.

[0160] Step S500: Generate an emergency warning instruction containing an event location identifier based on the emergency event type and spatiotemporal feature information, and send the emergency warning instruction to the target emergency response terminal to initiate response operations.

[0161] The emergency event type refers to the type of emergency event existing in the target area as determined in step S400, such as fire, earthquake, flood, etc. Spatiotemporal characteristic information is feature information generated in step S450, including information such as area coverage and development trend prediction. Event location identifiers are identifiers used to accurately identify the location and scope of the emergency event. Emergency warning instructions are instructions generated based on the emergency event type and spatiotemporal characteristic information, used to convey relevant information about the emergency event to the target emergency response terminal and initiate response operations. The target emergency response terminal refers to the terminal equipment responsible for responding to and handling emergency events, such as the command center of a fire department or the emergency system of an earthquake monitoring department.

[0162] In one implementation, step S500 may specifically include the following steps S510-S560:

[0163] Step S510: Parse the set of preset early warning rules corresponding to the emergency event type, and extract the early warning level identifier and response strategy code associated with the emergency event type.

[0164] A pre-defined set of early warning rules is a collection of pre-set early warning rules for different types of emergency events. Each rule includes information such as the early warning level and response strategy corresponding to that type of emergency event. Early warning level identifiers are used to distinguish the severity of emergency events, such as Level 1 early warning, Level 2 early warning, etc. Response strategy codes are codes that encode emergency response strategies; different codes represent different response measures, such as evacuating residents and allocating relief supplies.

[0165] Analyzing the set of preset early warning rules corresponding to different emergency event types can be implemented as follows: Establish a mapping relationship between emergency event types and preset early warning rule sets. For example, map fire events to a fire early warning rule set, earthquake events to an earthquake early warning rule set, and so on. Then, based on the determined emergency event type, find the corresponding preset early warning rule set. Within the preset early warning rule set, determine the early warning level identifier and response strategy code based on the specific circumstances of the emergency event, such as the regional coverage and development trend in spatiotemporal characteristic information. For example, for a fire event, if its regional coverage is large and its development trend is rapid, the corresponding early warning level identifier might be Level 1, and the response strategy code might be "evacuate residents + allocate a large amount of fire-fighting resources".

[0166] Step S520: Based on the regional coverage range in the spatiotemporal feature information, calculate the center identifier and boundary extension parameters of the coverage area of ​​the emergency event within the current sensing data unit.

[0167] Regional coverage is information describing the geographical area affected by an emergency event at different points in time within spatiotemporal feature information. The coverage area center identifier refers to the identifier of the center location of the coverage area within the current sensing data unit, such as coordinates. Boundary expansion parameters are parameters used to describe the expansion of the emergency event's coverage area boundaries, such as expansion speed and direction.

[0168] The center identifier of the coverage area can be calculated by averaging the coordinates of all points within the coverage area based on the spatiotemporal feature information. For example, if the coverage area is a polygonal region represented by a series of coordinate points, the x and y coordinates of all points within the polygon are added together, and then divided by the number of coordinate points to obtain the center coordinates of the coverage area. Boundary expansion parameters can be determined based on the expansion direction and rate described in the spatiotemporal feature information. For example, if the spatiotemporal feature information gives an expansion direction of 30° east of south and an expansion rate of 2 square kilometers per hour, this information can be used as boundary expansion parameters to describe that the boundary of the emergency event coverage area expands in that direction at a rate of 2 square kilometers per hour.

[0169] Step S530: Extract the development trend prediction from the spatiotemporal feature information, combine it with the center identifier of the coverage area, and construct a dynamic development direction prediction function based on the linear fitting model of the historical development trajectory.

[0170] Development trend prediction is the prediction of the future direction and speed of an emergency event based on spatiotemporal characteristics. Historical development trajectory refers to the record of the development path of an emergency event over past time. A linear fitting model is a model that fits historical development trajectory data using linear equations; it can identify linear trends in the data. The dynamic development direction prediction function is a function constructed based on development trend prediction, coverage area center identification, and the linear fitting model of historical development trajectories; it is used to predict the development direction of an emergency event at different points in the future.

[0171] Predicting development trends from spatiotemporal feature information can be achieved by directly extracting relevant information such as trend prediction direction sequences, expansion directions, and expansion rate descriptions. Constructing a dynamic development direction prediction function by combining the center identifier of the coverage area and a linear fitting model of historical development trajectories can be implemented as follows: Linear regression is used to fit historical development trajectory data, resulting in a linear fitting equation. The basic idea of ​​linear regression is to find the optimal fitted line by minimizing the sum of squared distances from data points to the fitted line. For example, for the location coordinates of an emergency event in the past, linear regression can find a linear equation y = ax + b, where x represents time and y represents location coordinates. Then, the parameters of the linear fitting equation are adjusted based on the development trend prediction and the center identifier of the coverage area. For example, if the development trend prediction indicates that the development direction of the emergency event will change, the slope and intercept of the linear fitting equation can be adjusted according to the angle and magnitude of the change. Finally, the adjusted linear fitting equation is used as a dynamic development direction prediction function to predict the development direction of the emergency event at different future time points.

[0172] Step S540: Input the boundary extension parameters into the predefined coordinate mapping rules to generate a standardized set of location identifiers consistent with the geographic identification system of the target area.

[0173] Predefined coordinate mapping rules are pre-defined rules used to convert boundary extension parameters into coordinate representations consistent with the target area's geographic identification system. The target area's geographic identification system is a system used to identify geographical locations within the target area, such as latitude and longitude coordinate systems or Geographic Information Systems (GIS). A standardized set of location identifiers is a series of location identifiers obtained after transformation using coordinate mapping rules. These identifiers have a unified representation within the target area's geographic identification system, facilitating subsequent processing and use.

[0174] The steps for inputting boundary expansion parameters into predefined coordinate mapping rules are as follows: First, determine the predefined coordinate mapping rules. These rules can be designed based on the characteristics of the target area's geographic identification system (GIS), for example, converting the boundary expansion distance in kilometers into changes in latitude and longitude coordinates. Then, convert information such as the expansion speed and direction in the boundary expansion parameters into coordinate changes. For example, if the boundary expansion parameters indicate an eastward expansion at a speed of 2 kilometers per hour, the distance of 2 kilometers per hour is converted into latitude and longitude changes according to the target area's GIS coordinate transformation rules. Finally, apply these coordinate changes to the center marker of the covered area, generating a series of standardized location markers consistent with the target area's GIS. These location markers are then combined to form a standardized location marker set. For example, for a city flood event, the boundary expansion parameters are converted into changes in latitude and longitude coordinates using coordinate mapping rules, and these changes are applied to the center latitude and longitude coordinates of the flood-covered area, generating a standardized location marker set such as "(longitude 116.40, latitude 39.90), (longitude 116.41, latitude 39.90),...".

[0175] Step S550: Perform time dimension interpolation processing on the standardized location identifier set according to the dynamic development direction prediction function to generate a predicted identifier sequence of emergency event development paths within a subsequent preset time period.

[0176] Time-dimensional interpolation, based on a standardized set of location identifiers, inserts more location identifiers into subsequent preset time periods according to a dynamic development direction prediction function, thus providing a more detailed description of the emergency event's development path. The preset time period is a pre-defined time range, such as the next 24 hours or the next week. The predicted identifier sequence is a series of location identifiers generated after time-dimensional interpolation, representing the predicted development path of the emergency event within the subsequent preset time period.

[0177] In one implementation, step S550 may specifically include the following steps S551 to S555:

[0178] Step S551: Extract the location markers of adjacent timestamps from the standardized location marker set, and construct a two-dimensional coordinate point set containing timestamps and location markers.

[0179] The location markers in the standardized location marker set are location markers generated in step S540 that are consistent with the geographic identification system of the target area, and each location marker corresponds to a timestamp. The two-dimensional coordinate point set is a set of two-dimensional coordinate points composed of timestamps and location markers, where the horizontal axis represents the timestamp and the vertical axis represents the location marker.

[0180] Extracting location markers with adjacent timestamps from a standardized location marker set can be done by traversing the set and selecting adjacent markers sequentially according to their timestamps. For example, if the standardized location marker set contains location markers p1, p2, and p3 corresponding to timestamps t1, t2, and t3, then the adjacent timestamp marker pairs (p1, p2) and (p2, p3) can be extracted. Each location marker and its corresponding timestamp are combined into a two-dimensional coordinate point, such as (t1, p1), (t2, p2), and (t3, p3). All these two-dimensional coordinate points are then combined to construct a two-dimensional coordinate point set containing both timestamps and location markers.

[0181] Step S552: Use the cubic spline interpolation algorithm to perform curve fitting on the two-dimensional coordinate point set to generate an interpolation curve that reflects the continuity of the emergency event development path.

[0182] Cubic spline interpolation is an algorithm used for curve fitting. By constructing a cubic polynomial function between each adjacent data point, the entire curve has continuous first and second derivatives at the data points, thus ensuring the smoothness and continuity of the curve. The interpolated curve is obtained by fitting a set of two-dimensional coordinate points using cubic spline interpolation, and it can reflect the continuity of the development path of an emergency event.

[0183] Curve fitting of a two-dimensional coordinate point set using cubic spline interpolation can be implemented as follows: The two-dimensional coordinate point set is sorted according to the order of its x-coordinate (timestamp). Then, based on the principle of cubic spline interpolation, a cubic polynomial function is constructed between each adjacent data point. The general form of the cubic polynomial function is S(x) = a + bx + cx. 2 +dx 3Here, x is the timestamp, and a, b, c, and d are the coefficients of the polynomial. By satisfying the continuity conditions of adjacent polynomials at data points (equal function values, equal first derivatives, equal second derivatives) and boundary conditions (such as known first or second derivatives at the endpoints), the coefficients of the cubic polynomial for each interval can be solved. Finally, the cubic polynomials of all intervals are combined to form an interpolation curve. For example, for a two-dimensional coordinate point set of a hurricane event, an interpolation curve is constructed using a cubic spline interpolation algorithm. This curve can smoothly connect the location markers at each time point, reflecting the continuity of the hurricane's development path.

[0184] Step S553: ​​Determine the curvature adjustment parameters of the interpolation curve based on the dynamic development direction prediction function, and adaptively correct the local curvature of the interpolation curve.

[0185] Curvature adjustment parameters are used to adjust the curvature of the interpolation curve, and they can be determined based on the dynamic development direction prediction function. Local curvature refers to the degree of curvature of the interpolation curve in a specific local region. By adaptively correcting the local curvature, the interpolation curve can better reflect the actual development path of the emergency event.

[0186] Determining the curvature adjustment parameters of the interpolation curve based on the dynamic development direction prediction function can be implemented by analyzing the changes in the dynamic development direction prediction function at different time points. The dynamic development direction prediction function provides information on the development direction of an emergency event at different future time points, such as the angle and speed of change in direction. If the dynamic development direction prediction function indicates a significant change in the development direction of the emergency event at a certain time point, it means that the local curvature of the interpolation curve near that time point should be relatively large; conversely, if the change in development direction is small, the local curvature can be relatively small. Then, the curvature adjustment parameters are determined based on the changes in the dynamic development direction prediction function. A mapping relationship can be established to map the changes in the dynamic development direction prediction function to the curvature adjustment parameters. Finally, the curvature adjustment parameters are used to correct the local curvature of the interpolation curve. The adjustment of the local curvature can be achieved by adjusting the coefficients of the cubic polynomial of the interpolation curve in the corresponding region.

[0187] Step S554: Sample uniformly on the interpolation curve at preset time intervals to generate predicted positioning identifiers for each time point within the subsequent preset time period.

[0188] The preset time interval is a pre-defined time interval, such as every hour or every half day. The predicted location marker is a location marker obtained by sampling on the interpolation curve at the preset time interval, representing the predicted location of the emergency event at each time point within the subsequent preset time period.

[0189] Uniform sampling at preset time intervals on the interpolation curve can be implemented as follows: Determine the start and end times of subsequent preset time periods. Then, based on the preset time intervals, uniformly select a series of time points between the start and end times. For each selected time point, find the corresponding location marker on the interpolation curve. The coordinates of the corresponding location marker can be calculated by substituting the time points into the cubic polynomial function of the interpolation curve. Combine all these predicted location markers to generate a sequence of predicted location markers for each time point within the subsequent preset time period.

[0190] Step S555: Perform smoothing filtering on the predicted positioning markers to eliminate abnormal fluctuations in marker points generated during interpolation and generate a predicted marker sequence.

[0191] Smoothing filtering is a method used to remove abnormal fluctuations in data, making the data smoother and more stable. Abnormal fluctuation markers are those points that deviate from the normal development path during interpolation due to data noise, fitting errors, or other reasons. The predicted marker sequence is a series of predicted location markers obtained after smoothing filtering, which can more accurately reflect the development path of the emergency event within a subsequent preset time period.

[0192] Moving average filtering can be used to smooth predicted location markers. The basic idea of ​​moving average filtering is to smooth the data by calculating the average value within a certain window. Specifically, the size of the moving average window is first determined. The window size can be adjusted according to specific circumstances, for example, it can be set to 3 or 5. Then, the predicted location marker sequence is traversed. For each marker, a certain number of markers before and after it (the window size) are selected, and the average value of these markers is calculated. This average value is used as the smoothed value for that marker. The smoothed values ​​of all markers are combined to generate the smoothed predicted marker sequence. For example, for a predicted location marker sequence for a forest fire, if the moving average window size is set to 3, for each marker in the sequence, the average value of its two preceding and following markers and itself is calculated to obtain the smoothed predicted location marker sequence, eliminating abnormal fluctuations in markers generated during interpolation.

[0193] Step S560: Integrate the warning level identifier, response strategy code, and prediction identifier sequence to generate an emergency warning instruction containing an event location identifier chain with timestamp alignment. The representation of each identifier node in the event location identifier chain is consistent with the representation of the standardized location identifier set.

[0194] The warning level identifier is the warning level identifier associated with the emergency event type extracted in step S510. The response strategy code is the code of the corresponding response strategy. The predicted identifier sequence is the predicted location identifier sequence of the emergency event within a subsequent preset time period generated in step S555. Timestamp alignment refers to associating each identifier point in the predicted identifier sequence with its corresponding timestamp to ensure that each identifier point has clear time information. The event location identifier chain is a chain composed of a series of event location identifiers arranged in chronological order, with each identifier node representing the location of the emergency event at a certain point in time. The emergency warning instruction is an instruction that integrates the warning level identifier, response strategy code, and event location identifier chain, used to convey relevant information about the emergency event to the target emergency response terminal and initiate response operations.

[0195] In one implementation, step S560 may specifically include the following steps S561 to S566:

[0196] Step S561: Input the warning level identifier, response strategy code, and prediction identifier sequence into the instruction generation module, perform information association and matching processing, and generate the initial warning instruction content.

[0197] The instruction generation module is used to fuse and process warning level identifiers, response strategy codes, and prediction identifier sequences. Information association and matching processing refers to associating and matching the warning level identifiers, response strategy codes, and prediction identifier sequences, ensuring a clear correspondence between them in the initial warning instruction content. The initial warning instruction content is the instruction content generated after information association and matching processing; it contains relevant information about the warning level identifiers, response strategy codes, and prediction identifier sequences.

[0198] The information association and matching process of the warning level identifier, response strategy code, and predicted identifier sequence input into the instruction generation module can be implemented as follows: An association table of the warning level identifier, response strategy code, and predicted identifier sequence can be established in the instruction generation module. This association table records the correspondence between the warning level identifier, response strategy code, and each identifier point in the predicted identifier sequence. For example, for a Level 1 warning level identifier, the corresponding response strategy code is "Comprehensive Rescue + Evacuation of Residents," and this correspondence can be associated with each identifier point in the predicted identifier sequence. Then, based on the association table, the warning level identifier, response strategy code, and predicted identifier sequence are combined into the initial warning instruction content. The initial warning instruction content can adopt a structured format, such as "Warning Level: Level 1; Response Strategy: Comprehensive Rescue + Evacuation of Residents; Predicted Identifier Sequence: [(t1,p1),(t2,p2),…]".

[0199] Step S562: Perform timestamp alignment on the prediction identifier sequence in the initial warning instruction content to ensure that each prediction identifier node corresponds one-to-one with the corresponding timestamp information.

[0200] Timestamp alignment is the process of accurately associating each predicted identifier node in the predicted identifier sequence with its corresponding timestamp information. A predicted identifier node is each location identifier in the predicted identifier sequence, and the timestamp information is the time information corresponding to each predicted identifier node. Timestamp alignment makes the initial warning instructions more accurate and clear, facilitating understanding and processing by the target emergency response terminal.

[0201] The timestamp alignment process for the predicted identifier sequence in the initial warning instruction can be implemented as follows: Check whether each predicted identifier node in the predicted identifier sequence already contains the corresponding timestamp information. If the predicted identifier node only contains a location identifier but no timestamp information, the corresponding timestamp needs to be obtained from the time point information used during sampling in step S554. Then, the timestamp information is associated with the predicted identifier node to ensure that each predicted identifier node has a unique corresponding timestamp. For example, if the predicted identifier sequence is [p1, p2, p3] and the corresponding timestamp information is [t1, t2, t3], they are combined into the form [(t1, p1), (t2, p2), (t3, p3)] to achieve timestamp alignment.

[0202] Step S563: Extract key information elements from the initial warning instruction content, including the warning level identifier, response strategy code, and prediction identifier sequence after timestamp alignment.

[0203] Key information elements are the most important information in the initial warning instruction, including the warning level identifier, response strategy code, and the predicted identifier sequence after timestamp alignment. This information plays a crucial role in enabling the target emergency response terminal to perform emergency response operations.

[0204] Key information elements can be extracted from the initial warning instruction by parsing its structured format. For example, if the initial warning instruction is in the format "Warning Level: Level 1; Response Strategy: Full-scale Rescue + Evacuation of Residents; Predictive Identifier Sequence: [(t1,p1),(t2,p2),...]", then string matching and segmentation operations can be used to extract "Level 1" as the warning level identifier, "Full-scale Rescue + Evacuation of Residents" as the response strategy code, and [(t1,p1),(t2,p2),...] as the timestamp-aligned predictive identifier sequence.

[0205] Step S564: Standardize the format of key information elements, unify the representation and data structure of information elements, and generate standardized early warning instruction content.

[0206] Format standardization is the process of unifying and standardizing the representation and data structure of key information elements. A unified representation and data structure makes early warning instructions easier for emergency response terminals to process and parse. Standardized early warning instructions are instructions generated after format standardization, possessing a unified format and data structure.

[0207] Standardizing the format of key information elements can be implemented as follows: First, determine a unified representation format and data structure. For example, warning level identifiers can be uniformly represented by numbers, such as 1 for Level 1 warning; response strategy codes can be encoded using unified coding rules; and prediction identifier sequences can be represented using a unified coordinate system and data type. Then, the key information elements are converted according to the unified representation format and data structure. For example, the warning level identifier "Level 1" is converted to the number 1, and the response strategy code "Comprehensive Rescue + Evacuation of Residents" is converted to its corresponding coded value. Finally, the converted key information elements are combined into standardized warning instruction content. For example, standardized warning instruction content can be represented in JSON format: {"Warning Level": 1, "Response Strategy Code": "001", "Prediction Identifier Sequence": [(t1, p1), (t2, p2), ...]}.

[0208] Step S565: Verify the content of the standardized early warning instruction, verify the completeness and logical consistency of the information elements, and generate an emergency early warning instruction that has passed the verification.

[0209] Verification is the process of checking and validating the content of standardized early warning instructions, primarily verifying the completeness and logical consistency of information elements. Completeness of information elements means that the standardized early warning instruction content contains all necessary information elements, such as early warning level identifiers, response strategy codes, and prediction identifier sequences. Logical consistency means that the relationships between information elements are logically sound; for example, the correspondence between early warning level identifiers and response strategy codes is correct. An emergency early warning instruction that passes verification is one that, after verification, has complete information elements and is logically consistent.

[0210] Validating standardized early warning instructions can be implemented as follows: First, check if the standardized early warning instruction content contains all necessary information elements. This can be achieved by checking the key-value pairs in the JSON format of the standardized early warning instruction content. For example, check if it contains keys such as "early warning level," "response strategy code," and "prediction identifier sequence." Then, verify the logical consistency between the information elements. This can be done by querying a predefined association table to verify the correct correspondence between the early warning level identifier and the response strategy code. For example, if the response strategy code corresponding to a level 1 early warning level identifier should be "001," then check if the standardized early warning instruction content conforms to this correspondence. If the information elements in the standardized early warning instruction content are complete and logically consistent, a valid emergency early warning instruction is generated; otherwise, the standardized early warning instruction content needs to be corrected.

[0211] Step S566: Encrypt the verified emergency warning command to generate a final emergency warning command containing encrypted information.

[0212] Encryption processing involves encrypting the verified emergency warning instructions to ensure the security and confidentiality of the instruction content. Encrypted information is the information obtained after encryption, which prevents the instruction content from being stolen or tampered with during transmission. The final emergency warning instruction is an instruction containing encrypted information, which can be securely sent to the target emergency response terminal.

[0213] Encrypting verified emergency warning commands can be done using symmetric encryption algorithms, such as AES (Advanced Encryption Standard). AES is a symmetric-key encryption algorithm that uses the same key for both encryption and decryption. Specifically, an encryption key is first generated. This key can be a randomly generated binary sequence. Then, the verified emergency warning command is encrypted using the AES algorithm and the encryption key. This encryption process converts the verified emergency warning command into a string of ciphertext. Finally, the encrypted ciphertext is used as encrypted information and combined with necessary encryption metadata (such as the encryption algorithm identifier and key length) to form the final emergency warning command containing encrypted information. For example, the final emergency warning command can be in the following format: {"Encryption Algorithm":"AES","Key Length":128","Encrypted Information":"xxxxxx"}, where "xxxxxx" is the encrypted ciphertext.

[0214] Sending the final emergency warning instruction to the target emergency response terminal can be achieved through network communication technology. For example, a wireless communication network (such as 4G or 5G) can be used to send the final emergency warning instruction to the communication module of the target emergency response terminal. After receiving the final emergency warning instruction, the target emergency response terminal uses the corresponding decryption algorithm and key to decrypt the encrypted information, obtain the original emergency warning instruction content, and then initiates response operations, such as allocating rescue resources and organizing personnel evacuation, to deal with the emergency events existing in the target area.

[0215] It is understood that the various algorithms involved in the above descriptions of the embodiments of the present invention, such as word segmentation algorithms, filtering algorithms, cubic spline interpolation algorithms, symmetric encryption algorithms, etc., can all be obtained from relevant content in the prior art. To save space, they will not be elaborated on in the embodiments of the present invention. In addition, those skilled in the art can supplement the details based on common knowledge in the art when implementing the solutions of the present invention. For example, they can use normalization to eliminate dimensional conflicts before feature fusion, use interpolation to eliminate dimensional differences, reasonably set thresholds based on historical data, experience or business scenario requirements, train the model based on a general model training method, set the number of layers in the model structure based on actual needs, select activation functions, etc. The present invention will not describe the implementation process in a complicated manner.

[0216] Please see Figure 2 , Figure 2 This is a schematic diagram of a computer device provided in an embodiment of the present invention. The computer device includes at least a processor 101, a communication interface 102, and a memory 103. The processor 101, communication interface 102, and memory 103 can be connected via a bus or other means. The processor 101 (or Central Processing Unit, CPU) is the computing and control core of the computer device, capable of parsing various instructions and processing various data within the computer device. The communication interface 102 may optionally include a standard wired interface or a wireless interface (such as Wi-Fi, mobile communication interface, etc.), and can be used to send and receive data under the control of the processor 101; the communication interface 102 can also be used for data transmission and interaction within the computer device. The memory 103 is a storage device in the computer device used to store programs and data. It is understood that the memory 103 here can include the computer device's built-in memory, or it can include extended memory supported by the computer device. The memory 103 provides storage space, which stores the computer device's operating system; this invention does not limit the storage space.

[0217] In one embodiment, the processor 101 executes the emergency event identification method based on a neural network model provided in the above embodiments of the present invention by running a computer program in the memory 103.

Claims

1. An emergency event identification method based on a neural network model, characterized in that, The method includes: Acquire a multi-source sensing data set covering the target area. The multi-source sensing data set contains multiple sensing data units with timestamps. Each sensing data unit consists of environmental sensing data and interactive feedback data. The multi-source sensing data set is subjected to feature encoding processing to obtain a multi-dimensional feature representation set of the sensing data unit, the multi-dimensional feature representation set including temporal correlation features and semantic correlation features; The pre-trained emergency event recognition neural network model is invoked to perform event recognition processing on the multidimensional feature representation set, generating the event recognition result of the perception data unit, wherein the event recognition result includes the predicted probability distribution of different emergency event categories; Determining the type of emergency event present in the target area and the spatiotemporal characteristics of the emergency event in the sensing data unit based on the event identification results includes: parsing the predicted probability distribution in the event identification results, detecting probability peak points exceeding a preset threshold and their corresponding sensing data unit timestamp information; determining the start and end time points of the emergency event based on the timestamp information corresponding to the probability peak points, combined with the continuous change trend of the predicted probability of the sensing data units before and after, and generating time boundary description information; extracting environmental sensing data of the sensing data unit corresponding to the probability peak points, determining the initial occurrence area identifier and diffusion direction description of the emergency event in the sensing data unit; constructing a development trajectory model of the emergency event in the continuous sensing data unit based on the initial occurrence area identifier, the diffusion direction description, and the time boundary description information; and generating spatiotemporal characteristic information including regional coverage and development trend prediction through the development trajectory model. Based on the emergency event type and the spatiotemporal feature information, an emergency warning instruction containing an event location identifier is generated, and the emergency warning instruction is sent to the target emergency response terminal to initiate response operations.

2. The emergency event identification method based on a neural network model according to claim 1, characterized in that, The feature encoding process performed on the multi-source sensing data set to obtain a multi-dimensional feature representation set of the sensing data unit includes: The multi-source sensing data set is cleaned to remove noise interference from the sensing data units, resulting in a denoised sensing data set. The environmental sensing data in the denoised sensing dataset is subjected to temporal correlation extraction processing to generate the state evolution features of the sensing data unit under continuous time stamps as temporal correlation features. The state evolution features include a description of the change trend between adjacent time stamps. Semantic association extraction processing is performed on the interactive feedback data in the denoised perception dataset to generate semantic association features of the interactive content in the perception data unit. The semantic association features include a contextual description of the interactive intent. The temporal correlation features and the semantic correlation features are input into the feature integration module for correlation analysis processing to obtain an integrated feature set with a unified representation. Based on the association strength evaluation results of different dimensions of features in the integrated feature set, the fusion emphasis of the temporal association features and the semantic association features in the feature encoding process is dynamically adjusted to generate the multidimensional feature representation set.

3. The emergency event identification method based on a neural network model according to claim 1, characterized in that, The process of calling the pre-trained emergency event recognition neural network model to perform event recognition processing on the multi-dimensional feature representation set, generating the event recognition result of the perception data unit, includes: The multidimensional feature representation set is input into the feature processing layer of the emergency event recognition neural network model for hierarchical feature fusion processing to generate a fusion feature vector with multi-granularity representation. The context modeling layer of the emergency event recognition neural network model performs context dependency modeling on the fused feature vector to generate context-enhanced features in the time and semantic dimensions. Based on the probability prediction layer of the emergency event recognition neural network model, the context enhancement features are processed for event probability prediction to generate a probability distribution set containing the predicted probability values ​​corresponding to each emergency event category. The probability distribution set is subjected to category confidence calibration processing to adjust the confidence weight of the predicted probability values ​​of different emergency event categories, thereby generating a calibrated probability distribution set. The event identification result of the sensing data unit is generated based on the calibrated probability distribution set, and the event identification result includes the predicted probability distribution of each emergency event category.

4. The emergency event identification method based on a neural network model according to claim 1, characterized in that, The generation of an emergency warning instruction containing an event location identifier based on the emergency event type and the spatiotemporal feature information includes: The preset early warning rule set corresponding to the emergency event type is parsed, and the early warning level identifier and response strategy code associated with the emergency event type are extracted; Based on the regional coverage range in the spatiotemporal feature information, calculate the center identifier and boundary expansion parameters of the coverage area of ​​the emergency event within the current sensing data unit; The development trend prediction is extracted from the spatiotemporal feature information, and combined with the center identifier of the coverage area, a dynamic development direction prediction function is constructed based on the linear fitting model of the historical development trajectory. The boundary extension parameters are input into a predefined coordinate mapping rule to generate a standardized set of location identifiers consistent with the geographic identification system of the target area. The standardized location identifier set is interpolated in the time dimension according to the dynamic development direction prediction function to generate a predicted identifier sequence of emergency event development paths within a subsequent preset time period. By integrating the warning level identifier, response strategy code, and prediction identifier sequence, an emergency warning instruction containing an event location identifier chain with timestamp alignment is generated. The representation of each identifier node in the event location identifier chain is consistent with the representation of the standardized location identifier set.

5. The emergency event identification method based on a neural network model according to claim 2, characterized in that, The step of performing time-series correlation extraction processing on the environmental sensing data in the denoised sensing data set to generate state evolution features of the sensing data units at continuous time stamps as time-series correlation features includes: Extract the areas of change in environmental sensing data between adjacent sensing data units, and analyze the differences in attribute parameters of the areas of change. Based on the differences in the attribute parameters, the direction and degree of state change of the sensing data unit between adjacent timestamps are described; Based on the description of the direction and degree of state change, a state evolution analysis model for the sensing data unit is constructed. The state evolution analysis model is used to calculate a consistent description of the state change trend of the sensing data unit under continuous timestamps. The state evolution process of the sensing data unit is subjected to periodic pattern recognition processing, and the state repetition features of the sensing data unit within a preset time window are extracted as state evolution features.

6. The emergency event identification method based on a neural network model according to claim 2, characterized in that, The step of performing semantic association extraction processing on the interactive feedback data in the denoised perception data set to generate semantic association features of the interactive content in the perception data unit includes: The text content in the interactive feedback data is segmented into multiple interactive word units. A pre-trained semantic encoding model is invoked to perform context encoding on the multiple interactive word units, generating a semantic vector representation of the interactive feedback data; Analyze the semantic association strength between different interactive word units in the semantic vector representation, and generate an association graph containing semantic association weights; Based on the relationship graph, the semantic features of the core interactive intent in the interactive feedback data are extracted as semantic association features.

7. The emergency event identification method based on a neural network model according to claim 3, characterized in that, The process of performing context dependency modeling on the fused feature vector through the context modeling layer of the emergency event recognition neural network model to generate context-enhanced features in both temporal and semantic dimensions includes: The fused feature vector is input into the time dependency sublayer of the context modeling layer to perform time series dependency analysis and generate time-dimensional context dependency features. The fused feature vector is input into the semantic dependency sublayer of the context modeling layer to perform semantic content contextual dependency analysis and generate semantic dimension contextual dependency features. Cross-dimensional correlation analysis is performed on the context-dependent features of the time dimension and the context-dependent features of the semantic dimension to generate cross-dimensional context-dependent features; Based on the cross-dimensional context association features, the context dependency features of the time dimension and the context dependency features of the semantic dimension are weighted and fused to generate fused context dependency features. The fused context-dependent features are subjected to dimensional alignment processing to generate context-enhanced features in the temporal and semantic dimensions.

8. The emergency event identification method based on a neural network model according to claim 1, characterized in that, The generation of spatiotemporal feature information containing regional coverage and development trend predictions through the development trajectory model includes: Density clustering analysis is performed on the development trajectory identifier sequence output by the development trajectory model to identify the boundaries of dense regions where trajectory points are distributed. The coverage parameters are calculated based on the center identifiers of the dense region boundaries to generate a regional coverage description with the center identifiers and coverage parameters as elements. Extract the set of directional identifiers of the development trajectory identifier sequence under continuous timestamps, calculate the angular change rate and identifier movement gradient of the directional identifiers of adjacent timestamps based on the time interval of each unit, and construct a development trend direction stability evaluation function based on time unit; Based on the directional stability evaluation function, the directional identifier set is subjected to sliding window weighted averaging to generate a directional smoothed trend prediction directional sequence, and the cumulative slope change of the trend prediction directional sequence within a preset time window is calculated. A dynamic region expansion model is constructed based on the coverage parameters and the cumulative slope change of the trend prediction direction sequence to predict the expansion direction and expansion rate of the region coverage in subsequent sensing data units. The expansion direction and expansion rate are described in the input trajectory prediction function to generate a probability distribution map containing the trajectory path within the subsequent time window, and the largest connected region in the probability distribution map with a confidence level exceeding a preset threshold is extracted as the boundary of the predicted development path. Based on the description of the area coverage and the predicted development path boundary, spatiotemporal feature information including a spatial coverage identifier sequence and a trend direction change vector is generated. The spatial coverage identifier sequence is used to mark the real-time coverage area of ​​the emergency event in the sensing data unit, and the trend direction change vector is used to describe the predicted offset angle and velocity change of the emergency event's development direction.

9. A computer device, characterized in that, include: A memory, wherein a computer program is stored; A processor is configured to load the computer program to implement the emergency event identification method based on a neural network model as described in any one of claims 1-8.

Images