Disaster modeling and prediction method for human flow dense area of smart city

Abstract

The application provides a disaster modeling and prediction method for a crowd-intensive area of a smart city, comprising analyzing multi-source heterogeneous data involved in a disaster evolution process of a crowd-intensive area of a city, and designing a multi-dimensional unified representation of disasters of the crowd-intensive area of the city; processing disaster time series data of the crowd-intensive area of the city, mining a correlation relationship of the disaster time series data of the crowd-intensive area of the city through a time series correlation rule algorithm; constructing a correlation event graph based on the correlation relationship; classifying and predicting a multi-dimensional disaster representation of the crowd-intensive area of the city based on a prediction model of the multi-source heterogeneous data, and corresponding the result to the correlation event graph to predict the disaster of the crowd-intensive area of the city. The application uses a time series correlation rule mining algorithm to extract a disaster hazard correlation relationship graph of the crowd-intensive area of the city which integrates human knowledge, and then combines a model prediction of time series data, so that the prediction result is more interpretable and reliable.

Description

Technical Field

[0001] This invention relates to the field of disaster prediction in smart cities, and more specifically, to a method for disaster modeling and prediction in densely populated areas of smart cities. Background Technology

[0002] This invention is applied to disaster prediction in densely populated areas of smart cities. In order to solve the problems of complex and ever-changing urban scenarios and numerous types of disaster events, it is necessary to analyze and model the evolution of disaster events. By constructing a prediction model, the invention can accurately obtain disaster early warning information for a certain location in the city, thereby providing risk point prediction results and informing relevant personnel in advance to facilitate risk investigation.

[0003] To date, there are a few related technical solutions and research results, but the current technical solutions have many shortcomings.

[0004] Patent document CN114419869B discloses a method and system for urban disaster early warning based on time-series multidimensional prediction, relating to the field of disaster early warning. The method includes: acquiring current disaster-related data for a target city; the disaster-related data is multidimensional data composed of multiple factors related to disaster events; one factor corresponds to one single-dimensional data point; each single-dimensional data point corresponds to a time series; inputting the current disaster-related data into an urban disaster early warning model to obtain a future disaster probability index for the target city; the future disaster probability index is used for disaster early warning for the target city; wherein the urban disaster early warning model is obtained by parallel training of a first single-dimensional characterizer, a second single-dimensional characterizer, and a multidimensional characterizer using historical disaster-related data. However, the above scheme relies on data characterization training for early warning, resulting in low reliability.

[0005] Patent document CN115239129A discloses a method and system for constructing a risk point correlation chain in urban rail transit operations. It collects and processes risk point data under specific accident scenarios, determines the correlation relationships between these risk points based on the Apriori algorithm for association rules, and constructs a risk propagation chain based on these relationships. Based on this risk propagation chain, a graph theory model is used to draw the risk point propagation correlation chain under specific accident scenarios, ultimately constructing a proactive risk control method for preventing accidents. However, the above solution does not involve model optimization. Summary of the Invention

[0006] To address the shortcomings of existing technologies, the purpose of this invention is to provide a method for disaster modeling and prediction in densely populated areas of smart cities.

[0007] A disaster modeling and prediction method for densely populated areas in smart cities, provided by the present invention, includes the following steps:

[0008] Step S1: Analyze the multi-source heterogeneous data involved in the disaster evolution process of various densely populated urban areas, and design a multi-dimensional unified representation method for disasters in densely populated urban areas;

[0009] Step S2: Process the disaster time series data of densely populated urban areas, and mine the correlation relationships of disaster time series data of densely populated urban areas through time series association rule algorithm;

[0010] Step S3: Construct an association event graph based on the aforementioned relationships;

[0011] Step S4: Based on the prediction model of the multi-source heterogeneous data, classify and predict the disaster representation of multi-dimensional urban densely populated areas, and map the results to the associated event graph to predict disasters in urban densely populated areas.

[0012] Preferably, step S1 includes:

[0013] Step S101: Obtain relevant disaster data in densely populated urban areas, and for the time period before and after the occurrence of a certain type of disaster, obtain time series data of all relevant events at the incident location;

[0014] Step S102: For the time series data, analyze the heterogeneous multi-source data from the perspectives of disaster type, hazard level, hazard characteristics, and hazard granularity in densely populated urban areas, and divide the corresponding time series data into several groups and assign them label information.

[0015] Preferably, in step S102, for the type of disaster in the current densely populated urban area, the corresponding historical time series data is manually analyzed to hierarchically mark its potential hazard level.

[0016] Preferably, step S2 includes:

[0017] Step S201: Using a self-organizing mapping network, the disaster time series data is dynamically divided into multiple granularities based on data attributes;

[0018] Step S202: Mining frequent itemsets and association rules in disaster time series data using the Apriori algorithm.

[0019] Preferably, step S201 includes:

[0020] Step S2011: Randomly select the weights of M input samples x to initialize the weights of the competition layer w;

[0021] Step S2012: Calculate x for each input node i Competing node w j cos distance d ij,Right now:

[0022]

[0023] Where: ||x i ||2 represents x i The second-order distance, ||w j ||2 represents w j The second moment;

[0024] Step S2013: Based on the cosine distance, select the node index k of the winning competition layer corresponding to each input node:

[0025]

[0026] Step S2014: Calculate the neighborhood radius R of the winning neuron, and the set of neurons W that fall within the neighborhood radius:

[0027]

[0028] W = {w j ||w j -w k ||2 <r m}

[0029] Where r is the initial radius, t is the number of iterations, and t m It is the maximum number of iterations, r m It is the maximum radius of the neighborhood;

[0030] Step S015: For each competing node w in the neuron set W j Based on the current input node x i Update the weights:

[0031] w j :=w j +η(t)(x i -w j )

[0032] Where η(t) is a function that decreases monotonically with t, indicating that the learning rate gradually decreases as the number of iterations increases;

[0033] Step S2016: Proceed to step S2012 and iterate until convergence;

[0034] Step S2017: After the algorithm converges, for each input node x i Both methods can yield corresponding clustering results, dividing time intervals that are adjacent in time and have the same clustering results into the same data subset.

[0035] Preferably, step S202 includes:

[0036] Step S2021: For each data subset D k Calculate each association rule I x →I y Support S in subdata k :

[0037]

[0038] Where σ k (I x ∪I y ) is a subset of data D k Contains I x ∪I y The number of N k It is a subset of data D k The total number of events in;

[0039] Step S2022: Automatically calculate the minimum support threshold S for the association rule. min :

[0040]

[0041] Where K represents the total number of subsets;

[0042] Step S2023: For each data subset D k Calculate each association rule I x →I y Confidence level C in subdata k :

[0043]

[0044] Step S2024: Automatically calculate the minimum confidence threshold C for the association rule. min :

[0045]

[0046] Step S2025: For each association rule I x →I y If it meets the support level S k (I x →I y )>S min (I x →I y And the confidence level satisfies C k (I x →I y C min (I x →I yIf the current rule is a frequent item association rule, then the final result will be a frequent itemset containing all frequent items.

[0047] Preferably, step S3 includes:

[0048] Step S301: Based on the obtained frequent itemsets, establish all events in them as nodes;

[0049] Step S302: Treat all association rules in the frequent itemset as unidirectional edges and connect them to the corresponding nodes;

[0050] Step S303: Use the confidence and support of the association rule as the two weights of the edge. If there are multiple duplicate edges, take the average weight.

[0051] Preferably, step S4 includes:

[0052] Step S401: Represent the multi-source heterogeneous data in step S1, extract the features of the time series data through the models of each modality, and perform supervised training of the model according to the data labels to obtain the degree of harm information of the corresponding hazard features.

[0053] Step S402: Input the trained multi-source heterogeneous data model with different types of hazard data to obtain the hazard level prediction results, and complete the disaster prediction of densely populated urban areas by querying the correlation event graph.

[0054] Preferably, step S401 includes:

[0055] Step S4011: For single-dimensional numerical time series data, input it into a recurrent neural network to encode time series information, and map it to obtain the probability of the hazard level label of the corresponding urban densely populated area;

[0056] For multidimensional matrix-type time series, it is input into a three-dimensional convolutional neural network to understand the temporal information contained in several image frames within a certain time period, and to map it to the probability of the hazard level label of the corresponding urban densely populated area.

[0057] Several models will be trained using time series data for several types of hazards.

[0058] Step S4012: Based on the true labels of the hazard level of disasters in the marked densely populated urban areas, use the minimum cross-entropy loss as the objective function to optimize the models of each modality and obtain the trained disaster prediction model for densely populated smart city areas.

[0059] Preferably, step S402 includes:

[0060] Step S4021: Input the time series data generated by the potentially hazardous events that can be collected in the current scenario into the corresponding model to obtain the predicted degree of hazard of the time series;

[0061] Step S4022: By obtaining the hazard type and hazard degree of several related times, query the associated event graph, and use the depth-first search algorithm for matching. If the number of successful queries on a search branch exceeds a certain threshold, a disaster warning is issued.

[0062] Compared with the prior art, the present invention has the following beneficial effects:

[0063] 1. This invention targets multi-source heterogeneous time-series data, employs multiple models for characterization and encoding, and ultimately optimizes each model by identifying the hazard level through classification labels, thereby facilitating the optimization and training of each model.

[0064] 2. This invention uses a unified temporal association rule mining algorithm to extract disaster hazard association maps of densely populated urban areas incorporating human knowledge, and then combines them with time-series data for model prediction, making the prediction results more interpretable and reliable. Attached Figure Description

[0065] Other features, objects, and advantages of the present invention will become more apparent from the following detailed description of non-limiting embodiments with reference to the accompanying drawings:

[0066] Figure 1 This is a flowchart illustrating the overall process of disaster modeling and prediction methods for densely populated areas in smart cities, as described in this embodiment of the invention.

[0067] Figure 2 This is an architecture diagram of the SOM network described in an embodiment of the present invention.

[0068] Figure 3 This is a schematic diagram illustrating the correlation between subway station hazard events as described in an embodiment of the present invention. Detailed Implementation

[0069] The present invention will now be described in detail with reference to specific embodiments. These embodiments will help those skilled in the art to further understand the present invention, but do not limit the invention in any way. It should be noted that those skilled in the art can make several changes and improvements without departing from the concept of the present invention. These all fall within the protection scope of the present invention.

[0070] like Figures 1 to 3As shown, according to the present invention, a disaster modeling and prediction method for densely populated areas in smart cities first collects multi-source heterogeneous time series data on the disaster evolution process in densely populated areas. Then, through supervised labeling of hazard levels and training of corresponding disaster prediction models, the method predicts the relevant hazards during the disaster evolution process. Further, using prior knowledge and an association rule mining algorithm, the method extracts a correlation graph of disaster events from the time series data. Finally, by predicting potential hazards related to the disaster and traversing and querying the correlation graph, the method completes the prediction of densely populated areas in smart cities.

[0071] To further explain, the disaster modeling and prediction method for densely populated areas in smart cities according to the present invention includes the following steps:

[0072] Step S1: Analyze the multi-source heterogeneous data involved in the disaster evolution process of various densely populated urban areas, and design a multi-dimensional unified representation of disasters in densely populated urban areas to describe the correlation events of disasters in densely populated urban areas from different tensor perspectives;

[0073] Step S2: Process the disaster time series data of densely populated urban areas, and mine the correlation relationship of historical disaster data of densely populated urban areas through time series association rule algorithm;

[0074] Step S3: Construct a disaster prediction graph for densely populated urban areas based on correlations to further explore the implicit knowledge of disasters in densely populated urban areas;

[0075] Step S4: Based on the prediction model using multi-source heterogeneous data, classify and predict the disaster representation of densely populated urban areas in multiple dimensions, and map the results to the associated event graph to accurately predict disasters in densely populated urban areas.

[0076] The temporal association rule algorithm is based on the Apriori algorithm, and the prediction model for multi-source heterogeneous data is based on recurrent neural networks and convolutional neural networks.

[0077] Step S1 includes:

[0078] Step S101: Obtain relevant disaster data in densely populated urban areas, and for the time period before and after the occurrence of a certain type of disaster, obtain the time series information of all relevant events at the incident location;

[0079] Step S102: For the current multi-source heterogeneous historical time series data of disasters in densely populated urban areas, analyze the heterogeneous multi-source data from the perspectives of disaster type, hazard level, hazard characteristics, and hazard granularity in densely populated urban areas, divide the corresponding time series data into several groups and assign label information;

[0080] In step S102, for the type of disaster in densely populated urban areas, the corresponding historical time series data is manually analyzed to hierarchically label the potential hazard level. Specifically, for stampede disasters in public places such as rail transit, the current pedestrian flow sequence information in the area is analyzed and its congestion level is classified to calculate the pedestrian density. For example, by setting up two checkpoints M and N, if the number of people passing through the two checkpoints is N... a and N b If the ratio of people passing through two checkpoints simultaneously is p, then the estimated total number of people is N. a +(1-p)N b By setting thresholds to classify congestion levels, the time-series data of pedestrian flow is labeled. More generally, other hazardous time-series information is labeled with the degree of hazard using artificial methods.

[0081] Step S2 includes:

[0082] Step S201: Through a self-organizing map network, the time series information is dynamically divided into multiple granularities according to the data attributes, so as to minimize the data difference within the same interval and maximize the data difference between adjacent intervals.

[0083] Step S202: Using the Apriori algorithm, we can mine frequent itemsets and association rules from the time series data of urban natural disasters.

[0084] Step S201 includes:

[0085] Step S2011: Randomly select the weights of M input samples x to initialize the weights of the competition layer w;

[0086] Step S2012: Calculate x for each input node i Competing node w j cos distance d ij ,Right now:

[0087]

[0088] Where: ||x i ||2 represents x i The second-order distance, ||w j ||2 represents w j The second moment;

[0089] Step S2013: Based on the cosine distance, select the node index k of the winning competition layer corresponding to each input node:

[0090]

[0091] Step S2014: Calculate the neighborhood radius R of the winning neuron, and the set of neurons W that fall within the neighborhood radius:

[0092]

[0093] W = {w j ||w j -w k ||2 <r m}

[0094] Where r is the initial radius, t is the number of iterations, and t m It is the maximum number of iterations, r m It is the maximum radius of the neighborhood;

[0095] Step S015: For each competing node w in the neuron set W j Based on the current input node x i Update the weights:

[0096] w j :=w j +η(t)(x i -w j )

[0097] Where η(t) is a function that decreases monotonically with t, indicating that the learning rate gradually decreases as the number of iterations increases;

[0098] Step S2016: Proceed to step S2012 and iterate until convergence;

[0099] Step S2017: After the algorithm converges, for each input node x i Both methods can yield corresponding clustering results, dividing time intervals that are adjacent in time and have the same clustering results into the same data subset.

[0100] Step S202 includes:

[0101] Step S2021: For each data subset D k Calculate each association rule I x →I y Support S in subdata k :

[0102]

[0103] Where σ k (I x ∪I y ) is a subset of data D k Contains I x ∪I y The number of N kIt is a subset of data D k The total number of events in the event.

[0104] Step S2022: Automatically calculate the minimum support threshold S for the association rule. min :

[0105]

[0106] Where K represents the total number of subsets.

[0107] Step S2023: For each data subset D k Calculate each association rule I x →I y Confidence level C in subdata k :

[0108]

[0109] Step S2024: Automatically calculate the minimum confidence threshold C for the association rule. min :

[0110]

[0111] Step S2025: For each association rule I x →I y If it meets the support level S k (I x →I y )>S min (I x →I y And the confidence level satisfies C k (I x →I y C min (I x →I y If the rule is set to 0, then the current rule is a frequent item association rule. Ultimately, a frequent itemset containing all frequent items can be obtained.

[0112] According to claim 1, a method for disaster modeling and prediction in densely populated areas of a smart city is provided, wherein step S3 includes:

[0113] Step S301: Based on the frequent itemsets obtained in step S2, establish all events in them as nodes.

[0114] Step S302: Treat all association rules in the frequent itemset as unidirectional edges and connect them to the corresponding nodes.

[0115] Step S303: Use the confidence and support of the association rule as the two weights of the edge. If there are multiple duplicate edges, take the average weight.

[0116] Step S4 includes:

[0117] Step S401: Represent the multi-source heterogeneous data in Step 1, extract the time series features through the models of each modality, and perform supervised training of the models according to the data labels to obtain the degree of harm information of the corresponding hazard features.

[0118] Step S402: Input the trained multi-source heterogeneous data model into different types of hazard data to obtain the hazard degree prediction results, and complete the disaster prediction in densely populated urban areas by querying the correlation event graph;

[0119] Step S401 includes:

[0120] Step S4011: For single-dimensional numerical time series, input them into a recurrent neural network to encode the time series information, and map them to obtain the probability of the hazard level label of the corresponding urban densely populated area. For multi-dimensional matrix time series, such as video information, input them into a three-dimensional convolutional neural network to understand the time series information contained in several image frames within a certain time period, and map them to the probability of the hazard level label of the corresponding urban densely populated area.

[0121] Specifically, several models will be trained using time series data for several types of hazards.

[0122] Step S4012: Based on the true labels of the hazard levels of disasters in densely populated urban areas marked in Step 1, use the minimum cross-entropy loss as the objective function to optimize the models of each modality, and obtain the trained disaster prediction model for densely populated smart city areas.

[0123] Step S402 includes:

[0124] Step S4021: Input the time series data generated by the potentially hazardous events that can be collected in the current scenario into the corresponding model to obtain the predicted degree of hazard of the time series;

[0125] Step S4022: By obtaining the hazard type and hazard degree of several relevant times, query the association graph constructed in step 3, and use the depth-first search algorithm for matching. If the number of successful queries on a search branch exceeds a certain threshold, a disaster warning is issued.

[0126] Example 1

[0127] This embodiment 1 provides a model training method for disaster modeling and prediction in densely populated areas of smart cities, including:

[0128] Step S1: Analyze the multi-source heterogeneous data involved in the disaster evolution process of various densely populated urban areas, and design a multi-dimensional unified representation of disasters in densely populated urban areas to describe the correlation events of disasters in densely populated urban areas from different tensor perspectives;

[0129] Step S2: Based on the prediction model using multi-source heterogeneous data, classify and predict the disaster representation of densely populated urban areas in multiple dimensions, and map the results to the associated event graph to accurately predict disasters in densely populated urban areas.

[0130] Step S1 includes:

[0131] Step S101: Obtain relevant disaster data in densely populated urban areas, and for the time period before and after the occurrence of a certain type of disaster, obtain the time series information of all relevant events at the incident location;

[0132] Specifically, for densely populated urban areas, image and video data such as surveillance footage at nearby latitude and longitude coordinates can be collected; weather data such as temperature, humidity, and wind force; and sensor data generated by public facilities; these can be represented as time-series data at a certain time step T, with the dimension (N,T).

[0133] Step S102: For the current multi-source heterogeneous historical time series data of disasters in densely populated urban areas, analyze the heterogeneous multi-source data from the perspectives of disaster type, hazard level, hazard characteristics, and hazard granularity in densely populated urban areas, divide the corresponding time series data into several groups and assign label information.

[0134] Specifically, step S102 includes:

[0135] Step S1021: For the types of disasters occurring in densely populated urban areas, the corresponding historical time-series data is manually analyzed to hierarchically classify their potential hazard levels. Specifically, for stampede disasters in public places such as rail transit, the current pedestrian flow sequence information within the area is analyzed to classify its congestion level and calculate its pedestrian density. For example, by setting up two checkpoints M and N, if the number of people passing through the two checkpoints is N... a and N b If the ratio of people passing through two checkpoints simultaneously is p, then the estimated total number of people is N. a +(1-p)N bBy setting thresholds to classify congestion levels, the time-series data of pedestrian flow is labeled. More generally, other hazardous time-series information is labeled with the degree of hazard using artificial methods.

[0136] Specifically, step S4 includes:

[0137] Step S401: Represent the multi-source heterogeneous data in Step 1, extract the time series features through the models of each modality, and perform supervised training of the models according to the data labels to obtain the degree of harm information of the corresponding hazard features.

[0138] Step S402: Input the trained multi-source heterogeneous data model with different types of hazard data to obtain the hazard level prediction results, and complete the disaster prediction of densely populated urban areas by querying the correlation event graph.

[0139] Specifically, step S401 includes:

[0140] Step S4011: For single-dimensional numerical time series, input them into a recurrent neural network to encode the time series information, and map them to obtain the probability of the hazard level label of the corresponding urban densely populated area. For multi-dimensional matrix time series, such as video information, input them into a three-dimensional convolutional neural network to understand the time series information contained in several image frames within a certain time period, and map them to the probability of the hazard level label of the corresponding urban densely populated area.

[0141] Specifically, several models will be trained using time series data for different hazard types. If disaster event A... i The multi-source heterogeneous time series dataset is x i ={x1,x2,…,x n},in This represents the time series data for hazard type j corresponding to disaster type i within time period T. i The corresponding hazard level is marked as y i ={y1,y2,…,y n}, then input it into the corresponding hazard treatment model M j In the middle, the output is the predicted hazard level classification probability M. j (x j );

[0142] Step S4012: Based on the true labels of the hazard level of disasters in densely populated urban areas marked in Step 1, use the minimum cross-entropy loss as the objective function to optimize the models of each modality and obtain the trained disaster prediction model for densely populated smart city areas.

[0143] Specifically targeting x j The tag y j , will yj In the classification prediction vector M j (x j The corresponding label prediction value is obtained from the data, and the cross-entropy loss is used to obtain the prediction value for the hazard event A. i Each loss i Each model is optimized using the stochastic gradient descent algorithm for each loss;

[0144] Specifically, step S202 includes:

[0145] Step S4021: Time series data x generated from potentially hazardous events that can be collected in the current scenario. j Input into the corresponding model M j In the process, the predicted class number argmax(M) of the time series is obtained. j (x j This is mapped to the severity level type.

[0146] Step S4022: By obtaining the hazard type and hazard degree of several relevant times, query the association graph constructed in step 3, and use the depth-first search algorithm for matching. If the number of successful queries on a search branch exceeds a certain threshold, a disaster warning is issued.

[0147] Example 2

[0148] This embodiment 1 provides a method for mining disaster association rules and constructing a map of densely populated areas in a smart city, including:

[0149] Step S1: Analyze the multi-source heterogeneous data involved in the disaster evolution process of various densely populated urban areas, and design a multi-dimensional unified representation of disasters in densely populated urban areas to describe the correlation events of disasters in densely populated urban areas from different tensor perspectives;

[0150] Step S2: Process the disaster time series data of densely populated urban areas, and mine the correlation relationship of historical disaster data of densely populated urban areas through time series association rule algorithm;

[0151] Step S1 includes:

[0152] Step S101: Obtain relevant disaster data in densely populated urban areas, and for the time period before and after the occurrence of a certain type of disaster, obtain the time series information of all relevant events at the incident location;

[0153] Step S102: For the current multi-source heterogeneous historical time series data of disasters in densely populated urban areas, analyze the heterogeneous multi-source data from the perspectives of disaster type, hazard level, hazard characteristics, and hazard granularity in densely populated urban areas, divide the corresponding time series data into several groups and assign label information;

[0154] Specifically, in step S102, for the type of disaster in densely populated urban areas, the corresponding historical time series data is manually analyzed to hierarchically label its potential hazard level. Specifically, for stampede disasters in public places such as rail transit, the current pedestrian flow sequence information in the area is analyzed to classify its congestion level and calculate its pedestrian density. For example, by setting up two checkpoints M and N, if the number of people passing through the two checkpoints is N... a and N b If the ratio of people passing through two checkpoints simultaneously is p, then the estimated total number of people is N. a +(1-p)N b By setting thresholds to classify congestion levels, the time-series data of pedestrian flow is labeled. More generally, other hazardous time-series information is labeled with the degree of hazard using artificial methods.

[0155] Step S2 includes:

[0156] Step S201: Through a self-organizing map network, the time series information is dynamically divided into multiple granularities according to the data attributes, so as to minimize the data difference within the same interval and maximize the data difference between adjacent intervals.

[0157] Step S202: Using the Apriori algorithm, we can mine frequent itemsets and association rules from the time series data of urban natural disasters.

[0158] Specifically, step S201 includes:

[0159] Step S2011: Randomly select the weights of M input samples x to initialize the weights of the competition layer w;

[0160] Step S2012: Calculate x for each input node i Competing node w j cos distance d ij ,Right now:

[0161]

[0162] Where: ||x i ||2 represents x i The second-order distance, ||w j ||2 represents w j The second moment;

[0163] Step S2013: Based on the cosine distance, select the node index k of the winning competition layer corresponding to each input node:

[0164]

[0165] Step S2014: Calculate the neighborhood radius R of the winning neuron, and the set of neurons W that fall within the neighborhood radius:

[0166]

[0167] W = {w j ||w j -w k ||2 <r m}

[0168] Where r is the initial radius, t is the number of iterations, and t m It is the maximum number of iterations, r m It is the maximum radius of the neighborhood;

[0169] Step S015: For each competing node w in the neuron set W j Based on the current input node x i Update the weights:

[0170] w j :=w j +η(t)(x i -w j )

[0171] Where η(t) is a function that decreases monotonically with t, indicating that the learning rate gradually decreases as the number of iterations increases;

[0172] Step S2016: Proceed to step S2012 and iterate until convergence;

[0173] Step S2017: After the algorithm converges, for each input node x i Both methods can yield corresponding clustering results, dividing time intervals that are adjacent in time and have the same clustering results into the same data subset.

[0174] Specifically, step S202 includes:

[0175] Step S2021: For each data subset D k Calculate each association rule I x →I y Support S in subdata k :

[0176]

[0177] Where σ k (I x ∪I y ) is a subset of data D k Contains I x ∪I y The number of N k It is a subset of data D k The total number of events in the event.

[0178] Step S2022: Automatically calculate the minimum support threshold S for the association rule. min :

[0179]

[0180] Where K represents the total number of subsets.

[0181] Step S2023: For each data subset D k Calculate each association rule I x →I y Confidence level C in subdata k :

[0182]

[0183] Step S2024: Automatically calculate the minimum confidence threshold C for the association rule. min :

[0184]

[0185] Step S2025: For each association rule I x →I y If it meets the support level S k (I x →I y )>S min (I x →I y And the confidence level satisfies C k (I x →I y C min (I x →I y If the rule is set to 0, then the current rule is a frequent item association rule. Ultimately, a frequent itemset containing all frequent items can be obtained.

[0186] Those skilled in the art will understand that, besides implementing the system and its various devices, modules, and units provided by this invention in the form of purely computer-readable program code, the same functions can be achieved entirely through logical programming of the method steps, making the system and its various devices, modules, and units of this invention function in the form of logic gates, switches, application-specific integrated circuits, programmable logic controllers, and embedded microcontrollers. Therefore, the system and its various devices, modules, and units provided by this invention can be considered as a hardware component, and the devices, modules, and units included therein for implementing various functions can also be considered as structures within the hardware component; alternatively, the devices, modules, and units for implementing various functions can be considered as both software modules implementing the method and structures within the hardware component.

[0187] Specific embodiments of the present invention have been described above. It should be understood that the present invention is not limited to the specific embodiments described above, and those skilled in the art can make various changes or modifications within the scope of the claims, which do not affect the essence of the present invention. Unless otherwise specified, the embodiments and features described in this application can be arbitrarily combined with each other.

Claims

1. A method for disaster modeling and prediction in densely populated areas of smart cities, characterized in that, Includes the following steps: Step S1: Analyze the multi-source heterogeneous data involved in the disaster evolution process of various densely populated urban areas, and design a multi-dimensional unified representation method for disasters in densely populated urban areas; Step S2: Process the disaster time series data of densely populated urban areas, and mine the correlation relationships of disaster time series data of densely populated urban areas through time series association rule algorithm; Step S3: Construct an association event graph based on the aforementioned relationships; Step S4: Based on the prediction model of the multi-source heterogeneous data, classify and predict the disaster representation of multi-dimensional urban densely populated areas, and map the results to the associated event graph to predict disasters in urban densely populated areas. Specifically: Step S4011: For single-dimensional numerical time series data, input it into a recurrent neural network to encode time series information, and map it to obtain the probability of the hazard level label of the corresponding urban densely populated area; For multidimensional matrix-type time series, it is input into a three-dimensional convolutional neural network to understand the temporal information contained in several image frames within a certain time period, and to map it to the probability of the hazard level label of the corresponding urban densely populated area. Several models will be trained using time series data for several types of hazards. Step S4012: Based on the true labels of the hazard level of disasters in the marked densely populated urban areas, use the minimum cross-entropy loss as the objective function to optimize the models of each modality and obtain the trained disaster prediction model for densely populated smart city areas. Step S4021: Input the time series data generated by the potentially hazardous events that can be collected in the current scenario into the corresponding model to obtain the predicted degree of hazard of the time series; Step S4022: By obtaining the hazard type and hazard degree of several related times, query the related event graph, and use the depth-first search algorithm for matching. If the number of successful queries on a search branch exceeds a certain threshold, a disaster warning is issued. Step S1 includes: Step S101: Obtain relevant disaster data in densely populated urban areas, and for the time period before and after the occurrence of a certain type of disaster, obtain time series data of all relevant events at the incident location; Step S102: For the time series data, analyze the heterogeneous multi-source data from the perspectives of disaster type, hazard level, hazard characteristics, and hazard granularity in densely populated urban areas, and divide the corresponding time series data into several groups and assign them label information; Step S2 includes: Step S201: Using a self-organizing mapping network, the disaster time series data is dynamically divided into multiple granularities based on data attributes; Step S202: Mining frequent itemsets and association rules in disaster time series data using the Apriori algorithm.

2. The disaster modeling and prediction method for densely populated areas in smart cities according to claim 1, characterized in that, In step S102, for the type of disaster in the current densely populated urban area, the corresponding historical time series data is manually analyzed to hierarchically mark the potential hazard level.

3. The disaster modeling and prediction method for densely populated areas in smart cities according to claim 1, characterized in that, Step S201 includes: Step S2011: Randomly select Input Samples The weights are used to initialize the competition layer. The weights; Step S2012: Calculate each input node Competing nodes cos distance ,Right now: in: express The second-order moment, express The second moment; Step S2013: Based on the cosine distance, select the node index k of the winning competition layer corresponding to each input node: Step S2014: Calculate the neighborhood radius of the winning neuron. and the set of neurons falling within the neighborhood radius. : in It is the initial radius set. It is the number of iterations. It is the maximum number of iterations. It is the maximum radius of the neighborhood; Step S015: For the neuron set Each competing node Based on the current input node Update the weights: in It is a follow A monotonically decreasing function indicates that the learning rate gradually decreases as the number of iterations increases; Step S2016: Proceed to step S2012 and iterate until convergence; Step S2017: After the algorithm converges, for each input node... Both methods can yield corresponding clustering results, dividing time intervals that are adjacent in time and have the same clustering results into the same data subset.

4. The disaster modeling and prediction method for densely populated areas in smart cities according to claim 1, characterized in that, Step S202 includes: Step S2021: For each data subset Calculate each association rule Support in subdata : in It is a subset of data Includes The number of It is a subset of data The total number of events in; Step S2022: Automatically calculate the minimum support threshold for association rules. : in This indicates the total number of subsets. Step S2023: For each data subset Calculate each association rule Confidence in subdata : Step S2024: Automatically calculate the minimum confidence threshold for association rules. : Step S2025: For each association rule If it meets the support level And the confidence level is satisfied If the current rule is a frequent item association rule, then the final result will be a frequent itemset containing all frequent items.

5. The disaster modeling and prediction method for densely populated areas in smart cities according to claim 1, characterized in that, Step S3 includes: Step S301: Based on the obtained frequent itemsets, establish all events in them as nodes; Step S302: Treat all association rules in the frequent itemset as unidirectional edges and connect them to the corresponding nodes; Step S303: Use the confidence and support of the association rule as the two weights of the edge. If there are multiple duplicate edges, take the average weight.

Images