Disaster risk data mining analysis method and system based on FIPSO and FINN

Abstract

The application provides a disaster risk data mining and analysis method and system based on FIPSO and FINN, which comprises the following steps: collecting and preprocessing multi-source heterogeneous data corresponding to the disaster risk to be analyzed; based on the FIPSO algorithm and the FINN neural network, the preprocessed multi-source heterogeneous data is mined and analyzed to obtain the analysis result of the disaster risk; wherein the FIPSO algorithm is used to initialize the parameters of the FINN neural network, guide the iterative optimization of the FINN neural network, and trigger the Levy flight mechanism when the iterative optimization is stuck, so as to guide the FINN neural network to jump out of the local optimum; the FINN neural network comprises a perception layer, a hidden layer and an output layer connected in sequence; the hidden layer is provided with a parameter disturbance mechanism, which is used to guide the FINN neural network to jump out of the local optimum when the output error of the output layer is greater than a preset threshold; and the application can deeply mine potential risk information from multi-source heterogeneous disaster data, and significantly improve the accuracy of major emergency disaster risk prediction and evaluation.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of disaster emergency management and data mining technology, and more particularly, to a disaster risk data mining analysis method and system based on FIPSO and FINN. BACKGROUND

[0002] In the field of urban disaster emergency management, it is crucial to quickly and accurately obtain disaster information and implement effective response measures to protect the safety of residents' lives and property. With the development of sensor technology and communication technology, real-time collection and analysis of multi-modal data (such as meteorological data, geographic information data, power system operation data, etc.) has become an important means of disaster monitoring. Multi-modal data contains rich disaster feature information, and how to efficiently fuse and process these data to mine their potential value is a key problem faced by current disaster emergency monitoring technology.

[0003] Currently, the prior art (CN118503903A) discloses a flood disaster emergency monitoring method and system. This scheme first collects multi-source heterogeneous data such as video, image, and text, utilizes the complementary advantages between multi-source heterogeneous data, captures inter-modal correlations through self-attention mechanisms, fuses graph neural network automatically learned power grid topology structure knowledge, and realizes intelligent flood monitoring through end-to-end training of deep learning. Although this prior art combines a Transformer encoder with a graph neural network, its model optimization mechanism is single, and it only relies on fixed loss functions (cross-entropy, mean square error, etc.) for supervised training. However, disaster data has the characteristics of high dimensionality, nonlinearity, and dynamic change, and this prior art is difficult to adaptively adjust model parameters, easily falling into local optimum, and unable to accurately capture the potential correlations between multi-modal data in complex disaster scenarios. In addition, this prior art constructs a flood disaster emergency model around the power line topology structure, which is relatively single in function, only focusing on the monitoring of the power system in flood disasters, and cannot be applied to other disaster monitoring. SUMMARY

[0004] The application provides a disaster risk data mining analysis method and system based on FIPSO (Fully Informed Particle Swarm Optimization, improved particle swarm optimization algorithm) and FINN (Fast, Scalable Quantized Neural Network Inference on FPGAs, biological heuristic neural network), which can be widely applied to earthquake, fire and other major emergency disaster risk prediction and evaluation through hierarchical classification data management and intelligent model fusion, and has more advantages in disaster type coverage, data full life cycle management and actual application value.

[0005] To solve the above technical problems, the technical solutions of the application are as follows:

[0006] A disaster risk data mining analysis method based on FIPSO and FINN, comprising the following steps:

[0007] Collecting multi-source heterogeneous data corresponding to the disaster risk to be analyzed and performing preprocessing;

[0008] Based on the FIPSO algorithm and the FINN neural network, the multi-source heterogeneous data after preprocessing is mined and analyzed to obtain the analysis result of the disaster risk;

[0009] The FIPSO algorithm is used to initialize the parameters of the FINN neural network, guide the iterative optimization of the FINN neural network, and trigger the Levy flight mechanism when the iterative optimization is stuck, so as to guide the FINN neural network to jump out of the local optimum;

[0010] The FINN neural network comprises a perception layer, a hidden layer and an output layer connected in sequence; the hidden layer is provided with a parameter disturbance mechanism for guiding the FINN neural network to jump out of the local optimum when the output error of the output layer is greater than a preset threshold.

[0011] Preferably, the multi-source heterogeneous data at least includes any one or more of meteorological data, geological data, social and economic data and historical disaster data.

[0012] Preferably, the preprocessing is performed by distributed computing, and the preprocessing at least includes data cleaning, format conversion and data standardization;

[0013] The data cleaning at least includes removing duplicate data, correcting error data and filling missing data;

[0014] The format conversion includes converting data in different formats into a unified format;

[0015] The data standardization comprises: performing normalization processing on the data according to a preset data standard.

[0016] Preferably, in the FIPSO algorithm, the position vector of the i th particle at the t th iteration is defined as , the parameters corresponding to the FINN neural network, and the velocity vector is defined as , wherein represents the position sub-vector of the i th particle at the t th iteration of the n th element; represents the velocity sub-vector of the i th particle at the t th iteration of the n th element at the corresponding position; n represents the number of parameters of the FINN neural network;

[0017] The position vector is updated according to the following formula :

[0018]

[0019] , wherein is an inertia weight used to balance the global exploration and local exploitation ability; and are respectively a first and a second learning factor used to control the degree of learning of the particle to the individual optimal position and the global optimal position ; and are respectively a first and a second random number;

[0020] The initial parameters of the FINN neural network are obtained according to the position vector obtained after the last update.

[0021] Preferably, in the FIPSO algorithm, an information concentration parameter is introduced, which represents the information concentration from the i th position to the j th position at the t th iteration, and the update formula is:

[0022]

[0023] , wherein is an information volatilization coefficient used to avoid excessive accumulation of information; m is the total number of particles, is the information increment left by the k th particle on the path in the current iteration, and the calculation method is:

[0024] If the particle k passes through the path , , otherwise, ;

[0025] wherein Q is a constant, representing pheromone intensity; is the path length of the kth particle in this iteration;

[0026] The particle search direction is determined according to the current pheromone concentration parameter in the iteration process.

[0027] Preferably, in the FINN neural network, the output of the output layer is is represented as:

[0028]

[0029] wherein, is the input data of the perception layer; is the weight matrix from the perception layer to the hidden layer; is the weight matrix from the hidden layer to the output layer; is the activation function of the hidden layer; is the activation function of the output layer;

[0030] The wall collision neuron is further arranged in the hidden layer, and when the output error of the output layer exceeds the preset error threshold for s consecutive times, the parameter disturbance mechanism of the hidden layer is triggered, and the weight corresponding to the wall collision neuron is adjusted according to the following formula:

[0031]

[0032] wherein, and are the weights corresponding to the wall collision neuron after adjustment and before adjustment respectively; is the disturbance coefficient; is the output error of the output layer; represents the gradient; is the sign function; represents the absolute value.

[0033] Preferably, in the FIPSO algorithm, when the output error of the output layer of the FINN neural network decreases by less than δ for u consecutive iterations, it is determined that the iteration optimization falls into stagnation, and the Levy flight mechanism is triggered; wherein u and δ are preset first and second parameters respectively;

[0034] The Levy flight mechanism adjusts the particle position according to the following formula:

[0035]

[0036] wherein, is a step control parameter, represents element-wise multiplication, is a random step length subject to Levy distribution, and its probability density function is:

[0037]

[0038] in, For a random step size that follows a Lévy distribution The probability density function; These are the characteristic parameters of the Lévy distribution, satisfying... , For gamma function, Let be the random step size of the Lévy distribution.

[0039] Preferably, the output error of the output layer is calculated according to the MSE (mean square error) formula.

[0040] Preferably, the method further includes:

[0041] Multiple databases are constructed using distributed storage technology to manage preprocessed multi-source heterogeneous data in a hierarchical and classified manner; the databases include at least an original database, a resource database, a subject database, and a business database;

[0042] The raw database is used to store unprocessed raw data; the resource database is used to store preprocessed multi-source heterogeneous data; the subject database is used to classify and integrate multi-source heterogeneous data according to different disaster types; the business database is used to customize data storage according to preset business needs; when the raw database acquires new data, it synchronizes the new data to other databases in real time.

[0043] This invention also provides a disaster risk data mining and analysis system based on FIPSO and FINN, which applies the above-mentioned methods and includes:

[0044] Data preprocessing module: used to collect multi-source heterogeneous data corresponding to the disaster risks to be analyzed and perform preprocessing;

[0045] Data mining and analysis module: used to mine and analyze preprocessed multi-source heterogeneous data based on the FIPSO algorithm and FINN neural network to obtain the analysis results of the disaster risk;

[0046] In the data mining and analysis module, the FIPSO algorithm is used to initialize the parameters of the FINN neural network, guide the FINN neural network to iteratively optimize, and trigger the Levy flight mechanism when the iterative optimization stalls, guiding the FINN neural network to escape local optima.

[0047] The FINN neural network includes a perceptual layer, a hidden layer, and an output layer connected in sequence. The hidden layer is equipped with a parameter perturbation mechanism to guide the FINN neural network out of local optima when the output error of the output layer exceeds a preset threshold.

[0048] Compared with the prior art, the beneficial effects of the technical scheme of the present application are:

[0049] The present application provides a kind of FIPSO and FINN-based disaster risk data mining analysis method and system, first, the corresponding multi-source heterogeneous data of disaster risk to be analyzed are collected and preprocessed;After that, based on FIPSO algorithm and FINN neural network, the multi-source heterogeneous data after preprocessing are mined and analyzed, and the analysis result of disaster risk is obtained;Wherein, FIPSO algorithm is used to initialize the parameters of FINN neural network, guide FINN neural network iteration optimization, and when iteration optimization stagnates, trigger the levi flight mechanism, guide FINN neural network to jump out of local optimum;FINN neural network includes perception layer, hidden layer and output layer connected in turn;Hidden layer is provided with parameter disturbance mechanism, for guiding FINN neural network to jump out of local optimum when the output error of output layer is greater than preset threshold;

[0050] The present application innovatively integrates improved particle swarm optimization algorithm (FIPSO) and biological heuristic neural network (FINN), simulates the adaptive response mechanism of flea in complex environment, so that the model can dynamically adjust parameters and break through local optimal solution, thereby deeply mining potential risk information from multi-source heterogeneous disaster data, significantly improving the accuracy of major emergency disaster risk prediction and evaluation, providing more reliable data support and technical support for disaster prevention and emergency decision-making, and can be widely applied to major emergency disaster risk prediction and evaluation such as earthquake and fire, and has more advantages in disaster type coverage, data full life cycle management and actual application value. BRIEF DESCRIPTION OF DRAWINGS

[0051] Figure 1 A FIPSO and FINN-based disaster risk data mining analysis method flowchart provided for embodiment 1.

[0052] Figure 2 A FIPSO and FINN-based disaster risk data mining analysis method flowchart provided for embodiment 2.

[0053] Figure 3 A hybrid intelligent model architecture diagram provided for embodiment 2.

[0054] Figure 4 A FIPSO and FINN collaborative optimization schematic diagram provided for embodiment 2.

[0055] Figure 5 A special connection schematic diagram triggered by wall neuron provided for embodiment 2.

[0056] Figure 6 A levi flight mechanism trigger flowchart provided for embodiment 2.

[0057] Figure 7 FIG. 1 is a structural diagram of a FIPSO and FINN-based disaster risk data mining and analysis system according to an embodiment of the present application. DETAILED DESCRIPTION

[0058] The drawings are only used for illustrative purposes and should not be construed as limiting the patent;

[0059] In order to better illustrate the embodiments, some components in the drawings may be omitted, enlarged or reduced, and do not represent the actual size of the product;

[0060] It is understandable to some skilled persons in the art that some well-known structures and their descriptions in the drawings may be omitted.

[0061] The technical solutions of the present application will be further described below in combination with the drawings and embodiments.

[0062] Embodiment 1

[0063] As shown in Figure 1 , the present embodiment provides a FIPSO and FINN-based disaster risk data mining and analysis method, comprising the following steps:

[0064] S11: Collecting multi-source heterogeneous data corresponding to the disaster risk to be analyzed and performing preprocessing;

[0065] S12: Based on the FIPSO algorithm and the FINN neural network, the preprocessed multi-source heterogeneous data is mined and analyzed to obtain the analysis result of the disaster risk;

[0066] Wherein, the FIPSO algorithm is used to initialize the parameters of the FINN neural network, guide the iterative optimization of the FINN neural network, and when the iterative optimization is stuck, trigger the Levy flight mechanism to guide the FINN neural network to jump out of the local optimum;

[0067] The FINN neural network comprises a perception layer, a hidden layer and an output layer connected in sequence; the hidden layer is provided with a parameter perturbation mechanism for guiding the FINN neural network to jump out of the local optimum when the output error of the output layer is greater than a preset threshold.

[0068] In the specific implementation process, first, multi-source heterogeneous data related to major sudden disaster risks is collected, such as meteorological data, geological data, social and economic data, and historical disaster data, etc.

[0069] Then, the collected multi-source heterogeneous data is preprocessed, such as cleaning, conversion and standardization;

[0070] After data mining analysis, the present embodiment extracts potential disaster risk information from the preprocessed multi-source heterogeneous data and predicts and evaluates major disaster risks by means of a hybrid intelligent model combining the improved particle swarm optimization algorithm (FIPSO) and the biological heuristic neural network (FINN).

[0071] In the present embodiment, the improved particle swarm optimization algorithm (FIPSO) and the biological heuristic neural network (FINN) are combined in the following way: the FIPSO algorithm optimizes the initial weights and topology structure of the FINN neural network, laying a good foundation for model training; in the training process of the FINN neural network, if the prediction accuracy stagnates, the FIPSO algorithm triggers the Levy flight mechanism to guide the network out of the local optimum; at the same time, the FIPSO algorithm simulates the spatiotemporal characteristics of disaster data through the pheromone concentration field to dynamically adjust the search direction; the perception layer of the FINN neural network dynamically adjusts the receptive field of neurons according to the data density (the FINN neural network perception layer draws on the principles of biological perception and adaptation, and can automatically adjust the range of data that can be perceived by neurons according to the data density, so as to realize the dynamic adjustment of the receptive field of neurons), and the hidden layer is provided with "wall collision neurons" that are activated to perform parameter perturbation when the prediction error in a certain area continuously rises; through continuous training and optimization, the model extracts potential disaster risk information from data and predicts and evaluates major disaster risks;

[0072] Finally, the analysis results such as disaster risk information, prediction results and evaluation reports obtained through data mining analysis are output to the user in the form of visual charts and text reports;

[0073] Traditional disaster risk analysis models often fall into local optimal solutions and are difficult to adapt to complex and variable disaster data characteristics; in the present method, the improved particle swarm optimization algorithm (FIPSO) and the biological heuristic neural network (FINN) are combined innovatively in data mining analysis: the FIPSO algorithm simulates the spatiotemporal characteristics of disaster data through the pheromone concentration field to dynamically adjust the search direction, optimizes the initial weights and topology structure of the FINN neural network, and triggers the Levy flight mechanism to guide the network out of the local optimum when the network training accuracy stagnates; the perception layer of the FINN neural network can dynamically adjust the receptive field of neurons according to the data density, and the "wall collision neurons" in the hidden layer are activated to perform parameter perturbation when the prediction error continuously rises;

[0074] The present method enables the intelligent model to have stronger global search and self-adaptive adjustment capabilities through the synergistic work of the FINN neural network and the FIPSO algorithm, can deeply mine potential risk information from multi-source heterogeneous disaster data, significantly improves the accuracy of major disaster risk prediction and evaluation, and thus obtains much stronger analysis capabilities than traditional single models.

[0075] Embodiment 2

[0076] As Figure 2 shown, the embodiment provides a disaster risk data mining and analysis method based on FIPSO and FINN, comprising the following steps:

[0077] S21: Collecting multi-source heterogeneous data corresponding to the disaster risk to be analyzed and performing preprocessing;

[0078] S22: Using a distributed storage technology to construct multiple databases for hierarchical classification management of the preprocessed multi-source heterogeneous data; the databases at least include an original database, a resource database, a theme database and a business database;

[0079] The original database is used to store raw data without preprocessing; the resource database is used to store preprocessed multi-source heterogeneous data; the theme database is used to classify and integrate multi-source heterogeneous data according to different disaster types; the business database is used to customize data storage according to preset business requirements; when the original database acquires new data, the new data is synchronized to other databases in real time;

[0080] S23: Based on the FIPSO algorithm and the FINN neural network, the preprocessed multi-source heterogeneous data is mined and analyzed to obtain the analysis result of the disaster risk;

[0081] The FIPSO algorithm is used to initialize the parameters of the FINN neural network, guide the iterative optimization of the FINN neural network, and trigger the Levy flight mechanism when the iterative optimization stagnates, guiding the FINN neural network to jump out of the local optimum;

[0082] The FINN neural network includes a perception layer, a hidden layer and an output layer connected in sequence; the hidden layer is provided with a parameter perturbation mechanism for guiding the FINN neural network to jump out of the local optimum when the output error of the output layer is greater than a preset threshold;

[0083] In the embodiment, the multi-source heterogeneous data at least includes any one or more of meteorological data, geological data, social and economic data and historical disaster data;

[0084] In the embodiment, the preprocessing is performed using distributed computing, and the preprocessing at least includes data cleaning, format conversion and data standardization;

[0085] The data cleaning at least includes removing duplicate data, correcting erroneous data and filling missing data;

[0086] The format conversion includes converting data in different formats into a unified format;

[0087] The data standardization comprises: performing normalization processing on the data according to a preset data standard;

[0088] In the FIPSO algorithm, the position vector of the i th particle at the t th iteration is defined as The velocity vector is defined as wherein, represents the position sub-vector of the i th particle at the t th iteration of the n th element; represents the velocity sub-vector of the i th particle at the t th iteration of the n th element at the corresponding position; n represents the number of parameters of the FINN neural network;

[0089] The position vector is updated according to the following formula :

[0090]

[0091] wherein, is an inertia weight, used to balance the global exploration and local development capabilities; and are respectively the first and second learning factors, used to control the degree of learning of the particle to the individual optimal position and the global optimal position ; and are the first and second random numbers;

[0092] The initial parameters of the FINN neural network are obtained according to the position vector obtained after the last update;

[0093] In the FIPSO algorithm, the pheromone concentration parameter is introduced, which represents the pheromone concentration from the i th position to the j th position at the t th iteration, and the update formula is:

[0094]

[0095] wherein, is a pheromone evaporation coefficient, used to avoid excessive accumulation of pheromone; m is the total number of particles, is the pheromone increment left by the k th particle on the path in the current iteration, The calculation method is:

[0096] If the particle k passes through the path , otherwise, ;

[0097] wherein, Q is a constant, representing the pheromone intensity; path length of the kth particle in the current iteration;

[0098] determining a search direction of the particle according to the current pheromone concentration parameter in the iteration process;

[0099] In the FINN neural network in the embodiment, the output of the output layer is is expressed as:

[0100]

[0101] wherein, is the input data of the perception layer; is the weight matrix of the perception layer to the hidden layer; is the weight matrix of the hidden layer to the output layer; is the activation function of the hidden layer; is the activation function of the output layer;

[0102] The wall collision neuron is also arranged in the hidden layer, when the output error of the output layer exceeds the preset error threshold for s consecutive times, the parameter disturbance mechanism of the hidden layer is triggered, and the weight corresponding to the wall collision neuron is adjusted according to the following formula:

[0103]

[0104] wherein, and are the weights corresponding to the wall collision neuron after adjustment and before adjustment respectively; is the disturbance coefficient; is the output error of the output layer; represents the gradient; is the sign function; represents the absolute value;

[0105] In the FIPSO algorithm in the embodiment, when the output error of the output layer of the FINN neural network decreases by less than δ for u consecutive iterations, it is determined that the iteration optimization falls into stagnation, and the Levy flight mechanism is triggered; wherein, u and δ are respectively the first and second preset parameters;

[0106] The Levy flight mechanism adjusts the particle position according to the following formula:

[0107]

[0108] wherein, is a step control parameter, represents element-wise multiplication, is a random step length subject to Levy distribution, and the probability density function thereof is:

[0109]

[0110] wherein, is the probability density function of the random step obeying the Levy distribution; is a characteristic parameter of the Levy distribution, satisfying , is the Gamma function, is the value of the random step of the Levy distribution;

[0111] In this embodiment, the output error of the output layer is calculated according to the MSE (Mean Square Error) formula.

[0112] In the specific implementation process, first, through various ways such as API interface, web crawler, sensor, etc., real-time or periodic collection of multi-source heterogeneous data related to major sudden disaster risks is performed, such as meteorological data, geological data, social and economic data, and historical disaster data, etc., and the above multi-source heterogeneous data covers information such as natural environment, geographical conditions, social and economic influence factors, and historical laws of disaster occurrence, thereby providing an original data basis for subsequent analysis;

[0113] Then, the collected multi-source heterogeneous data is preprocessed, such as cleaning, conversion, and standardization; wherein, the data cleaning process includes removing duplicate data, correcting error data according to data logic rules, and filling missing data through interpolation method / statistical estimation, etc.; conversion is to convert data in different formats into a unified data format using ETL tools; standardization processing is to uniformly standardize and normalize the coding, naming, unit, etc. of the data according to the preset data standard; in this step, the distributed computing technology is adopted to improve the data processing efficiency;

[0114] Then, data storage is performed, and the distributed storage technology is adopted to construct databases at various levels, and through the construction of a hierarchical and classified data management system of original library, resource library, theme library, and business library, scientific storage and efficient calling of data are realized;

[0115] ​The embodiment is based on different service and scene requirements, and classifies the received data; the original library is used for storing the original collected data without processing, and the original state of the data is retained; the resource library receives the basic data after standardization processing, and provides standardized data support for subsequent analysis; the theme library integrates the data according to different disaster themes such as earthquakes and floods, to facilitate targeted analysis; the service library stores data according to specific service requirements such as emergency dispatch and risk assessment; secondly, the embodiment establishes a data index and metadata management system; the data index is used for quickly positioning and retrieving data, and the metadata management system is used for recording information such as the source of the data, the collection time, the processing process and the storage location; at the same time, data update rules are set, when the original library obtains new data, the data processing process is automatically triggered, and the updated data is synchronized to the resource library, the theme library and the service library, to ensure the timeliness and consistency of the data;

[0116] In the embodiment, the original library is used for directly receiving and storing the original and unprocessed heterogeneous data obtained from various sensing devices, government databases, the Internet and other channels; and in this level, the static data is mainly encrypted by using the data encryption standard, different data access levels are allocated according to user roles based on the RBAC (role-based access control) model, to realize fine-grained permission control; through the operation log recording function, combined with the blockchain technology, the data usage is tracked throughout and cannot be tampered with;

[0117] In the embodiment, the resource library is the key link of data quality guarantee after the data in the original library is subjected to preliminary processing such as cleaning, standardization, deduplication and format unification, to form standardized data resources; through the ETL (extraction, transformation and loading) data pipeline, the data is extracted from the original library, a series of quality rules (such as processing missing values, unifying the coordinate system and standardizing the measurement unit) are applied for transformation, and finally loaded into the standard data table of the resource library; a distributed columnar database or a large-scale parallel processing data warehouse is adopted, to support efficient data cleaning and transformation tasks;

[0118] In the embodiment, the theme library is the key link of data quality guarantee after the data in the original library is subjected to preliminary processing such as cleaning, standardization, deduplication and format unification, to form standardized data resources; through the ETL (extraction, transformation and loading) data pipeline, the data is extracted from the original library, a series of quality rules (such as processing missing values, unifying the coordinate system and standardizing the measurement unit) are applied for transformation, and finally loaded into the standard data table of the resource library; a distributed columnar database or a large-scale parallel processing data warehouse is adopted, to support efficient data cleaning and transformation tasks;

[0119] In this embodiment, the service library provides customized, model-ready data sets for specific predictive analysis models, and the data is usually processed into the form of feature vectors and labels; relevant topic data is extracted from the topic library for further feature engineering, normalization, sampling, etc., and finally a directly usable data set is generated; since this library data contains key features for model training, field-level encryption and data desensitization strategies need to be implemented to prevent sensitive information (such as the precise location of critical infrastructure) from being leaked;

[0120] The four-level database system described above ensures efficient, secure, and controllable data flow between the libraries through strict data flow processing logic and integrated interface definitions, providing a solid data foundation for disaster risk mining analysis and prediction;

[0121] After data mining analysis, the hybrid intelligent model of the improved particle swarm optimization algorithm (FIPSO) and the biological heuristic neural network (FINN) is used to extract potential disaster risk information from the preprocessed multi-source heterogeneous data and predict and evaluate major sudden disaster risks;

[0122] As shown in Figure 3 In this embodiment, the improved particle swarm optimization algorithm (FIPSO) and the biological heuristic neural network (FINN) are combined as follows: the FIPSO algorithm optimizes the initial weights and topology structure of the FINN neural network, laying a good foundation for model training; during the training process of the FINN neural network, if the prediction accuracy stagnates, the FIPSO algorithm triggers the Levy flight mechanism to guide the network out of the local optimum; at the same time, the FIPSO algorithm simulates the spatiotemporal characteristics of disaster data through the pheromone concentration field to dynamically adjust the search direction; the perception layer of the FINN neural network dynamically adjusts the neuron receptive field according to the data density, and the hidden layer is provided with "wall-hitting neurons" that are activated to perform parameter perturbation when the prediction error in a certain area continuously rises; through continuous training and optimization, the model continuously extracts potential disaster risk information from the data and predicts and evaluates major sudden disaster risks;

[0123] Specifically, as shown in Figure 4 First, the FIPSO algorithm framework is constructed, the particle position update formula and the pheromone concentration field update mechanism are designed, and the structure characteristics of the FINN neural network are combined to realize collaborative optimization of the two, thereby extracting potential risk information from disaster data and performing accurate prediction and evaluation;

[0124] In the FIPSO algorithm, the position vector of the i-th particle at the t-th iteration is defined as corresponding to the parameters of the FINN neural network, and the velocity vector is defined as wherein, represents the position sub-vector of the nth element of the ith particle at the tth iteration, corresponding to the weight and topology structure parameters of the FINN neural network, and the physical meaning is the specific value of the ith particle in the solution space corresponding to the weight or topology structure parameter of the FINN neural network at the tth iteration; represents the velocity sub-vector of the nth element of the ith particle at the tth iteration in the corresponding position, corresponding to the change rate of the weight and topology structure parameters of the FINN neural network, and the physical meaning is the change speed and direction of the ith particle in the solution space corresponding to the FINN weight or topology structure parameter at the tth iteration; n represents the number of parameters of the FINN neural network;

[0125] In the embodiment, each sub-vector in the position vector corresponds to the weight of the perception layer to the hidden layer, the hidden layer to the output layer, and the topology structure parameters such as the number of neurons and the connection mode in the FINN neural network, so as to realize the specific correspondence with the weight and topology structure parameters of the FINN neural network, and the position updating formula is:

[0126]

[0127] wherein, is the inertia weight, used to balance the global exploration and local development capabilities; , is the learning factor, controlling the degree of learning of the particle to the individual optimal position and the global optimal position G; , is a random number in the interval [0, 1]; the formula adjusts the particle velocity to guide it to search for better parameters in the solution space, and provides a better initial state for the FINN neural network;

[0128] is used to simulate the spatial and temporal characteristics of disaster data and construct the pheromone concentration field , representing the pheromone concentration from the ith position to the jth position in the solution space at the tth iteration, and the updating formula is:

[0129]

[0130] wherein, is the pheromone evaporation coefficient, which avoids the premature algorithm caused by excessive accumulation of pheromone; m is the total number of particles, is the pheromone increment left by the kth particle on the path (i, j) in this iteration, and the calculation method is:

[0131] If the particle k passes through the path (i, j), then , otherwise, =0 ;

[0132] Wherein, Q is a constant, the value range is between 0.1 to 10, representing the pheromone intensity, L k The path length of the kth particle in this iteration; the update of the pheromone concentration field enables the particle to dynamically adjust the direction according to historical experience (pheromone concentration) during the search process, and preferentially explore the data characteristics corresponding to the high-risk area;

[0133] In this embodiment, the construction of the pheromone concentration field and the update according to the above formula simulate the spatiotemporal characteristics of disaster data, allowing the particle to dynamically adjust the direction according to historical experience during the search process to preferentially explore the data characteristics of the high-risk area. The particle search correlates the FINN network parameter iterative update by optimizing the position vector corresponding to the FINN neural network weight and topological structure parameter;

[0134] When the FINN neural network is trained, let the weight matrix of the perception layer input to the hidden layer be W ih , the weight matrix of the hidden layer to the output layer be W ho , the input data be X, the hidden layer activation function be , and the output layer activation function be Then the calculation of the network output Y is:

[0135]

[0136] This embodiment also sets a "wall collision neuron" in the hidden layer. When the prediction error E of a certain area exceeds the threshold value (generally 0.01-1.0) for s times (generally set to 3-10 times) in a row, the special connection of the activated wall collision neuron is activated. The prediction error is calculated using the mean square error:

[0137]

[0138] Wherein, N is the number of samples, Y n is the network prediction output, is the true value;

[0139] At this time, parameter perturbation is performed by adjusting the weight of the special connection, and the calculation formula is:

[0140]

[0141] Wherein, and are the weights of the wall collision neuron after adjustment and before adjustment, respectively; is the perturbation coefficient, with a value range of 0.001-0.1; is the output error of the output layer; represents the gradient; is the sign function; denotes absolute value; this formula makes the neural network jump out of the local optimal solution by introducing disturbance related to error gradient, adjusting the weight of special connection corresponding to the bump neuron in the hidden layer instead of all parameters, as shown in Figure 5 FIG. 3 is a schematic diagram of special connection corresponding to the bump neuron;

[0142] In addition, in the process of cooperation between the FIPSO algorithm and the FINN neural network, when the prediction accuracy of the FINN neural network is continuously improved for u times (generally set to 5-15 times) in the iteration and the improvement is less than (generally 0.001-0.05), the FIPSO algorithm triggers the Levy flight mechanism to adjust the particle position:

[0143]

[0144] wherein is a step control parameter, which needs to be determined according to the characteristics of disaster data and the balance between global exploration and local development of the model, and the specific range is between 0.01 and 0.5, denotes element-wise multiplication, is a random step length obeying Levy distribution, and the probability density function is:

[0145]

[0146] wherein, is the probability density function of the random step length obeying Levy distribution; is a characteristic parameter of Levy distribution and 0 ≤ 2, is a gamma function, x is a random variable, i.e., the value of the random step length of Levy distribution, is a sine function term, is a denominator attenuation term; the Levy flight mechanism generates long-distance jumps to make the algorithm search for better solutions in the global range and avoid the FINN neural network from falling into a local optimum, and the process is as shown in Figure 6

[0147] ​The idea of the method is to fuse the improved particle swarm optimization algorithm (FIPSO) and the biological heuristic neural network (FINN) to realize the deep mining and accurate analysis of major disaster risks; the FIPSO algorithm simulates the movement of particles in the solution space, adjusts the particle position by using inertia weight, learning factor and other parameters, simulates the spatio-temporal characteristics of disaster data by combining the pheromone concentration field, dynamically adjusts the search direction, optimizes the initial weight and topology structure of the FINN neural network, and triggers the Levy flight mechanism to guide it out of the local optimum when the network training is stuck; the FINN neural network dynamically adjusts the neuron receptive field based on data density, and when the prediction error of a certain area continuously rises, the "wall collision neuron" of the hidden layer activates special connections for parameter disturbance; the two work together, FIPSO provides better initial parameters for FINN and assists global search, and FINN adjusts the structure adaptively during training, extracts potential risk features from hierarchical and classified disaster data through continuous iteration optimization, and finally realizes accurate prediction and evaluation of major disaster risks;

[0148] The operation process of the fusion intelligent model proposed in this embodiment has certain similarity with the flea reaction, FIPSO algorithm and FINN neural network work together, and continuously adjust and optimize to find the optimal solution, so as to realize accurate prediction and evaluation of major disaster risks; in this process, the FIPSO algorithm uses the pheromone concentration field to simulate the spatio-temporal characteristics of disaster data to dynamically adjust the particle search direction, the FINN neural network dynamically adjusts the neuron receptive field according to the data density, and the "wall collision neuron" of the hidden layer activates special connections for parameter disturbance when the prediction error continuously rises, which reflects the characteristics of self-adaptive adjustment of the model according to its own state and external environment changes;

[0149] This embodiment also periodically evaluates the prediction performance of the fusion model, automatically adjusts the inertia weight of the FIPSO algorithm and the learning rate of the FINN neural network when the model performance decreases, and incrementally trains the model according to new disaster data to continuously optimize the model parameters and improve the accuracy of major disaster risk prediction and evaluation;

[0150] Finally, the analysis results such as disaster risk information, prediction results and evaluation reports obtained by data mining analysis are output to the user in the form of visual charts and text reports; in this embodiment, data can be output through various types of visual charts such as line charts, column charts, heat maps and maps, so as to intuitively display the spatio-temporal distribution and development trend of disaster risks; at the same time, the user can customize the report format and content to meet the individual needs of different users; in addition, warning conditions are set, when the analysis results reach the warning conditions, disaster risk warning information is automatically generated and pushed to relevant personnel through SMS, email, pop-up window and other ways;

[0151] The prior art often focuses on single-link protection in data security, lacks whole-process protection, and the scheme also sets a data security management mechanism throughout the whole process of data collection, storage, processing and output; in the data transmission and storage process, a combination of symmetric encryption and asymmetric encryption is used to highly encrypt and protect the data; based on the RBAC model, different data access levels are allocated according to user roles to realize fine-grained permission control; through the operation log recording function, combined with the blockchain technology, the data usage is tracked throughout the whole process and cannot be tampered with; compared with the traditional technology, this all-round and multi-level data security protection system can more effectively guarantee the security of data in the whole processing and application process, and prevent data leakage and illegal access;

[0152] Traditional disaster risk analysis models often fall into local optimal solutions and are difficult to adapt to complex and variable disaster data characteristics; in the data mining analysis, the improved particle swarm optimization algorithm (FIPSO) and the biological heuristic neural network (FINN) are innovatively combined, the FIPSO algorithm simulates the spatiotemporal characteristics of disaster data by constructing the pheromone concentration field, dynamically adjusts the search direction, optimizes the initial weight and topology structure of the FINN neural network, and when the network training accuracy stagnates, triggers the Levy flight mechanism to guide it to jump out of the local optimum; the perception layer of the FINN neural network can dynamically adjust the receptive field of neurons according to the data density, and the "wall collision neuron" of the hidden layer activates special connections to perform parameter perturbation when the prediction error continuously rises;

[0153] The method makes the intelligent model have stronger global search and adaptive adjustment ability through the cooperative work of the FINN neural network and the FIPSO algorithm, can deeply mine potential risk information from multi-source heterogeneous disaster data, significantly improves the accuracy of major emergency disaster risk prediction and evaluation, and thus obtains stronger analysis ability far beyond traditional single models.

[0154] Embodiment 3

[0155] As shown in Figure 7 The embodiment provides a disaster risk data mining and analysis system based on FIPSO and FINN, which applies the method described in embodiment 2 and comprises:

[0156] The data preprocessing module 301 is used for collecting and preprocessing multi-source heterogeneous data corresponding to the disaster risk to be analyzed;

[0157] The data storage management module 302 is used for constructing multiple databases by using distributed storage technology, and classifying and managing the preprocessed multi-source heterogeneous data; the databases at least include an original database, a resource database, a theme database and a business database;

[0158] The original library is used to store raw data without preprocessing; the resource library is used to store preprocessed multi-source heterogeneous data; the theme library is used to classify and integrate multi-source heterogeneous data according to different disaster types; the business library is used to store data according to preset business requirements; when the original library acquires new data, the new data is synchronized to other databases in real time;

[0159] The data mining and analysis module 303 is used to mine and analyze the preprocessed multi-source heterogeneous data based on the FIPSO algorithm and the FINN neural network, and obtain the analysis result of the disaster risk;

[0160] The FIPSO algorithm is used to initialize the parameters of the FINN neural network, guide the iterative optimization of the FINN neural network, and trigger the Levy flight mechanism when the iterative optimization stagnates, guiding the FINN neural network to jump out of the local optimum;

[0161] The FINN neural network includes a perception layer, a hidden layer and an output layer connected in sequence; the hidden layer is provided with a parameter perturbation mechanism for guiding the FINN neural network to jump out of the local optimum when the output error of the output layer is greater than a preset threshold;

[0162] The data output module 304 is used to output the analysis result of the disaster risk to the user.

[0163] In the specific implementation process, first, the data preprocessing module 301 collects and preprocesses multi-source heterogeneous data corresponding to the disaster risk to be analyzed;

[0164] Then, the data storage and management module 302 uses distributed storage technology to build multiple databases to manage the preprocessed multi-source heterogeneous data in a hierarchical manner; the databases at least include an original library, a resource library, a theme library and a business library;

[0165] After that, the data mining and analysis module 303 mines and analyzes the preprocessed multi-source heterogeneous data based on the FIPSO algorithm and the FINN neural network, and obtains the analysis result of the disaster risk; wherein the FIPSO algorithm is used to initialize the parameters of the FINN neural network, guide the iterative optimization of the FINN neural network, and trigger the Levy flight mechanism when the iterative optimization stagnates, guiding the FINN neural network to jump out of the local optimum; the FINN neural network includes a perception layer, a hidden layer and an output layer connected in sequence; the hidden layer is provided with a parameter perturbation mechanism for guiding the FINN neural network to jump out of the local optimum when the output error of the output layer is greater than a preset threshold;

[0166] Finally, the data output module 304 outputs the analysis result of the disaster risk to the user.

[0167] To sum up, the core improvement points of the present application are:

[0168] 1) The improved particle swarm optimization algorithm (FIPSO) is deeply integrated with the biological heuristic neural network (FINN), the FIPSO algorithm optimizes the initial weight and topology structure of the FINN neural network, lays the foundation for model training, and realizes the complementary advantages through dynamic interaction:

[0169] The FIPSO algorithm simulates the search characteristics of group intelligence, simulates the spatial and temporal distribution characteristics of disaster data by constructing pheromone concentration field, dynamically adjusts the search direction of particles, optimizes the initial weight and topology structure of the FINN neural network, provides a better starting point for model training, and when the network training accuracy stagnates, triggers the "Levy flight" mechanism similar to the flea breaking through the limit, guiding the network to jump out of the local optimum;

[0170] The FINN neural network learns from the biological perception and adaptation principle, and its perception layer can automatically adjust the receptive field of neurons according to the data density, just like biological adjustment of perception range according to environmental changes; The "wall collision neuron" is set in the hidden layer, when the prediction error of a certain area continues to rise, the special connection path is activated to disturb the network parameters, simulating the behavior adjustment of organisms after hitting the wall, enhancing the self-adaptive ability of the network to complex disaster data. The two work together to realize the deep mining and accurate prediction and evaluation of major emergency disaster risk information, and break through the limitations of traditional models in processing high-dimensional and dynamic disaster data.

[0171] 2) By integrating the improved particle swarm optimization algorithm (FIPSO) and the biological heuristic neural network (FINN), the deep mining and accurate analysis of major emergency disaster risk are realized;

[0172] Through API interface, web crawler, sensor and other diversified channels, real-time or periodic collection of meteorological, geological, social and economic and other multi-source heterogeneous data is realized, and a comprehensive data foundation is constructed; Distributed computing technology is adopted, combined with data cleaning (de-duplication, error correction, completion), format conversion and standardization processing procedures to ensure data quality and consistency and improve data processing efficiency; Based on business needs, four-level database system of original library, resource library, theme library and business library is constructed, distributed storage technology and metadata management system are adopted to realize hierarchical and classified storage, fast retrieval and dynamic update of data; Finally, FIPSO provides better initial parameters for FINN and assists global search, FINN adjusts the structure adaptively in training, extracts potential risk features from hierarchical and classified stored disaster data through continuous iteration optimization, and finally realizes accurate prediction and evaluation of major emergency disaster risk.

[0173] The same or similar reference signs correspond to the same or similar parts;

[0174] The terms describing the positional relationship in the drawings are used only for illustrative purposes and should not be construed as limiting the present patent;

[0175] Obviously, the above-mentioned embodiments of the present application are only examples for clearly illustrating the present application, and are not intended to limit the implementation manners of the present application. Based on the above description, other different forms of changes or variations can be made by those skilled in the art. Here, it is not necessary and also impossible to enumerate all the implementation manners. Any modification, equivalent replacement and improvement, etc. made within the spirit and principle of the present application should be included in the protection scope of the claims of the present application.

Claims

1. A FIPSO and FINN based disaster risk data mining analysis method, characterized in that, The method comprises the following steps: Collecting multi-source heterogeneous data corresponding to disaster risk to be analyzed and preprocessing; The multi-source heterogeneous data comprises meteorological data, geological data, social and economic data, and historical disaster data; Based on the FIPSO algorithm and the FINN neural network, the preprocessed multi-source heterogeneous data is mined and analyzed to obtain the analysis result of the disaster risk; The FIPSO algorithm is used to initialize the parameters of the FINN neural network, guide the iterative optimization of the FINN neural network, and trigger the Levy flight mechanism when the iterative optimization stagnates, guiding the FINN neural network to jump out of the local optimum; The FINN neural network comprises a perception layer, a hidden layer, and an output layer connected in sequence; the hidden layer is provided with a parameter perturbation mechanism for guiding the FINN neural network to jump out of the local optimum when the output error of the output layer is greater than a preset threshold; In the FIPSO algorithm, when the output error of the output layer of the FINN neural network is reduced by less than δ for u consecutive iterations, it is determined that the iterative optimization has stagnated, and the Levy flight mechanism is triggered; wherein u and δ are respectively preset first and second parameters; The Levy flight mechanism adjusts the particle position according to the following formula: wherein is a step size control parameter, denotes element-wise multiplication, is a random step size following a Levy distribution with probability density function wherein is a random step length obeying a Levy distribution is a probability density function; is a characteristic parameter of the Levy distribution, satisfying , is a gamma function, is a random step length taking values from a Levy distribution.

2. The FIPSO and FINN based disaster risk data mining analysis method according to claim 1, wherein, The preprocessing is performed using distributed computing, and the preprocessing at least comprises data cleaning, format conversion, and data standardization; The data cleaning at least comprises removing duplicate data, correcting erroneous data, and filling missing data; The format conversion comprises converting data in different formats into a unified format; The data standardization comprises normalizing the data according to a preset data standard. 3.The FIPSO and FINN based disaster risk data mining and analysis method according to claim 1, wherein, In the FIPSO algorithm, the position vector of the ith particle at the tth iteration is defined as corresponding to the parameters of the FINN neural network, while the velocity vector is defined as wherein, denotes the position sub-vector of the nth element of the ith particle at the tth iteration; denotes the velocity sub-vector of the nth element of the ith particle at the tth iteration at the corresponding position; n denotes the number of parameters of the FINN neural network; The position vector is updated according to the following equation : wherein, is an inertia weight used to balance the global exploration and local exploitation capabilities; and are first and second learning factors, respectively, used to control the degree of learning of the particle towards the individual optimal position and the global optimal position , respectively; and are first and second random numbers; The initial parameters of the FINN neural network are obtained according to the position vector obtained after the last update.

4. The FIPSO and FINN-based disaster risk data mining and analysis method according to claim 3, characterized in that, In the FIPSO algorithm, a pheromone concentration parameter is introduced , which represents the pheromone concentration from the i-th position to the j-th position at the t-th iteration, and its update formula is: wherein, is the pheromone evaporation coefficient, used to avoid excessive accumulation of pheromone; m is the total number of particles, is the increment of pheromone left by the kth particle on the path in this iteration, is calculated as: If particle k traverses path then , else ; wherein Q is a constant, representing the intensity of pheromone; is the path length of the kth particle in this iteration; The particle search direction is determined according to the current pheromone concentration parameter in the iteration process.

5. The FIPSO and FINN based disaster risk data mining analysis method according to claim 1, wherein, In the FINN neural network, the output of the output layer is represented as: wherein, is input data to the perception layer; is a weight matrix from the perception layer to the hidden layer; is a weight matrix from the hidden layer to the output layer; is an activation function of the hidden layer; is an activation function of the output layer; The hidden layer is also provided with a wall collision neuron, and when the output error of the output layer exceeds the preset error threshold for s consecutive times, the parameter perturbation mechanism of the hidden layer is triggered, and the weight of the wall collision neuron is adjusted according to the following formula: wherein, and are the weights of the corresponding wall-kicking neurons after and before adjustment, respectively; is a disturbance coefficient; is the output error of the output layer; denotes the gradient; is a sign function; denotes the absolute value.

6. The FIPSO and FINN based disaster risk data mining analysis method according to claim 1, wherein, The output error of the output layer is calculated according to the MSE mean square error formula.

7. The FIPSO and FINN-based disaster risk data mining and analysis method according to any one of claims 1 to 6, characterized in that, The method further comprises: A plurality of databases are constructed using distributed storage technology to hierarchically and classifiedly manage the preprocessed multi-source heterogeneous data; the databases at least comprise an original database, a resource database, a theme database, and a business database; The original database is used to store raw data without preprocessing; the resource database is used to store preprocessed multi-source heterogeneous data; the theme database is used to classify and integrate multi-source heterogeneous data according to different disaster types; the business database is used to customize data storage according to preset business requirements; when the original database obtains new data, the new data is synchronized to other databases in real time.

8. A FIPSO and FINN based disaster risk data mining analysis system, applying the method of any one of claims 1-7, characterized in that, Comprise: A data preprocessing module for collecting multi-source heterogeneous data corresponding to disaster risk to be analyzed and preprocessing; A data mining and analysis module for mining and analyzing the preprocessed multi-source heterogeneous data based on the FIPSO algorithm and the FINN neural network to obtain the analysis result of the disaster risk; In the data mining analysis module, the FIPSO algorithm is used to initialize parameters of the FINN neural network, guide iterative optimization of the FINN neural network, and trigger a Levy flight mechanism to guide the FINN neural network to jump out of a local optimum when the iterative optimization falls into stagnation. The FINN neural network comprises a perception layer, a hidden layer and an output layer connected in sequence, and the hidden layer is provided with a parameter perturbation mechanism for guiding the FINN neural network to jump out of a local optimum when an output error of the output layer is greater than a preset threshold.

Images