Park fire prediction system and method based on multi-source data fusion analysis

Abstract

The invention provides a park fire prediction system and method based on multi-source data fusion analysis, and relates to the field of fire prediction. Comprising a data acquisition module used for acquiring multi-source data including video data, environment sensor data, user feedback data and historical fire event data in a park, a data preprocessing module used for carrying out preprocessing operations including data cleaning, format unification, abnormal value processing and data enhancement on the acquired multi-source data, and a data processing module used for processing the acquired multi-source data. The consistency and integrity of the data are ensured; the analysis decision module performs fire identification and risk analysis on the preprocessed data, specifically comprises an image and video identification unit, a distributed intelligent agent cooperation unit and a biological inspiration algorithm optimization unit, and generates a fire prediction result and an emergency decision suggestion; and the feedback optimization module is used for continuously improving the accuracy and the stability of the system through trusted data fusion and a self-adaptive updating mechanism based on user feedback and actual fire event data.

Description

Technical Field

[0001] The present invention relates to the field of fire prediction, and in particular to a park fire prediction system and method based on multi-source data fusion analysis. Background Art

[0002] With the acceleration of urbanization, park safety issues have received increasing attention. Fire, as one of the main safety hazards in parks, has become a research hotspot in prevention and control technology. The development of technology has enabled fire prediction systems to warn of fires by analyzing environmental data and surveillance videos, providing technical support for park safety management.

[0003] Currently, many fire prediction systems rely mainly on a single data source, such as environmental sensors or video surveillance. These systems often lack comprehensive analysis capabilities and are unable to fully capture the complex signs before a fire occurs. For example, some technologies only use temperature sensors to monitor ambient temperature changes, while ignoring other key factors such as smoke and humidity, which limits their accuracy and reliability in early fire detection.

[0004] The shortcomings of existing technologies lie in the limitations of their data processing and analysis capabilities, as well as the lack of multi-source data fusion, which leads to the emergence of information islands and the inability to form a unified fire risk assessment. In addition, the analysis results of a single data source often lack depth and breadth and cannot provide sufficient early warning information and emergency decision-making support. Summary of the invention

[0005] 1. Technical issues to be resolved

[0006] In view of the shortcomings of the prior art, the present invention provides a campus fire prediction system and method based on multi-source data fusion analysis, which aims to integrate video surveillance, environmental sensor data and user feedback, and use deep learning models and distributed intelligent agent technology to solve the limitations of the data processing and analysis capabilities of the above-mentioned prior art, as well as the shortcomings of multi-source data fusion, which lead to the emergence of information islands and the inability to form a unified fire risk assessment. The analysis results of a single data source often lack depth and breadth and cannot provide sufficient early warning information and emergency decision support.

[0007] (II) Technical solution

[0008] To achieve the above objectives, the present invention is implemented through the following technical solutions: a park fire prediction system and method based on multi-source data fusion analysis, including: a data acquisition module, a data preprocessing module, an analysis and decision module, and a feedback optimization module;

[0009] The data acquisition module is used to collect multi-source data in the park, including video data, environmental sensor data, user feedback data and historical fire event data;

[0010] The data preprocessing module performs preprocessing operations such as data cleaning, format unification, outlier processing and data enhancement on the collected multi-source data to ensure the consistency and integrity of the data;

[0011] The analysis and decision-making module: performs fire identification and risk analysis on the pre-processed data, specifically including an image and video recognition unit, a distributed intelligent agent collaboration unit, and a bio-inspired algorithm optimization unit, and generates fire prediction results and emergency decision-making suggestions;

[0012] The feedback optimization module optimizes system parameters and prediction models in real time based on user feedback and actual fire event data, and continuously improves the accuracy and stability of the system through trusted data fusion and adaptive update mechanism.

[0013] Preferably, the data acquisition module is the basis of the system. Through the installation of various sensors and video surveillance equipment in the park, the sensors include temperature, smoke and humidity sensors, and continuously monitor the environmental changes of the park. The video surveillance equipment uses high-resolution cameras to capture video streams of various areas of the park in real time. These video streams are encrypted and transmitted to the data preprocessing module to ensure the security and real-time nature of the data; the environmental sensors are arranged at key locations in the park, regularly collect data on temperature, humidity and smoke concentration, and send the data to the data preprocessing module in real time through wireless networks; at the same time, user feedback data is collected through mobile applications or on-site terminals, and staff can report abnormal conditions and fire event information at any time to ensure that the system can respond to potential fire risks in a timely manner.

[0014] Preferably, the data preprocessing module performs necessary cleaning and formatting of the collected multi-source data to improve the accuracy of subsequent analysis. In the data cleaning step, the data quality is ensured by removing invalid or erroneous data points. Assume that the original data set is D = {d1, d2, ···, d n}, for each data point d i , using an outlier detection algorithm to perform preliminary screening; wherein the outlier detection can be based on statistical methods, such as assuming that the data follows a normal distribution, for data point d i , if its value is greater than μ+kσ or less than μ-kσ, then d i The outliers are removed, where μ is the mean, σ is the standard deviation, and k is a constant determined based on experience or domain knowledge, generally k = 3; then the data from different devices are converted into a unified format to ensure that the data can be seamlessly compatible and integrated; then the data format collected by device A is set to F A =(a1,a2,···,a m ), the data format collected by device B is F B=(b1,b2,···,b p ), through a specific transformation function T, so that F A =T(F B ) or F B =T(F A ) to achieve the unification of data formats.

[0015] Preferably, after completing the format unification, the data enhancement technology is applied to randomly rotate, scale and add noise to the video data to enhance the adaptability of the model. For a certain frame image I in the video, the rotation matrix of its rotation operation is set to R(θ), where θ is the rotation angle, and θ is in [-θ max ,θ max ] is randomly selected within the range and set according to the actual situation. The scaling factor of the scaling operation is s, and s is in [s min ,s max ] is randomly selected within the range, s min and max is the preset value, add noise N, the noise can obey Gaussian distribution μ n and σ n It can be set according to the noise intensity requirement, and the image I′ after data enhancement can be expressed as: I′=R(θ)*(s*I+N); through these steps, effective preprocessing of multi-source data is achieved, providing high-quality and uniformly formatted data for subsequent analysis and model training.

[0016] Preferably, in the analysis and decision-making module, the fire feature recognition unit performs real-time recognition of fire features through deep learning technology; first, the convolutional neural network (CNN) is used to extract spatial features of the video data to identify image features such as flames and smoke, and these features are then sent to the recurrent neural network (RNN), which captures the time series features between the video data through its hidden layer and identifies the change trend of the fire features over time. In this process, the output of the CNN is used as the input of the RNN, so that the RNN can use the spatial features extracted by the CNN to analyze the dynamic changes of the fire over time; the specific operation is that in the convolution layer of the CNN, for an input video data X and a convolution kernel K, the dimension of the video data is H*W*C, H is the height, W is the width, and C is the number of channels; the dimension of the convolution kernel is h*w*C*n, where h and w are the height and width of the convolution kernel, and n is the number of convolution kernels. The convolution operation can be expressed as:

[0017]

[0018] Among them, Y is the output feature map, σ is the activation function, ij is the coordinate of the output feature map, l represents the number of layers, and b is the bias term; in the pooling layer, assuming that the maximum pooling is used and the pooling window size is k*k, the calculation formula is: Where Z is the output after pooling; then the recurrent neural network (RNN) is used to capture the time series features between video data to identify the temporal trend of fire characteristics. The specific operation is as follows: first, for a simple RNN unit, given the input, x t In the video data sequence, t represents the time step and the hidden state h t The update formula is:

[0019]

[0020] Among them, W ih and W hh is the weight matrix, b h is a bias term; when processing video data sequences to predict the trend of fire characteristics, assuming that the output y t With the hidden state h t There is a linear relationship, then: t =W hy h t +b y ; Among them, W hy Output weight matrix, b y is the output bias term.

[0021] Preferably, in order to further enhance the comprehensiveness of feature extraction, the visual Transformer (ViT) performs global feature encoding on the video data and calculates the correlation between each image region through a multi-head self-attention mechanism. The specific operation is as follows: first, the input video data I is divided into N image blocks X P (P=1,2,…,N), each image block is represented by a vector after linear mapping. Assuming that the input image block dimension is P*P*C, a vector with dimension D is obtained after linear mapping E, then z0=[x1E,x2E,…,x n E]+E pos ; Among them, E pos is a position embedding vector used to encode the position information of the image block; in the multi-head self-attention mechanism, for the self-attention calculation of a single head, assuming that the query vector key Q, vector K and value vector V are all of dimension N*D, the attention calculation is V; where d k is the dimension of the key vector, and the multi-head self-attention is the concatenation of the calculation results of multiple heads and then linear transformation; after all features are extracted, the transfer learning technology is used to fine-tune the model on a specific fire-related data set to improve the system's adaptability to the park environment and prediction accuracy.

[0022] Preferably, the distributed intelligent agent collaboration unit is composed of regional agents, global agents and coordination agents to form an efficient data processing and analysis system; the regional agent is responsible for the collection and preliminary analysis of video recognition results and environmental data in the region, and sends the processed information to the global agent; the global agent aggregates the data from the regional agents, conducts an overall fire risk analysis in the park, and forms a comprehensive report; the coordination agent directs the collaboration and resource allocation between regional agents based on the global analysis results to achieve rapid response and information sharing, and at the same time introduces a semantic-based adaptive communication protocol to dynamically adjust the communication bandwidth according to the importance and urgency of the information to ensure the priority transmission of critical information.

[0023] Preferably, the bio-inspired algorithm optimization unit is based on a non-dominated sorting genetic algorithm (NSGA-II) and a reinforcement learning mechanism to optimize the fire emergency response strategy. The bio-inspired algorithm defines multiple optimization objectives, such as minimizing evacuation time, maximizing resource utilization, and reducing casualties. It stratifies population individuals through non-dominated sorting and uses crowding distance calculation to maintain population diversity. The formula for crowding distance calculation is:

[0024]

[0025] Among them, M is the number of targets, f is the objective function value, and the natural selection process is simulated by genetic operations of selection, crossover and mutation to search for the optimal solution that meets multiple goals such as minimizing evacuation time, maximizing resource utilization and reducing casualties. At the same time, combined with the reinforcement learning technology, according to the effect of personnel evacuation and resource utilization in actual fire incidents, the parameters and weights of the optimization algorithm are dynamically adjusted to ensure the effectiveness of emergency decision-making in different scenarios. In addition, the introduction of emotional computing technology enables the system to adjust the evacuation strategy according to the emotional state of on-site personnel, thereby improving the rationality and safety of decision-making.

[0026] Preferably, the core of the feedback optimization module is to ensure the continuous optimization of the system by means of trusted data fusion and model adaptive update; the construction of the trusted data fusion mechanism is crucial, and for user feedback and actual fire event data, credibility assessment is carried out from multiple dimensions such as data source reliability, data integrity, consistency between data and historical data, and stability of the data collection environment; specifically, assuming that the data source reliability weight is w1, the integrity weight is w2, the consistency weight with historical data is w3, and the collection environment stability weight is w4, for a certain data D i , its credibility C i The calculation formula is: C i =w1*r i +w2*c i+w3*h i +w4*s i , where r i 、c i 、h i 、s i The scores of the data in terms of source reliability, completeness, consistency with historical data, and stability of the collection environment (scores range from 0 to 1). In this way, each piece of data is given a credibility weight, and high-credibility data is used first in subsequent analysis. When the data uncertainty is high, the system will judge based on the pre-set uncertainty threshold τ. If the mean credibility of a certain type of data is It actively requests feedback from users to further improve the accuracy of the data; then it uses model fusion and online learning technology to receive new data in real time during the operation of the system. new , an online learning algorithm based on a variant of stochastic gradient descent, updates the model parameters θ according to the new data, and the update formula is Among them, α is the learning rate, L is the loss function, and the model fusion strategy combined with weighted average fusion is used to dynamically adjust the weight and structure of the model to better adapt to changes in the park environment and respond to emerging fire risks.

[0027] Preferably, the data enhancement technology plays a vital role in the preprocessing module to improve the adaptability of the subsequent analysis module to data changes; in the specific implementation process, the collected video data is randomly rotated, scaled, cropped and noise added, and these enhancement operations ensure that the generated new data still retains the characteristics of the original data by setting appropriate parameters and ranges, while providing a wider range of training samples for the deep learning model. For example, when rotating the video data, the rotation angle is set between -30° and 30° to increase the model's ability to recognize fire characteristics from different viewing angles; this process is automated, so that data enhancement not only improves the robustness of the model, but also saves time for manual processing.

[0028] (III) Beneficial effects

[0029] The present invention provides a park fire prediction system and method based on multi-source data fusion analysis. It has the following beneficial effects:

[0030] 1. By integrating multi-source data and using deep learning models to identify fire characteristics, it is possible to accurately detect and predict signs of fire. Real-time analysis of video data and rapid processing of sensor data enable the system to issue early warnings in the early stages of a fire, significantly improving the accuracy of fire identification. In addition, the deep learning model is continuously optimized through transfer learning technology and can continuously adapt to changes in the park environment, ensuring the system's efficient response under different conditions and reducing losses after a fire.

[0031] 2. The system uses a distributed intelligent agent architecture, combined with a non-dominated sorting genetic algorithm (NSGA-II) and reinforcement learning to optimize emergency response strategies. It can dynamically adjust resource allocation according to the fire risk situation in the park. This process not only takes into account minimizing evacuation time and maximizing resource utilization, but also combines emotional computing technology to analyze the emotions of on-site personnel and optimize evacuation routes. Through efficient collaboration between intelligent agents, the system can make quick decisions when a fire occurs, reasonably allocate emergency resources, and ensure the safety of personnel and the protection of park property. At the same time, through online learning and feedback optimization mechanisms, it ensures that the emergency response plan is further improved after each application to cope with fire incidents of different types and scales. DETAILED DESCRIPTION

[0032] The technical solutions in the embodiments of the present invention are described clearly and completely below. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of them. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0033] The embodiment of the present invention provides a park fire prediction system and method based on multi-source data fusion analysis, wherein the basis of the system is a data acquisition module, which realizes real-time acquisition of multi-source data through multiple environmental sensors and video surveillance equipment in the park. First, temperature, humidity and smoke sensors are deployed at various key locations in the park to monitor environmental changes in real time, and the collected data are transmitted to the central processing system through a wireless network. In addition, the video surveillance equipment uses a high-resolution camera to comprehensively monitor the park and capture possible signs of fire. These video data will be encrypted and transmitted to the image and video recognition module; the system also integrates a user feedback function, through which park staff use mobile applications or on-site terminals to report signs of fire or abnormal events, further enhancing the data acquisition capability of the system.

[0034] After the data acquisition module transfers the data to the data preprocessing module, data cleaning is performed first to ensure the quality of the data. Invalid and erroneous data points are removed through the outlier detection algorithm. Abnormal data collected by temperature, humidity or smoke sensors are mainly identified through statistical methods and eliminated. Then, sensor data from different devices are uniformly formatted to ensure that various types of data can be seamlessly integrated and adapted to subsequent analysis modules. In terms of video data, the system uses data enhancement techniques such as random rotation, scaling, and noise addition to enhance the model's adaptability to various scene changes. All cleaned and enhanced data are standardized to provide consistent and high-quality input for deep learning models.

[0035] The preprocessed high-quality data enters the analysis and decision-making module. First, the fire feature recognition unit performs deep learning analysis. This module uses a combined architecture of convolutional neural network (CNN), recurrent neural network (RNN) and visual Transformer (ViT) to comprehensively analyze the video data. CNN is used to extract spatial features from video data and identify fire features such as flames and smoke; RNN captures the time series features between video data to identify the time trend of fire development, and then combines the two. Finally, ViT analyzes the overall structure of the image at the global level and calculates the correlation between image regions through the self-attention mechanism, thereby improving the accuracy and comprehensiveness of fire feature extraction. Then, through transfer learning technology, pre-training is performed on a large-scale fire-related data set, and then the model is fine-tuned through park data to improve its adaptability to the park environment.

[0036] The data after fire feature recognition is transmitted to the distributed intelligent agent collaboration unit. This unit efficiently processes data and performs risk analysis through a three-layer structure of regional agents, global agents, and coordination agents. The regional agent is responsible for collecting and preliminarily analyzing the environmental data and video recognition results in the region, and sending the analysis results to the global agent. The global agent then aggregates the data from all regional agents and analyzes the overall fire risk in the park. The coordination agent coordinates the emergency response of each region based on the global analysis results, allocates resources and schedules tasks to ensure the efficiency of the emergency response. At the same time, in order to improve the efficiency of information transmission, a semantic-based adaptive communication protocol is adopted to dynamically adjust the communication bandwidth according to the urgency of the data, and differential privacy technology is introduced in data transmission to ensure the privacy and security of sensitive data.

[0037] In order to further optimize the emergency response strategy, the system applies a bio-inspired algorithm optimization unit. This module is based on the non-dominated sorting genetic algorithm (NSGA-II) and reinforcement learning technology. Combined with the actual situation of the park, it performs multi-objective optimization on the evacuation path and resource allocation in the fire emergency response. The non-dominated sorting algorithm is used to evaluate multiple optimization objectives, such as minimizing the evacuation time, maximizing resource utilization, and reducing casualties. The crowding distance calculation is used to maintain the diversity of the population and ensure the comprehensiveness of the search process. Reinforcement learning is combined in the optimization process, and the weights and strategies of the algorithm are adjusted according to real-time feedback to ensure that the optimal emergency plan is always selected in the ever-changing fire scenario.

[0038] The feedback optimization module serves as the adaptive adjustment core of the system. It can optimize the system's prediction and decision-making models in real time based on user feedback and actual fire event data. It evaluates the credibility of each piece of user feedback data through a trusted data fusion mechanism to ensure that high-credibility data is processed first. The system assigns a credibility weight to each piece of data by evaluating the reliability of the data source, the integrity of the data, the consistency of historical data, and the stability of the data collection environment. In scenarios with high data uncertainty, the system will actively request feedback information from users to further improve data accuracy. By using online learning technology and model fusion strategies, the system can adjust the weight and structure of the model in real time to better adapt to changes in the park environment and improve the accuracy and stability of fire predictions.

[0039] Finally, park managers can obtain fire predictions and emergency decision-making suggestions through the real-time visualization interface provided. Through the three-dimensional virtual scene, managers can clearly see the fire risks, personnel distribution and resource scheduling in the park. In an emergency, the system will automatically adjust the emergency plan according to the latest fire prediction data and provide managers with the most appropriate resource allocation suggestions. Through this comprehensive system, the park's fire response capabilities have been significantly improved. It can not only detect and predict fire risks in a timely manner, but also make emergency responses quickly when a fire occurs, thereby maximizing the safety of personnel and park property.

[0040] Although embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions and variations may be made to the embodiments without departing from the principles and spirit of the present invention, and that the scope of the present invention is defined by the appended claims and their equivalents.

Claims

1. A park fire prediction system based on multi-source data fusion analysis, comprising: The data acquisition module, data preprocessing module, analysis and decision module, and feedback optimization module are characterized by: The data acquisition module is used to collect multi-source data in the park, including video data, environmental sensor data, user feedback data and historical fire event data; The data preprocessing module performs preprocessing operations such as data cleaning, format unification, outlier processing and data enhancement on the collected multi-source data to ensure the consistency and integrity of the data; The analysis and decision-making module performs fire identification and risk analysis on the preprocessed data, including an image and video recognition unit, a distributed intelligent agent collaboration unit, and a bio-inspired algorithm optimization unit, and generates fire prediction results and emergency decision-making suggestions; The feedback optimization module is based on user feedback data and actual fire event data, and continuously improves the accuracy and stability of the system through trusted data fusion and adaptive update mechanism.

2. A campus fire prediction system based on multi-source data fusion analysis according to claim 1, characterized in that: The data preprocessing module includes: Perform data cleaning on video data and sensor data to remove invalid or erroneous data points; Convert data from different sources into a unified format to ensure data compatibility; Apply outlier detection algorithms to identify and handle outliers in environmental data; Perform data enhancement on video data, including random rotation, scaling, and noise addition, to improve the adaptability of subsequent analysis modules to data changes.

3. A campus fire prediction system based on multi-source data fusion analysis according to claim 2, characterized in that: The analysis and decision-making module includes: The image and video recognition unit uses a deep learning architecture model that integrates convolutional neural networks (CNN), recurrent neural networks (RNN), and visual transformers (ViT) to identify fire features in video data in real time, and introduces transfer learning to improve the adaptability of the recognition model in the park environment; Distributed intelligent agent collaboration unit, including a three-layer structure of regional agent, global agent and coordination agent, realizes distributed data processing and transmission in the park through an event-driven collaboration mechanism; The bio-inspired algorithm optimization unit optimizes the evacuation path and resource allocation plan of fire emergency response based on the non-dominated sorting genetic algorithm (NSGA-II) for multi-objective optimization and reinforcement learning mechanism.

4. The park fire prediction system based on multi-source data fusion analysis according to claim 3 is characterized by: The specific steps of the image and video recognition unit include: Using the CNN to extract spatial features from the video data collected in the data collection module, identifying image features of flames and smoke, and gradually extracting feature maps through its convolutional layer and pooling layer; The RNN uses the spatial features extracted by CNN to further capture the changing trend of time series features in video data; The ViT uses a multi-head self-attention mechanism to encode global features of video data, identify the correlation between different image regions and enhance the comprehensiveness of fire feature extraction; The recognition accuracy and generalization ability of the recognition model are improved by fine-tuning the video data through transfer learning technology.

5. The park fire prediction system based on multi-source data fusion analysis according to claim 4 is characterized by: The image and video recognition unit combines a multi-model fusion strategy, utilizes the characteristic advantages of different deep learning models, and fuses the fire recognition results of CNN, RNN and ViT models through a weighted average mechanism to improve the accuracy and robustness of recognition. The multi-model fusion strategy includes a dynamic weight allocation mechanism based on fire scene characteristics, which can automatically adjust the weight of each model according to the complexity of the environment.

6. The park fire prediction system based on multi-source data fusion analysis according to claim 3 is characterized by: The distributed intelligent agent collaboration unit specifically includes: Regional agent, responsible for the collection and preliminary analysis of the video recognition results and environmental data in the region; The global agent aggregates data from all regional agents to conduct an overall fire risk analysis within the park; Coordination agent, which coordinates emergency response and resource allocation among regions based on the global analysis results; A semantics-based adaptive communication protocol is introduced to automatically adjust the communication bandwidth and transmission frequency according to the urgency of the information.

7. The park fire prediction system based on multi-source data fusion analysis according to claim 3 is characterized in that: The bioinspired algorithm optimization unit optimizes the emergency response through the following steps: A non-dominated sorting genetic algorithm (NSGA-II) is used for multi-objective optimization, taking into account emergency goals such as evacuation time, resource utilization, and personnel safety; Introducing a reinforcement learning mechanism to dynamically adjust the weights and strategies of the optimization algorithm based on the evacuation effect and resource usage in actual fire events; By incorporating affective computing technology into the optimization algorithm, the emotional state of personnel is taken into consideration in the optimization of the evacuation route to ensure the rationality and safety of the evacuation route.

8. The method for predicting park fires based on multi-source data fusion analysis according to claim 1 is characterized in that: The data collection module step is to collect video data, environmental sensor data, user feedback data and historical fire event data in the park to build a comprehensive fire prediction data set; The data preprocessing module steps include performing preprocessing operations such as data cleaning, format unification and outlier processing on the collected multi-source data to ensure the consistency and integrity of the data; The analysis and decision-making module step is to identify the fire through the image and video recognition unit, the distributed intelligent agent collaboration unit and the bio-inspired algorithm optimization unit for the data preprocessed by the data preprocessing module, and generate fire prediction results and emergency decision-making suggestions; The steps of the feedback optimization module are to collect user feedback data and actual fire event data, optimize system parameters and prediction models in real time, and continuously improve the accuracy and stability of the system through trusted data fusion and adaptive update mechanism.

Images