A fire risk prediction and early warning method and system based on big data analysis

Abstract

The application discloses a fire risk prediction and early warning method and system based on big data analysis, and relates to the technical fields of fire prevention and control and big data analysis. The method comprises the following steps: S1, acquiring multi-source heterogeneous data of a target region, wherein the multi-source heterogeneous data comprises fire history data, environmental parameter data, building information data and power load data, the environmental parameter data comprises temperature, humidity, air pressure and wind speed, and the building information data comprises building structure, material type and fire-fighting facility condition. The application acquires and pre-processes multi-source heterogeneous data such as fire history data, environmental parameter data, building information data and power load data of the target region, so as to guarantee data quality; time sequence characteristics of each fire risk factor under a plurality of preset time scales of hour level, day level and week level are extracted, and a multi-scale time sequence fusion model is combined to capture the evolution trend of the fire risk factor in continuous time and the deep time sequence correlation between the fire risk factor and fire occurrence.

Description

Technical Field

[0001] This invention relates to the field of fire prevention and control and big data analysis technology, specifically to a method and system for fire risk prediction and early warning based on big data analysis. Background Technology

[0002] Fire is a sudden and destructive disaster that poses a serious threat to people's lives and property. Traditional fire prevention and control mainly rely on manual inspections and real-time sensor monitoring. The former is inefficient and has limited coverage, while the latter often only triggers alarms after a fire has occurred, lacking the ability to predict risks beforehand. In recent years, with the development of the Internet of Things and big data technologies, analyzing multi-source information such as historical fire data, environmental parameters, and building information to achieve advanced prediction and graded early warning of fire occurrence probabilities has become an important research direction for improving public safety. These methods can uncover potential patterns in fire occurrence from large amounts of data, providing a scientific basis for the optimal allocation of fire prevention resources and precise intervention.

[0003] However, existing fire prediction models based on big data analytics generally suffer from insufficient utilization of temporal characteristics. Specifically, the formation of fire risk is often a dynamic and cumulative process, with related influencing factors such as temperature, humidity, and electrical load constantly changing over time. Moreover, the contribution of fluctuation patterns at different time scales to risk formation varies. Existing methods mostly use static or fixed-time-window data for modeling, making it difficult to capture the evolution trend of risk factors over continuous time and their deep temporal correlation with fire occurrence. This results in insufficient accuracy in identifying short- to medium-term fire risks, easily leading to false alarms or missed alarms, and reducing the reliability and usability of early warning information. Summary of the Invention

[0004] To address the shortcomings of existing technologies, this invention provides a fire risk prediction and early warning method and system based on big data analysis, which solves the problems of insufficient utilization of time-series features, insufficient accuracy in judging short- and medium-term fire risks, and susceptibility to false alarms and missed alarms compared with existing technologies.

[0005] To achieve the above objectives, the present invention provides the following technical solution: a fire risk prediction and early warning method based on big data analysis, comprising: S1. Acquire multi-source heterogeneous data of the target area. The multi-source heterogeneous data includes fire history data, environmental parameter data, building information data and power load data. Environmental parameter data includes temperature, humidity, air pressure and wind speed. Building information data includes building structure, material type and fire protection facility status. S2. From multi-source heterogeneous data, extract time-series features at multiple preset time scales for each fire risk factor. The multiple preset time scales include hourly, daily, and weekly levels. The time-series features represent the dynamic change pattern of the fire risk factor at the corresponding time scale. The dynamic change pattern includes trend, periodicity, volatility, or autocorrelation. S3. Based on the temporal characteristics of multiple preset time scales, a multi-scale temporal fusion model is constructed. The multi-scale temporal fusion model is used to capture the evolution trend of fire risk factors in continuous time and their deep temporal correlation with fire occurrence. S4. Calculate the predicted fire risk value of the target area using a multi-scale time-series fusion model. Fire risk prediction value This indicates the probability of a fire occurring or the fire risk level, where the fire risk prediction value is... The calculation formula is: ; in, Indicates from the first Feature vectors extracted at each time scale and processed by a multi-scale temporal fusion model and For the model's weights and biases, For activation function, This is a vector concatenation operation; S5. Generate and output fire early warning information based on fire risk prediction values.

[0006] Furthermore, after acquiring the multi-source heterogeneous data of the target region, the process further includes: Preprocessing operations are performed on multi-source heterogeneous data. These operations include data cleaning, missing value imputation, outlier detection and handling, and normalization or standardization.

[0007] Furthermore, the step of extracting time-series features at multiple preset time scales from multi-source heterogeneous data for each fire risk factor includes: Sliding window sampling is performed on multi-source heterogeneous data, where the length of each sliding window corresponds to a preset time scale; Statistical characteristics are calculated for the data within each sliding window. These characteristics include the mean, maximum, minimum, standard deviation, slope, kurtosis, autocorrelation coefficient, and Fourier transform spectral characteristics to comprehensively describe the dynamic changes within that time scale.

[0008] Furthermore, the construction of the multi-scale temporal fusion model includes: Feature encoding is performed on time-series features at different time scales to obtain feature vectors for each time scale. The feature vectors at their respective time scales are concatenated or weighted and fused, and then input into a neural network model for nonlinear transformation to generate a fused feature representation. The neural network model includes recurrent neural networks, temporal convolutional networks, or network structures based on attention mechanisms.

[0009] Furthermore, the feature encoding of temporal features at different time scales includes: An independent temporal encoder is used to process the temporal feature sequence at each time scale. The temporal encoder includes a long short-term memory network unit, a gated recurrent unit, or a one-dimensional convolutional layer to extract and compress the sequence information at the corresponding time scale.

[0010] Furthermore, the generation and output of fire early warning information includes: Fire risk prediction value The fire risk level is determined by comparing it with multiple preset risk thresholds. Fire risk levels include at least three levels: low risk, medium risk, and high risk. Based on the determined fire risk level, a corresponding early warning signal or message is generated, which can be distributed via SMS, application push, email, or audible and visual alarms.

[0011] Furthermore, the method also includes: Based on actual fire event data, the multi-scale time series fusion model is trained and iteratively optimized. Training employs a supervised learning method, which minimizes the loss function between the predicted value and the actual label. The loss function can be either the cross-entropy loss function or the mean squared error loss function.

[0012] Furthermore, the method also includes: Upon receiving a fire warning, the system intelligently recommends corresponding fire prevention strategies or emergency response measures based on the fire risk prediction value and fire risk level. Fire prevention strategies include resource allocation, personnel evacuation route planning, or enhanced monitoring of key areas.

[0013] This invention also provides a fire risk prediction and early warning system based on big data analysis, comprising: The data acquisition module is used to acquire multi-source heterogeneous data of the target area, including fire history data, environmental parameter data, building information data, and electricity load data. The time-series feature extraction module is used to extract time-series features at multiple preset time scales from multi-source heterogeneous data for each fire risk factor. The time-series feature extraction module includes a sliding window sampling unit and a statistical feature calculation unit. The model building and prediction module is used to build a multi-scale time series fusion model based on the time series characteristics under multiple preset time scales, and to use the multi-scale time series fusion model to calculate the fire risk prediction value of the target area. The early warning generation module is used to generate and output fire early warning information based on fire risk prediction values. The early warning generation module can intelligently recommend fire prevention strategies or emergency response measures according to the risk level.

[0014] Furthermore, the model building and prediction module also includes: The feature encoding unit is used to encode the temporal features at different time scales to obtain the feature vectors at their respective time scales. The fusion processing unit is used to concatenate or weightedly fuse feature vectors from their respective time scales. The prediction unit is used to input the fused features into the neural network model for nonlinear transformation in order to calculate the fire risk prediction value.

[0015] Compared with the prior art, the beneficial effects of the present invention are as follows: This invention acquires and preprocesses multi-source heterogeneous data, including historical fire data, environmental parameter data, building information data, and electricity load data, of the target area to ensure data quality. It extracts time-series features at multiple preset time scales (hourly, daily, and weekly) for each fire risk factor, and combines this with a multi-scale time-series fusion model to capture the evolution trend of fire risk factors over continuous time and their deep temporal correlation with fire occurrence. This effectively solves the problems of insufficient utilization of time-series features, insufficient accuracy in short-to-medium-term fire risk identification, and susceptibility to false alarms and missed alarms in existing technologies, thus improving the reliability and usability of early warning information. Furthermore, it provides tiered early warnings based on fire risk prediction values ​​and distributes early warning information through multiple channels. It can also intelligently recommend corresponding fire prevention strategies or emergency response measures, further improving the efficiency of fire prevention and emergency response, and providing scientific support for fire prevention in the target area. Attached Figure Description

[0016] Figure 1 This is a flowchart of the method of the present invention; Figure 2 This is a flowchart of the multi-scale temporal feature extraction process of the present invention; Figure 3 This is a structural diagram of the multi-scale temporal fusion model of the present invention; Figure 4 This is a system structure diagram of the present invention. Detailed Implementation

[0017] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0018] Please see Figure 1-3 This invention provides a fire risk prediction and early warning method based on big data analysis, comprising: S1. Acquire multi-source heterogeneous data of the target area. The multi-source heterogeneous data includes fire history data, environmental parameter data, building information data and power load data. Environmental parameter data includes temperature, humidity, air pressure and wind speed. Building information data includes building structure, material type and fire protection facility status. S2. From multi-source heterogeneous data, extract time-series features at multiple preset time scales for each fire risk factor. The multiple preset time scales include hourly, daily, and weekly levels. The time-series features represent the dynamic change pattern of the fire risk factor at the corresponding time scale. The dynamic change pattern includes trend, periodicity, volatility, or autocorrelation. S3. Based on the temporal characteristics of multiple preset time scales, a multi-scale temporal fusion model is constructed. The multi-scale temporal fusion model is used to capture the evolution trend of fire risk factors in continuous time and their deep temporal correlation with fire occurrence. S4. Calculate the predicted fire risk value of the target area using a multi-scale time-series fusion model. Fire risk prediction value This indicates the probability of a fire occurring or the fire risk level, where the fire risk prediction value is... The calculation formula is: ; in, Indicates from the first Feature vectors extracted at each time scale and processed by a multi-scale temporal fusion model and For the model's weights and biases, For activation function, This is a vector concatenation operation; S5. Generate and output fire early warning information based on fire risk prediction values.

[0019] Specifically, when conducting fire risk prediction and early warning in a city's core commercial area, historical fire data is retrieved from the local fire department's fire accident database, covering information such as the time, location, cause, and extent of damage of fires in and around the commercial area over the past eight years. Environmental parameter data is collected through monitoring equipment deployed on the rooftops of various buildings, street green belts, and underground parking garages in the commercial area. Temperature and humidity are recorded every 5 minutes by high-precision temperature and humidity sensors, air pressure data comes from digital barometers connected to meteorological stations, and wind speed is collected in real time by outdoor anemometers. Building information data is obtained from the housing and construction department's archive system and property operation and maintenance platform, including the structural type of each building, such as frame structure or brick-concrete structure, the type of building materials, such as whether flammable insulation materials are used, and the status of fire protection facilities, such as the expiration date of fire extinguishers and the operating status of automatic sprinkler systems. Electricity load data is collected through the power company's smart meter network, obtaining the power consumption data of each building every 15 minutes.

[0020] After data acquisition, time-series features at three preset time scales—hourly, daily, and weekly—were extracted for key fire risk factors such as temperature, humidity, and electricity load. The hourly time scale uses a 60-minute sliding window to capture dynamic changes such as sudden increases in electricity load or temperature rises within a short period. The daily time scale uses a 24-hour sliding window to reflect the periodic changes between daytime peak electricity consumption and nighttime troughs. The weekly time scale uses a 168-hour sliding window to uncover the weekly trend of electricity load differences between weekdays and weekends. In the dynamic change patterns of the time-series features, trend features can be obtained through linear fitting, such as the rising or falling trend of temperature within a continuous 12-hour window; periodic features are obtained through Fourier transform analysis, such as the frequency of peak electricity load occurrences within 24 hours; volatility features are reflected by calculating the variance of the difference between adjacent window data, such as the dramatic fluctuations in humidity during rainy weather; and autocorrelation features are obtained by calculating the correlation coefficient between the data sequence and its own lagged sequence, such as the similarity of electricity load at the same time each day.

[0021] When constructing a multi-scale time-series fusion model based on extracted multi-scale time-series features, the time-series features at each time scale are first encoded separately, and then fusion is achieved through vector concatenation. Taking temperature features as an example, hourly temperature time-series features are encoded using Long Short-Term Memory (LSTM) network units, daily temperature time-series features are encoded using gated recurrent units (GRUs), and weekly temperature time-series features are encoded using one-dimensional convolutional layers. The output dimension of each encoder is set to a 32-dimensional feature vector, and all features are normalized using Min-Max before encoding. The normalization formula is: ; in These are the normalized eigenvalues. These are the original eigenvalues. This is the minimum value of the feature. This is the maximum value of this feature.

[0022] The encoded hourly, daily, and weekly feature vectors are concatenated into a 96-dimensional fusion vector using the concat operation. This fusion vector is then input into an attention-based neural network model for nonlinear transformation. This model enhances the contribution of features that significantly affect fire risk through attention weight allocation, thereby capturing the evolution trend of risk factors over continuous time and their deep temporal correlation with fire occurrence.

[0023] When calculating the predicted fire risk using the constructed model, the following formula is used: ; in Let t be the predicted fire risk value at time t, with a value ranging from 0 to 1, used to represent the probability of fire occurrence; The 32-dimensional feature vector extracted at the hourly timescale at time t and processed by the model is given. Let t be a 32-dimensional feature vector on a daily timescale. This is a 32-dimensional feature vector at time t with a week-scale time scale; The model weight matrix is ​​a 96-dimensional × 1-dimensional matrix, which consists of parameters obtained through model training and is used to perform linear transformation on the fusion vector. These are the model bias parameters; This is an activation function used to map the result of a linear transformation to the interval between 0 and 1. This is a vector concatenation operation that concatenates three 32-dimensional feature vectors into a 96-dimensional fused vector. Before concatenation, each feature vector has been normalized to ensure consistent dimensions and the rationality of the linear transformation.

[0024] according to When generating early warning information, risk assessment criteria are set based on the fire prevention and control needs of the commercial area. A low-risk warning is generated when the value is less than 0.3, prompting property management to strengthen daily inspections; when... A medium-risk warning is generated when the risk level is between 0.3 and 0.7, notifying fire patrol personnel to increase the frequency of area patrols; when... A high-risk warning is generated when the value exceeds 0.7, directly linking with the fire command center to prepare for emergency response. The warning information is simultaneously distributed through the commercial area's emergency management platform to the work apps of property management personnel, SMS terminals of merchants, and the area's public broadcasting system, achieving multi-channel coverage.

[0025] In this embodiment, after acquiring multi-source heterogeneous data of the target area, the method further includes: Preprocessing operations are performed on multi-source heterogeneous data. These operations include data cleaning, missing value imputation, outlier detection and handling, and normalization or standardization.

[0026] Specifically, after acquiring multi-source heterogeneous data from the target area, preprocessing operations are immediately carried out to ensure data quality. During data cleaning, for historical fire data, duplicate records are deleted, such as repeated entries of the same fire event in different databases, and data with incorrect formats, such as records with disordered date formats, are removed. For environmental parameter data, invalid values ​​caused by sensor malfunctions, such as negative temperature readings or extreme values ​​that exceed normal limits, are filtered out. When filling missing values, short-term missing data in environmental parameter data, such as missing humidity data for a certain 10 minutes, are filled using linear interpolation, that is, the missing value is calculated based on the linear trend of three valid data points before and after the missing period. Missing items in building information data, such as missing maintenance dates for fire protection facilities in some old buildings, are filled using the average maintenance dates for fire protection facilities of similar buildings in the area. Missing electricity load data is filled using the average electricity load for the same time period on adjacent dates. Outlier detection and handling employ the 3σ rule. First, the mean μ and standard deviation σ of each data sequence are calculated. Data exceeding the range of μ ± 3σ are identified as outliers. For abnormal peak values ​​in electrical load, such as power consumption significantly exceeding the building's normal maximum value during a certain period, the mean of the five normal data points before and after the outlier is used to replace the outlier. Normalization is performed using Z-score standardization, with the formula: ,in For standardized data, The original data is represented by μ, the mean of the data sequence is represented by σ, and the standard deviation of the data sequence is represented by σ. This processing ensures that all data are within a similar numerical range, avoiding interference with model training caused by differences in units such as temperature in degrees Celsius and electricity load in kilowatts.

[0027] In this embodiment, from multi-source heterogeneous data, time-series features at multiple preset time scales are extracted for each fire risk factor, including: Sliding window sampling is performed on multi-source heterogeneous data, where the length of each sliding window corresponds to a preset time scale; Statistical characteristics are calculated for the data within each sliding window. These characteristics include the mean, maximum, minimum, standard deviation, slope, kurtosis, autocorrelation coefficient, and Fourier transform spectral characteristics to comprehensively describe the dynamic changes within that time scale.

[0028] Specifically, when extracting time-series features at multiple preset time scales from multi-source heterogeneous data, the data is first sampled using a sliding window. For the hourly time scale, the sliding window length is set to 60 minutes with a step size of 10 minutes, meaning an hourly window containing 60 minutes of data is generated every 10 minutes to densely capture subtle changes in risk factors within a short period. For the daily time scale, the sliding window length is set to 24 hours with a step size of 1 hour, generating a daily window containing 24 hours of data every hour to reflect the cyclical changes in risk factors within a day. For the weekly time scale, the sliding window length is set to 168 hours with a step size of 6 hours, generating a weekly window containing 168 hours of data every 6 hours to uncover weekly trends. Statistical characteristics are calculated for the data within each sliding window. The mean is obtained by dividing the sum of all data points within the window by the number of data points, reflecting the overall level of the risk factor within that time scale. The maximum and minimum values ​​are extracted to represent the peak and trough values ​​of the data within the window, respectively, reflecting extreme cases. The standard deviation is obtained by calculating the square root of the sum of the squares of the differences between the data points within the window and the mean, reflecting the dispersion or volatility of the data. The slope measures the asymmetry of the data distribution, and the formula is: ; Where n is the number of data items in the window. For the i-th data point, 's' represents the average value of the data within the window, 's' represents the standard deviation, and a positive slope indicates a right-skewed data distribution, while a negative slope indicates a left-skewed distribution. Kurtosis is used to describe the steepness of a data distribution, and the formula is: ; A kurtosis greater than 3 indicates a steeper data distribution, while a kurtosis less than 3 indicates a flatter distribution. The autocorrelation coefficient is obtained by calculating the correlation between the data sequence within the window and its own k-step lagged sequence. The formula is as follows: , where k is the lag step number, used to capture the autocorrelation of the data; The Fourier transform spectral characteristics are obtained by performing a Fast Fourier Transform on the data within the window, as shown in the formula. , where F(k) is the spectral feature, f(n) is the time series data within the window, N is the number of data points, and j is the imaginary unit, used to identify the periodicity of the data.

[0029] In this embodiment, a multi-scale temporal fusion model is constructed, including: Feature encoding is performed on time-series features at different time scales to obtain feature vectors for each time scale. The feature vectors at their respective time scales are concatenated or weighted and fused, and then input into a neural network model for nonlinear transformation to generate a fused feature representation. The neural network model includes recurrent neural networks, temporal convolutional networks, or network structures based on attention mechanisms.

[0030] Specifically, when constructing a multi-scale temporal fusion model, the temporal features at different time scales are first encoded. For hourly temporal features, a Long Short-Term Memory (LSTM) network unit is used as the temporal encoder. The input dimension of this encoder is set to the number of extracted statistical features (e.g., 8 statistical features). The hidden layer dimension is set to 64. Through a gating mechanism, short-term dynamic change information in hourly features is selectively memorized, and a 64-dimensional hourly feature vector is output. For daily temporal features, a gated recurrent unit is used as the encoder. The input dimension is also the number of statistical features, and the hidden layer dimension is set to 64. This simplifies the structure of the LSM network to improve computational efficiency while effectively capturing the periodic changes of daily features, outputting a 64-dimensional daily feature vector. For weekly temporal features, a one-dimensional convolutional layer is used as the encoder. The kernel size is set to 3, and the number of kernels is set to 64. The trend information of weekly features is extracted through local convolution operations, outputting a 64-dimensional weekly feature vector. After encoding, the three 64-dimensional feature vectors are weighted and fused. The fusion weights are obtained adaptively through the model training process. For example, based on the contribution of features at each time scale to the fire prediction results, the final weights are determined as follows: hourly weight 0.4, daily weight 0.3, and weekly weight 0.3. The fusion formula is as follows: ,in To fuse feature vectors, These are hourly feature vectors. For daily-level feature vectors, The feature vectors are weekly-level features. Before fusion, each feature vector has been normalized using Min-Max to ensure consistent dimensions. The fused feature vector is then input into a temporal convolutional network for nonlinear transformation. This network contains three convolutional layers and two pooling layers. The convolutional layers use the ReLU activation function, and the pooling layers use max pooling. Through multiple nonlinear transformations, the deep correlations between features are further explored, ultimately generating a fused feature representation with a dimension of 128.

[0031] In this embodiment, feature encoding is performed on temporal features at different time scales, including: An independent temporal encoder is used to process the temporal feature sequence at each time scale. The temporal encoder includes a long short-term memory network unit, a gated recurrent unit, or a one-dimensional convolutional layer to extract and compress the sequence information at the corresponding time scale.

[0032] Specifically, when encoding time-series features at different time scales, an independent encoder is constructed using Long Short-Term Memory (LSTM) network units for hourly time-series feature sequences. The input to this encoder is a statistical feature sequence calculated using an hourly sliding window. The sequence length is set to 24, meaning it contains features from 24 consecutive hourly windows. The input dimension for each time step is 8, corresponding to 8 statistical features. The number of neurons in the encoder's hidden layer is set to 64. Through the synergistic effect of the input gate, forget gate, and output gate, key short-term dynamic information in the hourly sequence, such as sudden increases in electricity load and rapid temperature rises, is retained, while redundant noise data is forgotten. The final output is an hourly feature vector with a dimension of 64. This vector compresses the sequence information from 24 time steps, and normalization ensures dimensional uniformity. For daily time-series feature sequences, a gated recurrent unit (GRU) is used to construct the encoder. The sequence length is set to 7, which includes features from 7 consecutive daily windows. The input dimension for each time step is 8, and the number of hidden layer neurons is set to 64. The GRU simplifies the structure of the Long Short-Term Memory (LSTM) network by using update and reset gates, effectively capturing periodic changes in daily sequences such as daytime and nighttime environmental parameters and electricity load while reducing computational complexity. The output is a 64-dimensional daily feature vector. For weekly time-series feature sequences, a one-dimensional convolutional layer is used to construct the encoder. The sequence length is set to 4, which includes features from 4 consecutive weekly windows. The number of input channels is 8, corresponding to 8 statistical features. The kernel size is set to 3, the number of kernels is set to 64, and the convolution stride is set to 1. Local convolution operations are used to extract trend differences in risk factors between weekdays and weekends in the weekly sequence. Then, global average pooling is used to compress the convolution output into a 64-dimensional weekly feature vector.

[0033] In this embodiment, generating and outputting fire early warning information includes: Fire risk prediction value The fire risk level is determined by comparing it with multiple preset risk thresholds. Fire risk levels include at least three levels: low risk, medium risk, and high risk. Based on the determined fire risk level, a corresponding early warning signal or message is generated, which can be distributed via SMS, application push, email, or audible and visual alarms.

[0034] Specifically, when generating and outputting fire early warning information, the calculated fire risk prediction value is first... The risk was compared with three preset risk thresholds, which were set to 0.3 and 0.7 respectively. It was determined to be at a low risk level at that time. It was determined to be at a medium risk level at that time. When a high-risk level is triggered, a low-risk warning signal is generated. The warning message includes the current risk level, key risk factors such as slightly elevated electricity load, and recommended measures such as strengthening daily inspections of key areas. This warning message is automatically sent to the work email of property management personnel in the target area and simultaneously updated in the "Low Risk Alert" section of the regional emergency management APP. When a medium-risk level is triggered, a medium-risk warning message is generated. In addition to the risk level and analysis of key risk factors, it includes recommended emergency preparedness measures such as checking fire-fighting facilities and clearing evacuation routes. The warning message is pushed to the mobile phones of fire patrol personnel in the area via SMS, and simultaneously displayed as a pop-up on the homepage of the emergency management APP and on the electronic bulletin boards of buildings in the area. When a high-risk level is triggered, in addition to generating a warning message containing the risk level, detailed analysis of risk factors, and emergency response measures, the audible and visual alarm system is activated. Fire alarm bells and red warning lights in the target area are simultaneously activated. The warning message is sent simultaneously to the fire command center, regional property management, businesses, and residents via SMS, APP push, and email. The message sent to residents also includes the location information of the nearest emergency shelter.

[0035] In this embodiment, the method further includes: Based on actual fire event data, the multi-scale time series fusion model is trained and iteratively optimized. Training employs a supervised learning method, which minimizes the loss function between the predicted value and the actual label. The loss function can be either the cross-entropy loss function or the mean squared error loss function.

[0036] Specifically, when training and iteratively optimizing the multi-scale temporal fusion model, the first step is to collect multi-source heterogeneous data and corresponding actual fire event data from the target area over the past six years to construct a training dataset. Each sample in the dataset contains a feature vector of the multi-source heterogeneous data at a specific moment after preprocessing and temporal feature extraction, along with a label indicating whether a fire occurred within 24 hours of that moment. A fire occurrence is labeled as 1, and no fire occurrence as 0. The dataset is divided into training and testing sets in a 7:3 ratio. The training process employs a supervised learning method, with the optimization objective being to minimize the loss function between the predicted value and the actual label. The cross-entropy loss function is selected, and its formula is: ; Where L is the loss value and N is the number of training samples. Let i be the actual label of the i-th sample. The fire risk prediction value for the i-th sample is the model output. Training employed a stochastic gradient descent optimizer with a learning rate of 0.001 and a batch size of 32. After each training epoch, the model's prediction accuracy and F1 score were evaluated using a test set. If the evaluation metrics showed no improvement after five consecutive test set epochs, the learning rate was adjusted to 0.5 times its original value. Furthermore, the latest fire event data and multi-source heterogeneous data were collected quarterly to incrementally train the model and update its weights. and bias This ensures that the model can adapt to changes in the fire risk characteristics of the target area.

[0037] In this embodiment, the method further includes: Upon receiving a fire warning, the system intelligently recommends corresponding fire prevention strategies or emergency response measures based on the fire risk prediction value and fire risk level. Fire prevention strategies include resource allocation, personnel evacuation route planning, or enhanced monitoring of key areas.

[0038] Specifically, upon receiving a fire warning, the system intelligently recommends fire prevention strategies or emergency response measures based on the predicted fire risk value and corresponding risk level. When the warning is for low risk, considering current major risk factors such as slightly higher ambient temperature, the recommended fire prevention strategies include increasing the frequency of daily inspections of key flammable areas such as power distribution rooms and warehouses from twice a week to three times a week, while also reminding property management to check the water pressure of fire hydrants to ensure that basic fire protection facilities are in good condition. When a medium-risk warning is triggered, the recommended emergency response measures include resource allocation and personnel evacuation route planning. In terms of resource allocation, emergency supplies such as fire extinguishers and hoses from fire stations within a 3-kilometer radius are allocated to temporary storage points in the target area to ensure rapid deployment of supplies. In terms of personnel evacuation route planning, based on building distribution and road conditions, two main evacuation routes and one backup evacuation route are planned, avoiding dangerous areas such as gas stations and flammable material storage areas. Route maps are then pushed to residents and businesses in the area via the emergency management app. When the warning level is high risk, in addition to strengthening the monitoring of key areas such as adding temporary surveillance cameras in areas such as warehouses and power distribution rooms to transmit images to the fire command center in real time, firefighters are also arranged to be on duty on site, temporary checkpoints are set up at the main entrances and exits of the area, residents are reminded to reduce unnecessary outings, and fire prevention precautions and evacuation route guidance are broadcast every 30 minutes through the community broadcasting system.

[0039] Please see Figure 4 The present invention also provides a fire risk prediction and early warning system based on big data analysis, comprising: The data acquisition module is used to acquire multi-source heterogeneous data of the target area, including fire history data, environmental parameter data, building information data, and electricity load data. The time-series feature extraction module is used to extract time-series features at multiple preset time scales from multi-source heterogeneous data for each fire risk factor. The time-series feature extraction module includes a sliding window sampling unit and a statistical feature calculation unit. The model building and prediction module is used to build a multi-scale time series fusion model based on the time series characteristics under multiple preset time scales, and to use the multi-scale time series fusion model to calculate the fire risk prediction value of the target area. The early warning generation module is used to generate and output fire early warning information based on fire risk prediction values. The early warning generation module can intelligently recommend fire prevention strategies or emergency response measures according to the risk level.

[0040] Specifically, in practical applications, the fire risk prediction and early warning system based on big data analysis utilizes a data acquisition module that collects multi-source heterogeneous data through various interfaces. This module connects to the fire department's fire accident database via an API interface, automatically synchronizing the previous day's historical fire data every morning; it connects to environmental monitoring equipment deployed in the target area via the LoRa wireless communication protocol, receiving real-time environmental parameter data such as temperature, humidity, air pressure, and wind speed, with a data sampling frequency set every 5 minutes; it connects directly to the building information management system and property operation and maintenance platform of the housing and construction department via a database connection, obtaining updated data on building structure, material types, and fire protection facilities status at the beginning of each month; and it connects to the smart meter network of the power company via a dedicated power communication protocol, collecting electricity load data for each building every 15 minutes. The sliding window sampling unit in the time-series feature extraction module automatically generates corresponding sliding windows based on preset hourly, daily, and weekly time scales, sampling the preprocessed data collected by the data acquisition module; the statistical feature calculation unit incorporates algorithms for calculating statistical features such as average, maximum, and standard deviation, automatically calculating the corresponding statistical features for the data within each sliding window to generate a time-series feature sequence. The model building and prediction module loads the trained and optimized multi-scale temporal fusion model and inputs the multi-scale temporal features output from the temporal feature extraction module into the model. Through feature encoding, fusion, and nonlinear transformation, the model calculates the predicted fire risk value of the target area in real time. The early warning generation module receives the fire risk prediction value output by the model, compares it with the preset threshold to determine the risk level, and generates a corresponding early warning signal or message based on the risk level. It then calls the SMS gateway, APP push interface, email server, and audible and visual alarm control interface to distribute the early warning information to relevant personnel and equipment. At the same time, it calls the built-in strategy recommendation algorithm to generate corresponding fire prevention strategies or emergency response measures based on the risk level and current risk factors, and pushes them to relevant personnel such as firefighters and property management staff through the emergency management APP.

[0041] In this embodiment, the model building and prediction module further includes: The feature encoding unit is used to encode the temporal features at different time scales to obtain the feature vectors at their respective time scales. The fusion processing unit is used to concatenate or weightedly fuse feature vectors from their respective time scales. The prediction unit is used to input the fused features into the neural network model for nonlinear transformation in order to calculate the fire risk prediction value.

[0042] Specifically, the feature encoding unit in the model building and prediction module incorporates three independent encoder modules, corresponding to time-series feature encoding at hourly, daily, and weekly time scales, respectively. The hourly encoder module uses a long short-term memory network unit with built-in parameters including input dimension and hidden layer dimension. It can automatically adjust the calculation process according to the input hourly time-series feature sequence and output hourly feature vectors. The daily encoder module uses a gated recurrent unit with parameter settings matched to the hourly encoder to ensure that the output daily feature vector dimension is consistent with the hourly encoder. The weekly encoder module uses a one-dimensional convolutional layer with built-in convolutional kernels of different sizes for selection. It automatically selects an appropriate convolutional kernel size according to the length of the weekly time-series feature sequence and outputs weekly feature vectors with uniform dimensions. All three encoder modules include a normalization processing submodule to normalize the output feature vectors and ensure consistent dimensions. After receiving the feature vectors output by the three encoder modules, the fusion processing unit first performs Z-score standardization on each vector through the feature standardization submodule to eliminate dimensional differences. Then, the weighted fusion submodule performs weighted calculations on the standardized feature vectors based on the fusion weights obtained from model training to generate a fused feature vector. The fusion weights can be updated in real time based on the model's iterative optimization results. The prediction unit has a built-in temporal convolutional network model, which includes preset convolutional layers, pooling layers, and activation function parameters. After inputting the fused feature vectors output by the fusion processing unit into the model, it undergoes nonlinear transformation through multiple convolutional and pooling operations. Finally, it outputs the fire risk prediction value through a sigmoid activation function. The system also includes a built-in result verification submodule to check the reasonableness of the output prediction values, ensuring that the predicted values ​​are within a reasonable range of 0 to 1. Through the collaborative work of each subunit, the model building and prediction module can efficiently and accurately complete the processing of multi-scale time-series features and fire risk prediction, providing reliable prediction results for early warning generation.

[0043] In summary, this invention ensures data quality by acquiring and preprocessing multi-source heterogeneous data, including historical fire data, environmental parameter data, building information data, and electricity load data of the target area. It extracts time-series features at multiple preset time scales (hourly, daily, and weekly) for each fire risk factor and combines this with a multi-scale time-series fusion model to capture the evolution trend of fire risk factors over continuous time and their deep temporal correlation with fire occurrence. This effectively solves the problems of insufficient utilization of time-series features, insufficient accuracy in short-to-medium-term fire risk identification, and susceptibility to false alarms and missed alarms in existing technologies, thus improving the reliability and usability of early warning information. Furthermore, it provides tiered early warnings based on fire risk prediction values ​​and distributes early warning information through multiple channels. It can also intelligently recommend corresponding fire prevention strategies or emergency response measures, further improving the efficiency of fire prevention and emergency response, and providing scientific support for fire prevention in the target area.

[0044] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0045] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A fire risk prediction and early warning method based on big data analysis, characterized in that, include: S1. Acquire multi-source heterogeneous data of the target area. The multi-source heterogeneous data includes fire history data, environmental parameter data, building information data and power load data. Environmental parameter data includes temperature, humidity, air pressure and wind speed. Building information data includes building structure, material type and fire protection facility status. S2. From multi-source heterogeneous data, extract time-series features at multiple preset time scales for each fire risk factor. The multiple preset time scales include hourly, daily, and weekly levels. The time-series features represent the dynamic change pattern of the fire risk factor at the corresponding time scale. The dynamic change pattern includes trend, periodicity, volatility, or autocorrelation. S3. Based on the temporal characteristics of multiple preset time scales, a multi-scale temporal fusion model is constructed. The multi-scale temporal fusion model is used to capture the evolution trend of fire risk factors in continuous time and their deep temporal correlation with fire occurrence. S4. Calculate the predicted fire risk value of the target area using a multi-scale time-series fusion model. Fire risk prediction value This indicates the probability of a fire occurring or the fire risk level, where the fire risk prediction value is... The calculation formula is: ; in, Indicates from the first Feature vectors extracted at each time scale and processed by a multi-scale temporal fusion model and For the model's weights and biases, For activation function, This is a vector concatenation operation; S5. Generate and output fire early warning information based on fire risk prediction values.

2. The fire risk prediction and early warning method based on big data analysis according to claim 1, characterized in that, After acquiring the multi-source heterogeneous data of the target region, the method further includes: Preprocessing operations are performed on multi-source heterogeneous data. These operations include data cleaning, missing value imputation, outlier detection and handling, and normalization or standardization.

3. The fire risk prediction and early warning method based on big data analysis according to claim 1, characterized in that, The step involves extracting time-series features at multiple preset time scales from multi-source heterogeneous data for each fire risk factor, including: Sliding window sampling is performed on multi-source heterogeneous data, where the length of each sliding window corresponds to a preset time scale; Statistical characteristics are calculated for the data within each sliding window. These characteristics include the mean, maximum, minimum, standard deviation, slope, kurtosis, autocorrelation coefficient, and Fourier transform spectral characteristics to comprehensively describe the dynamic changes within that time scale.

4. The fire risk prediction and early warning method based on big data analysis according to claim 1, characterized in that, The construction of the multi-scale temporal fusion model includes: Feature encoding is performed on time-series features at different time scales to obtain feature vectors for each time scale. The feature vectors at their respective time scales are concatenated or weighted and fused, and then input into a neural network model for nonlinear transformation to generate a fused feature representation. The neural network model includes recurrent neural networks, temporal convolutional networks, or network structures based on attention mechanisms.

5. The fire risk prediction and early warning method based on big data analysis according to claim 4, characterized in that, The feature encoding of temporal features at different time scales includes: An independent temporal encoder is used to process the temporal feature sequence at each time scale. The temporal encoder includes a long short-term memory network unit, a gated recurrent unit, or a one-dimensional convolutional layer to extract and compress the sequence information at the corresponding time scale.

6. The fire risk prediction and early warning method based on big data analysis according to claim 1, characterized in that, The generation and output of fire early warning information includes: Fire risk prediction value The fire risk level is determined by comparing it with multiple preset risk thresholds. Fire risk levels include at least three levels: low risk, medium risk, and high risk. Based on the determined fire risk level, a corresponding early warning signal or message is generated, which can be distributed via SMS, application push, email, or audible and visual alarms.

7. The fire risk prediction and early warning method based on big data analysis according to claim 1, characterized in that, The method further includes: Based on actual fire event data, the multi-scale time series fusion model is trained and iteratively optimized. Training employs a supervised learning method, which minimizes the loss function between the predicted value and the actual label. The loss function can be either the cross-entropy loss function or the mean squared error loss function.

8. The fire risk prediction and early warning method based on big data analysis according to claim 1, characterized in that, The method further includes: Upon receiving a fire warning, the system intelligently recommends corresponding fire prevention strategies or emergency response measures based on the fire risk prediction value and fire risk level. Fire prevention strategies include resource allocation, personnel evacuation route planning, or enhanced monitoring of key areas.

9. A fire risk prediction and early warning system based on big data analysis, applied to the fire risk prediction and early warning method based on big data analysis as described in any one of claims 1-8, characterized in that, include: The data acquisition module is used to acquire multi-source heterogeneous data of the target area, including fire history data, environmental parameter data, building information data, and electricity load data. The time-series feature extraction module is used to extract time-series features at multiple preset time scales from multi-source heterogeneous data for each fire risk factor. The time-series feature extraction module includes a sliding window sampling unit and a statistical feature calculation unit. The model building and prediction module is used to build a multi-scale time series fusion model based on the time series characteristics under multiple preset time scales, and to use the multi-scale time series fusion model to calculate the fire risk prediction value of the target area. The early warning generation module is used to generate and output fire early warning information based on fire risk prediction values. The early warning generation module can intelligently recommend fire prevention strategies or emergency response measures according to the risk level.

10. The fire risk prediction and early warning system based on big data analysis according to claim 9, characterized in that, The model building and prediction module also includes: The feature encoding unit is used to encode the temporal features at different time scales to obtain the feature vectors at their respective time scales. The fusion processing unit is used to concatenate or weightedly fuse feature vectors from their respective time scales. The prediction unit is used to input the fused features into the neural network model for nonlinear transformation in order to calculate the fire risk prediction value.

Images