Fire-fighting data processing method and system based on big data

Abstract

The invention relates to the technical field of fire-fighting big data processing, and discloses a fire-fighting data processing method and system based on big data. The method comprises the steps of performing multi-scale space mapping on an original data stream collected by a fire-fighting sensor to form a three-dimensional data volume, and further analyzing an internal gradient of the three-dimensional data volume to construct a space flow field. Identifying flow field singular points in the field, constructing a singular point network, extracting key hubs, inputting the key hubs into a flow field generation model, and generating an extended flow field; and carrying out dynamic fusion and iterative evolution on the extended flow field and the original data volume to obtain a stable synthetic data field, finally separating an enhanced fluctuation component from the stable synthetic data field, recombining the enhanced fluctuation component with the high-frequency part of the original data, and outputting an enhanced result. According to the method, the field state structure and the evolution law of the data can be deeply analyzed, so that the accuracy of early recognition of the fire risk is improved, and the stability of a monitoring system under the condition that the data is incomplete or interfered is enhanced.

Description

Technical Field

[0001] This invention relates to the field of fire protection big data processing technology, specifically to a fire protection data processing method and system based on big data. Background Technology

[0002] Current fire monitoring technologies typically rely on widely deployed sensor networks to collect raw data such as temperature and smoke concentration. Mainstream processing methods focus on real-time threshold alarms for discrete data points, rule-based logical judgments, or simple spatiotemporal correlations of multi-source data. These solutions essentially treat data as independent observation sequences, identifying local anomalies through statistical or empirical models.

[0003] Existing technical solutions have shortcomings. Threshold and rule-based methods struggle to characterize the complex dynamic interactions and overall evolution processes among multiple physical quantities and nodes in a fire protection environment. They fail to understand the intrinsic structure and evolutionary dynamics of the data field at the system level, resulting in weak capabilities to address early hidden risks, chain reactions of multiple events, and data incompleteness. A technical solution is needed that can transform discrete data into a holistic dynamic field with inherent laws, and based on this, perform deep feature extraction and data reconstruction.

[0004] The core problem that this invention aims to solve is: how to break through the limitations of traditional discrete data processing modes and achieve in-depth analysis and enhancement of the overall field structure and evolution law of fire monitoring data, so as to improve the accuracy of risk identification and system robustness. Summary of the Invention

[0005] The purpose of this invention is to provide a fire data processing method and system based on big data to solve the problems mentioned in the background art.

[0006] To achieve the above objectives, the present invention provides a fire protection data processing method based on big data, the method comprising: Receive raw fire data streams from a cluster of sensors deployed in the fire monitoring area; The original fire data stream is subjected to multi-scale spatial mapping to form a three-dimensional data volume containing spatial nodes, event channels, and time scales; Analyze the data gradient distribution inside the three-dimensional data volume to construct a spatial flow field describing the direction and intensity of data flow; In the spatial flow field, flow field singularities with convergence or divergence characteristics are identified, and connection channels between the flow field singularities are established to form a singularity network. Extract the core singularities in the singularity network that satisfy the strength threshold and connectivity threshold, and use them as key data hubs; The key data hub is input into the trained flow field generation model, which generates an extended flow field through a deep neural network structure. The extended flow field and the basic components of the three-dimensional data volume are dynamically fused to form an enhanced data field; The enhanced data field is iteratively evolved to obtain a stable synthetic data field; The enhanced fluctuation component is separated from the synthetic data field, and the enhanced fluctuation component is recombined with the high-frequency component of the original fire data stream to output the enhanced fire monitoring data.

[0007] Preferably, the multi-scale spatial mapping of the original fire data stream includes: Define spatial scale parameters, which include macro-regional scale, meso-unit scale, and micro-viewpoint scale; For each spatial scale, extract the corresponding sensor readings, event labels, and time-stamped sequences from the raw fire data stream; The sensor readings at each spatial scale are mapped to spatial nodes, event tags are mapped to event channels, and time-scale sequences are mapped to time-scale scales. Spatial nodes, event channels, and time scales obtained by mapping at the macro-regional scale, meso-unit scale, and micro-viewpoint scale are spatially registered and hierarchically nested to form a three-dimensional data volume containing spatial nodes, event channels, and time scales.

[0008] Preferably, the step of analyzing the data gradient distribution within the three-dimensional data volume includes: Along the spatial node dimensions of the three-dimensional data volume, the rate of change of data values ​​between adjacent spatial nodes is calculated to obtain the spatial gradient; Along the event channel dimension of the three-dimensional data volume, calculate the rate of change of correlation strength between event channels to obtain the correlation gradient; The time gradient is obtained by calculating the rate of change of the data value over time along the time scale dimension of the three-dimensional data volume. The spatial gradient, correlated gradient, and temporal gradient are vector-synthesized to construct the gradient direction and gradient magnitude of each data point, forming a spatial flow field that describes the direction and intensity of data flow.

[0009] Preferably, identifying flow field singularities with converging or diverging characteristics in the spatial flow field includes: Calculate the gradient vector divergence at each location point in the spatial flow field; Points with negative gradient vector divergence and absolute values ​​exceeding a preset convergence threshold are identified and marked as flow field singularities with convergence characteristics. Points with positive gradient vector divergence exceeding a preset divergence threshold are identified and marked as flow field singularities with divergence characteristics. All flow field singularities with converging characteristics and those with diverging characteristics are collectively referred to as flow field singularities.

[0010] Preferably, establishing the connection channel between the flow field singularities includes: In the spatial flow field, streamline tracing is performed along the gradient direction starting from each flow field singularity with divergent characteristics; Tracing the streamlines until reaching a flow field singularity with converging characteristics, the streamline path between the flow field singularity with diverging characteristics and the flow field singularity with converging characteristics is marked as a connecting channel; Repeat the streamline tracing process starting from each flow field singularity with divergent characteristics until streamline tracing is completed for all flow field singularities with divergent characteristics. All the marked connection channels together with the flow field singularity form a singularity network.

[0011] Preferably, the extraction of core singularities in the singularity network that satisfy both the strength threshold and the connectivity threshold includes: Calculate the connection strength of each flow field singularity in the singularity network, whereby the connection strength is defined as the integral of the gradient magnitude of all connection channels connected to the flow field singularity; Calculate the network connectivity of each flow field singularity in the singularity network, where the network connectivity is defined as the number of other flow field singularities directly connected to the singularity. Set intensity and connectivity thresholds; Singularities in the flow field with connection strength greater than the strength threshold and network connectivity greater than the connectivity threshold are selected as key data hubs.

[0012] Preferably, inputting the key data hub into the trained flow field generation model includes: The flow field generation model employs a deep neural network with an encoder-decoder architecture; The encoder of the flow field generation model performs feature encoding on the key data hubs of the input to generate a latent spatial feature vector. The decoder of the flow field generation model decodes the potential spatial feature vector to generate an extended flow field that is associated with the key data hub in terms of spatial distribution and intensity characteristics. The spatial range covered by the extended flow field is greater than the spatial range of the original distribution of the key data hub; The construction steps of the flow field generation model include: Collect historical fire data streams and perform multi-scale spatial mapping on the historical fire data streams to form a historical three-dimensional data volume; Analyze the data gradient distribution within the historical three-dimensional data volume to construct the historical spatial flow field; In the historical spatial flow field, identify historical flow field singularities and establish connection channels between historical flow field singularities to form a historical singularity network; Extract historical key data hubs that satisfy the strength threshold and connectivity threshold from the historical singularity network; A deep neural network with an encoder-decoder architecture is constructed as a flow field generation model. The encoder consists of multiple convolutional and pooling layers, which are used to extract features and reduce dimensionality from the historical key data hubs of the input. The decoder consists of multiple deconvolutional and upsampling layers, which are used to reconstruct the extended flow field based on the feature vectors output by the encoder. Using the historical key data hub as training samples and the historical spatial flow field as the real label, the flow field generation model is trained under supervision. The model parameters are optimized by minimizing the difference between the generated flow field and the real flow field. During training, the gradient descent algorithm is used to iteratively update the model weights until the model converges, resulting in a trained flow field generation model.

[0013] Preferably, the dynamic fusion of the extended flow field and the fundamental components of the three-dimensional data volume includes: The three-dimensional data volume is decomposed into low-frequency trend components, mid-frequency mode components, and high-frequency detail components. The low-frequency trend component is used as the basic component of the three-dimensional data volume; In the spatiotemporal domain, the vector field of the extended flow field is superimposed point by point with the scalar field of the fundamental component; During the superposition process, the superposition weights are dynamically adjusted according to the gradient direction of the extended flow field at each point to generate an enhanced data field.

[0014] Preferably, separating the enhanced fluctuation component from the synthetic data field includes: The synthetic data field is decomposed to separate the slowly changing background field and the rapidly changing fluctuation field. The rapidly changing wave field is used as an enhanced wave component; The same field decomposition was performed on the original fire data stream to separate the original high-frequency components; The wave pattern of the enhanced wave component is aligned and fused with the detailed features of the original high-frequency component in the time and frequency domain to complete the reconstruction process and output the enhanced fire monitoring data.

[0015] Preferably, the present invention also includes a fire data processing system based on big data, the system including a memory, a processor, and a computer program stored in the memory and running on the processor, wherein when the processor executes the computer program, it implements the steps of the fire data processing method based on big data as described above.

[0016] Compared with the prior art, the beneficial effects of the present invention are: By mapping raw fire data streams to a three-dimensional data volume and analyzing its gradients to construct a spatial flow field, this method achieves continuous field-state characterization of discrete monitoring data. The distribution and changes of data in space are quantified as vectors with clear direction and intensity, thus characterizing the dynamic process of its overall flow and convergence. Identifying and connecting flow field singularities to form a network within this spatial flow field allows for direct location of event sources or core hubs of anomalous convergence and their associated paths from the topological structure of the data field. This shifts the basis of risk identification from parameter exceedances at single nodes to the detection of the overall structural characteristics and anomalous patterns of the data field.

[0017] By extracting key hubs from the singularity network to drive the flow field generation model, an extended flow field can be generated based on the learned intrinsic laws of the data field. The generated extended field is then fused with the original data volume and iteratively evolved until stable, ultimately recombining to output enhanced data. The model is used to simulate and autonomously improve the dynamics of the data field. When the original data is missing or disturbed, a more complete and smooth state representation can be reconstructed based on the regularity of the field. Simultaneously, the synthesized data field can reveal potential dynamic patterns not fully manifested in the original data, thereby improving the system's adaptability to complex environmental disturbances and its predictability of implicit evolutionary trends. Attached Figure Description

[0018] Figure 1 This is a schematic diagram illustrating the working principle of the fire data processing method based on big data described in this invention. Figure 2 A flowchart for multi-scale spatial mapping; Figure 3 A flowchart for establishing a connection channel for the flow field singularity; Figure 4 This refers to the distribution of the spatial flow field gradient direction and intensity vector during the flow field analysis stage. Figure 5 The curves show the changes in the dynamically superimposed weighting coefficients under different reference orientation angles. Detailed Implementation

[0019] 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.

[0020] Please see Figure 1 This invention provides a fire data processing method based on big data. The method includes: receiving a continuous stream of raw fire data from a sensor cluster deployed in a fire monitoring area, the data stream containing monitoring readings from various sensors; performing a multi-scale spatial mapping operation on the received raw fire data stream, which converts the raw data stream into a three-dimensional data volume containing spatial nodes, event channels, and time scales; analyzing the data gradient distribution within this three-dimensional data volume, constructing a spatial flow field that describes the direction and intensity of data flow by calculating gradients in different dimensions and synthesizing vectors; identifying flow field singularities with convergent or divergent characteristics in this spatial flow field, and establishing connection channels between these singularities based on streamline tracing, thereby forming a singularity network; extracting core singularities that simultaneously satisfy preset intensity and connectivity thresholds from this singularity network, these core singularities are defined as key data hubs; and inputting these key data hubs into a pre-trained flow field generation model, which employs a deep neural network structure and can generate a spatially extended flow field based on the input key data hubs. Subsequently, this extended flow field is dynamically fused with the base components of the previously obtained 3D data volume to generate an enhanced data field. This enhanced data field is then iteratively evolved until it reaches a stable state, resulting in a stable, evolved synthetic data field. The enhanced fluctuation component is separated from this synthetic data field and recombined with the high-frequency components of the original fire monitoring data stream, ultimately outputting the enhanced fire monitoring data.

[0021] Example 1: See Figure 2The system sets spatial scale parameters, including macro-regional scale, meso-unit scale, and micro-viewpoint scale. For each spatial scale, corresponding sensor readings, event tags, and time-scale sequences are extracted from the raw fire data stream. Sensor readings at each spatial scale are mapped to spatial nodes, event tags to event channels, and time-scale sequences to time-scale scales. The spatial nodes, event channels, and time-scale scales mapped at the macro-regional, meso-unit, and micro-viewpoint scales are spatially registered and hierarchically nested to form a three-dimensional data volume containing spatial nodes, event channels, and time-scale scales. Along the spatial node dimension of the three-dimensional data volume, the rate of change of data values ​​between adjacent spatial nodes is calculated to obtain the spatial gradient. Along the event channel dimension of the three-dimensional data volume, the rate of change of correlation strength between event channels is calculated to obtain the correlation gradient. Along the time-scale dimension of the three-dimensional data volume, the rate of change of data values ​​over time is calculated to obtain the temporal gradient. The calculated spatial gradient, correlation gradient, and temporal gradient are vector-synthesized to construct the gradient direction and gradient magnitude of each data point, forming a spatial flow field describing the direction and intensity of data flow.

[0022] In practical implementation, for a fire monitoring scenario in an industrial park covering ten square kilometers, the deployed sensor cluster includes 500 temperature sensors, 300 smoke sensors, and 100 flame sensors. Raw fire data streams are continuously input, containing the readings of each sensor every second, device status codes, and timestamps accurate to milliseconds. Spatial scale parameters are defined: the macro-regional scale corresponds to the geographical boundary of the entire industrial park; the meso-unit scale corresponds to the building outlines of twenty independent factories and warehouses within the park; and the micro-point location scale corresponds to the specific physical installation coordinates of each sensor. For the macro-regional scale, the average readings of all sensors are extracted from the raw fire data stream as the sensor reading at that scale, the "overall risk level of the park" is extracted as the event label, and a time-stamped sequence with hourly intervals is extracted. For the meso-unit scale, for each factory building, the set of all sensor readings located within that factory building is extracted from the raw fire data stream as the sensor reading for that unit, specific events such as "high temperature warning for factory A" and "smoke detection in warehouse B" are extracted as event labels, and a time-stamped sequence with minute intervals is extracted. For micro-level positioning, for each sensor installation point, the continuous reading of the sensor itself is extracted from the original fire data stream as the sensor reading, the event code triggered by the sensor is extracted as the event tag, and the original time stamp sequence with a second interval is extracted.

[0023] In practical implementation, the average sensor readings at the macro-regional scale are mapped to a spatial node representing the entire park, with the node value being the average temperature. The event label "Overall Risk Level of the Park" is mapped to an event channel connecting this macro-node with a virtual node in the higher-level monitoring center. The time-scaled sequence with hourly intervals is mapped to a macro-time scale. At the meso-unit scale, the set of sensor readings for each factory building is mapped to a spatial node representing that factory building, with the node value being the statistical characteristic value of the sensor readings within that factory building, such as the highest temperature. Specific event labels such as "High Temperature Warning for Factory A" are mapped to event channels connecting that factory building node with adjacent factory building nodes or macro-park nodes. The time-scaled sequence with minute intervals is mapped to a meso-time scale. At the micro-viewpoint scale, the readings of each sensor are mapped to a spatial node in precise coordinates. The sensor event code is mapped to an event channel connecting that sensor node with its corresponding factory building node. The raw time-scaled sequence with second intervals is mapped to a micro-time scale. All spatial nodes, event channels, and time scales obtained by mapping at the above three scales are spatially registered so that the coordinates of micro nodes are nested within the outline of the meso-level factory buildings, the outline of the meso-level factory buildings is nested within the boundary of the macro-level park, and the time scales are aligned hierarchically, ultimately forming a three-dimensional data volume containing multiple layers of spatial nodes, multi-dimensional event channels, and multi-level time scales.

[0024] In some embodiments, along the spatial node dimension of the three-dimensional data volume, the rate of change of data values ​​between adjacent spatial nodes is calculated to obtain the spatial gradient. Taking temperature data as an example, the temperature difference between the spatial nodes of factory A and the adjacent spatial nodes of factory B is calculated, and combined with the actual distance between the nodes, the spatial gradient vector and magnitude from the high-temperature node to the low-temperature node are obtained. Along the event channel dimension of the three-dimensional data volume, the rate of change of the correlation strength between event channels is analyzed to obtain the correlation gradient. For example, the probability change rate of the co-occurrence of the "high temperature warning of factory A" event channel and the "smoke detection of warehouse B" event channel within a specific time period is calculated to obtain the correlation gradient characterizing the direction of the evolution of event correlation. Along the time scale dimension of the three-dimensional data volume, the rate of change of data values ​​of the same spatial node over time is calculated to obtain the temporal gradient, such as the rate of increase or decrease of the temperature of a sensor node on a continuous time scale. In practice, spatial gradient, correlation gradient and temporal gradient are vector synthesized. The spatial gradient vector indicates the direction of diffusion in space, the correlation gradient vector indicates the direction of influence of event logic, and the temporal gradient vector indicates the direction of evolution in time. The three are vectored together to construct the gradient direction and gradient magnitude of each data point in the synthesized vector field. This synthesized vector field is the spatial flow field that describes the direction and intensity of data flow.

[0025] In some embodiments, the gradient vector composition operation can be expressed in the following form: in: Indicates spatial location and time The synthesized gradient vector. This represents the spatial gradient vector, which is calculated from the numerical difference between adjacent spatial nodes and the distance quotient. This represents the associated gradient vector, which is calculated from the change in the correlation strength between event channels over time. This represents the time gradient vector, calculated from the rate of change of data at the same node over time. , , These are the synthesis weight coefficients for the corresponding gradients, used to adjust the contribution of gradients in different dimensions during synthesis. Through the above calculations, each data point obtains a gradient vector with a clear direction and magnitude, and the gradient vectors of all data points together form the spatial flow field.

[0026] Optional, composite weighting coefficients , , The value of the gradient vector can be set based on prior knowledge of the monitoring scenario. In fire risk monitoring scenarios, the temporal gradient vector may be assigned a higher weight because a rapid rise in temperature is a more direct signal in the early stages of a fire. In correlation analysis scenarios, the correlation gradient vector may be assigned a higher weight to highlight the transmission relationship of the event chain. It can be understood that the calculation of spatial gradients depends on the spatial topological relationship of sensor nodes, the calculation of correlation gradients depends on the logical relationship defined by event labels, and the calculation of temporal gradients depends on a unified time-scaled sequence. The three-dimensional data volume provides a unified data foundation for these calculations, enabling gradients of different properties to be calculated and synthesized within the same mathematical framework.

[0027] Example 2: See Figure 3 The process involves calculating the gradient vector divergence at each point in the spatial flow field. Points with negative gradient vector divergence exceeding a preset convergence threshold are identified and marked as flow field singularities with convergence characteristics. Points with positive gradient vector divergence exceeding a preset divergence threshold are marked as flow field singularities with divergence characteristics. All marked flow field singularities with convergence and divergence characteristics are collectively referred to as flow field singularities. In the spatial flow field, starting from each flow field singularity with divergence characteristics, streamlines are traced along the gradient direction of that point. The streamlines are traced until a flow field singularity with convergence characteristics is reached, and the streamline path between this flow field singularity with divergence characteristics and this flow field singularity with convergence characteristics is marked as a connecting channel. The streamline tracing process starting from each flow field singularity with divergence characteristics is repeated until streamline tracing is completed for all flow field singularities with divergence characteristics. All marked connecting channels and flow field singularities together constitute a singularity network.

[0028] In practical implementation, based on the spatial flow field constructed in the aforementioned industrial park scenario, the gradient vector divergence at each discrete grid point or spatial node location is calculated for this flow field. The spatial flow field is a vector field defined on a three-dimensional data volume coordinate system, where each point has a direction and intensity represented by a composite gradient vector. Calculating the gradient vector divergence is a fundamental mathematical operation for identifying the "sources" and "sinks" of mass or information in the flow field. In practical implementation, the gradient vector divergence is calculated by measuring the net flux of the spatial flow field vector within a local, small region. A positive value indicates that the vector line diverges from that point, while a negative value indicates that the vector line converges towards that point. For a region dominated by both temperature and smoke concentration gradients, at a continuously heating equipment location inside the factory, the temperature gradient vector exhibits a radiating pattern from the center outwards, resulting in a positive gradient vector divergence. At a poorly ventilated raw material accumulation point in a corner of the factory, heat and smoke tend to accumulate rather than disperse, resulting in a negative gradient vector divergence.

[0029] In some embodiments, the preset convergence threshold and divergence threshold are two defined scalar values. Points with negative gradient vector divergence and absolute values ​​exceeding the preset convergence threshold are identified and marked as flow field singularities with convergence characteristics. For example, at a raw material accumulation point, the calculated gradient vector divergence is -1.2, and the preset convergence threshold is 0.8. Since the absolute value 1.2 is greater than 0.8, this point is marked as a flow field singularity with convergence characteristics. Points with positive gradient vector divergence and exceeding the preset divergence threshold are identified and marked as flow field singularities with divergence characteristics. For example, at a continuously heating equipment location, the calculated gradient vector divergence is +1.5, and the preset divergence threshold is 0.6. Since 1.5 is greater than 0.6, this point is marked as a flow field singularity with divergence characteristics. All marked flow field singularities with convergence and divergence characteristics are collectively referred to as flow field singularities. These singularities constitute special behavioral foci in the spatial flow field.

[0030] In some embodiments, gradient vector divergence The calculation formula is expressed as: in: Indicates position and time The spatial flow field vector at that location. , , These represent the components of the spatial flow field vector in the x-direction (space node dimension), y-direction (space node dimension), and t-direction (time scale dimension), respectively. This represents the rate of change of the x-component along the x-direction. This represents the rate of change of the y-component along the y-direction. This represents the rate of change of the component along the t-direction. Gradient vector divergence. The value of is the sum of the rates of change in these three directions. When the sum is significantly positive at a point, that point is a source of divergence; when it is significantly negative, that point is a sink of convergence.

[0031] Optionally, the specific values ​​of the preset convergence threshold and preset divergence threshold can be calibrated based on historical fire event data. Through statistical analysis of the spatial flow field divergence distribution under historical normal and abnormal event conditions, threshold levels that can effectively distinguish background noise from significant "source" and "sink" characteristics are determined. During continuous monitoring, these thresholds can also be dynamically adjusted based on the overall statistical characteristics of the real-time flow field. In specific implementation, after identifying all flow field singularities, streamline tracing is performed along the gradient direction from each flow field singularity with divergent characteristics in the spatial flow field. Streamline tracing is a numerical integration method; starting from a divergent singularity, it moves a small step along the gradient vector direction of that point to reach the next point, and then continues moving along the gradient vector direction of the new point, iterating in this way. The streamlines are traced until they reach a flow field singularity with converging characteristics. A streamline originating from an industrial high-temperature exhaust vent (diverging singularity) may follow a hot air diffusion path and eventually terminate in a poorly ventilated warehouse corner (converging singularity). The streamline path between the diverging and converging flow field singularities is marked as a connecting channel. The streamline tracing process is repeated from each flow field singularity with diverging characteristics. In a factory with multiple heat-generating devices, streamlines originating from multiple device points (diverging singularities) may converge to the same high-temperature risk area (converging singularity) or point to different smoke accumulation points, until streamline tracing is completed for all flow field singularities with diverging characteristics. All marked connecting channels, together with all previously identified flow field singularities, constitute a singularity network. This network describes the directional connectivity from the main "source" points to the main "sink" points in the spatial flow field.

[0032] It is understandable that the accuracy of streamline tracing is affected by the step size. An excessively large step size can cause streamlines to deviate from the actual path, while an excessively small step size increases the computational burden. The tracing process may encounter situations where known convergence singularities cannot be reached, such as when streamlines extend to the boundary of the monitoring area. Such paths can be marked as open channels or terminated according to preset rules. All successfully established connecting channels and flow field singularities together form a singularity network, which, in the form of a graph, depicts the core transmission framework of fire protection data in the spatial flow field. Optionally, the termination condition for streamline tracing is not limited to reaching a convergence singularity. The tracing process can also terminate when the streamline path length exceeds a preset maximum length, or when the streamline enters a "stagnant" region where the gradient vector magnitude is below a specific threshold. This incomplete path can be selected to be included in the connecting channels as needed.

[0033] Example 3: Calculate the connectivity strength of each flow field singularity in the singularity network. Connectivity strength is defined as the integral of the gradient magnitude of all connection channels connected to that singularity. Calculate the network connectivity of each flow field singularity in the singularity network. Network connectivity is defined as the number of other flow field singularities directly connected to that singularity. Set strength thresholds and connectivity thresholds. Select flow field singularities with connectivity strength greater than the strength threshold and network connectivity greater than the connectivity threshold as key data hubs.

[0034] In practical implementation, based on the singularity network already established in the aforementioned industrial park scenario, the singularity network consists of multiple flow field singularities and connecting channels linking these singularities. The connection strength of each flow field singularity in the singularity network is calculated, defined as the integral of the gradient magnitude of all connecting channels connected to that singularity. Specifically, for any flow field singularity, all connecting channels in the singularity network that originate or terminate at that singularity are found. For each connecting channel, the spatial flow field gradient magnitude distribution along the channel path is obtained. The gradient magnitude represents the intensity of data flow. The gradient magnitude is integrated along the connecting channel path, and the integration result represents the "flux" contribution of that connecting channel. The scalar value obtained by summing the integrals of all connecting channels associated with a flow field singularity is the connection strength of that singularity. For example, a flow field singularity located in a raw material warehouse is connected to three connecting channels from different heating devices. The integrals of the gradient magnitudes of the three channels are calculated separately and then summed, resulting in a connection strength of 85.6 for the warehouse singularity.

[0035] In some embodiments, the integral calculation of connection strength can be expressed as the following formula: in: Represents the singularity of the flow field The connection strength. Represents all singularities in the singularity network that are related to the flow field singularity. A collection of directly connected channels. Represents a set A specific connection channel within it. Indicates connection channel Spatial path. Indicates the connection channel path At a certain point The spatial flow field gradient vector at that location. Indicates at point The magnitude, or intensity, of the gradient vector at that point. Indicates along the path Integral infinitesimal. This indicates that the gradient magnitude is along the connecting channel. Perform a line integral along the entire path; the result represents the channel. The intensity contribution. Indicates that belonging to the set The integral results of all connected channels are summed to finally obtain the flow field singularity. Total connection strength .

[0036] In practical implementation, the network connectivity of each flow field singularity in the singularity network is calculated. Network connectivity is defined as the number of other flow field singularities directly connected to the singularity. Direct connection means that there is a connecting channel between two flow field singularities, regardless of whether the direction of the connecting channel is from one flow field singularity to another. For example, a flow field singularity acting as a convergence point receives connecting channels from three different divergence point flow field singularities; the number of other flow field singularities directly connected to this convergence point is three, and its network connectivity is three. A flow field singularity acting as a divergence point may emit connecting channels pointing to one or more convergence points; its network connectivity is equal to the number of convergence points it directly connects to. From the perspective of an undirected graph, only the number of adjacent singularities is counted without considering the connection direction; a singularity simultaneously connected to four other singularities has a network connectivity of four.

[0037] In practical implementation, a strength threshold and a connectivity threshold are set. The strength threshold is a threshold for connection strength. The strength threshold is a scalar lower bound, while the connectivity threshold is an integer lower bound for network connectivity. The strength threshold filters out flow field singularities that carry high "flux" in the flow field, while the connectivity threshold filters out flow field singularities that are in highly connected positions within the network. The strength threshold can be determined based on the statistical percentile of the historical singularity network connectivity distribution, while the connectivity threshold can be determined based on network topology characteristics. In one configuration, the strength threshold is set to 50.0, and the connectivity threshold is set to 2.

[0038] In practice, flow field singularities with connection strength and network connectivity exceeding both thresholds are selected as key data hubs. Each flow field singularity in the singularity network is traversed, and the connection strength of each singularity is... The network connectivity of each flow field singularity is compared with a connectivity threshold, compared to a strength threshold. A flow field singularity is only identified as a critical data hub if both its connectivity strength and network connectivity are greater than the strength threshold and the connectivity threshold, respectively. For example, flow field singularity A has a connectivity strength of 85.6, a network connectivity of 3, a strength threshold of 50.0, and a connectivity threshold of 2; therefore, flow field singularity A meets the criteria and is extracted as a critical data hub. Flow field singularity B has a connectivity strength of 40.2 and a network connectivity of 4; its connectivity strength does not exceed the strength threshold, so flow field singularity B is not considered a critical data hub. Flow field singularity C has a connectivity strength of 60.1 and a network connectivity of 1; its network connectivity does not exceed the connectivity threshold, so flow field singularity C is also not considered a critical data hub. All the extracted critical data hubs constitute the core set for subsequent steps.

[0039] In some embodiments, the intensity threshold and connectivity threshold can be set independently or in conjunction. One option is to use dynamic thresholds, where the intensity threshold is a multiple of the median connection intensity of all flow field singularities in the current singularity network, and the connectivity threshold is the average connectivity of all flow field singularity networks rounded up. Another option is to use fixed rules, where a set of empirical threshold parameters is pre-calibrated by analyzing the network state during historical fire events in a specific application scenario.

[0040] It is understandable that connectivity strength reflects the importance of flow field singularities in terms of data "flow," while network connectivity reflects their importance in terms of topological structure. The selection criteria, which simultaneously satisfy both thresholds, aim to identify key locations that are both high-flow hubs and high-connectivity nodes. These locations may correspond to core risk sources or critical transmission nodes in fire monitoring scenarios. The selection process involves deterministic calculations and comparisons, without relying on probabilistic judgments. The extraction of key data hubs simplifies the complex singularity network into a set of a few most important nodes. This set retains the core dynamic characteristics and topological framework of the original network, providing focused, high-quality input for subsequent flow field generation and data augmentation. The entire extraction process is based entirely on the quantitative attributes of flow field singularities and preset deterministic thresholds, ensuring repeatability and objectivity.

[0041] Example 4: The flow field generation model employs a deep neural network with an encoder-decoder architecture. The encoder of the flow field generation model encodes the features of the input key data hubs, generating latent spatial feature vectors. The decoder decodes these latent spatial feature vectors, generating an extended flow field that is associated with the key data hubs in terms of spatial distribution and intensity characteristics. This extended flow field covers a spatial range larger than the original distribution range of the key data hubs. The construction steps of the flow field generation model include: collecting historical fire-fighting data streams and performing multi-scale spatial mapping on these streams to form a historical three-dimensional data volume; analyzing the data gradient distribution within the historical three-dimensional data volume to construct a historical spatial flow field; identifying historical flow field singularities within the historical spatial flow field and establishing connection channels between these singularities to form a historical singularity network; and extracting historical key data hubs that satisfy intensity and connectivity thresholds from the historical singularity network. A deep neural network with an encoder-decoder architecture is constructed as the flow field generation model. The encoder consists of multiple convolutional and pooling layers for feature extraction and dimensionality reduction of the input historical key data hubs. The decoder consists of multiple deconvolutional and upsampling layers for reconstructing the extended flow field based on the feature vectors output by the encoder. Historical key data hubs are used as training samples, and historical spatial flow fields are used as ground truth labels to supervise the training of the flow field generation model. The model parameters are optimized by minimizing the difference between the generated flow field and the real flow field. During training, gradient descent is used to iteratively update the model weights until the model converges, resulting in the trained flow field generation model.

[0042] In practical implementation, the flow field generation model employs a deep neural network with an encoder-decoder architecture. The encoder in this architecture encodes the features of the input key data hubs, generating a latent spatial feature vector. The key data hubs are input as spatial coordinates and associated attribute vectors; for example, a key data hub can be represented as (x-coordinate, y-coordinate, connection strength, network connectivity, average gradient direction). A set of key data hubs constitutes a sparse spatial point set. The decoder in the encoder-decoder architecture decodes the latent spatial feature vector, generating an extended flow field that is associated with the key data hubs in terms of spatial distribution and intensity characteristics. The generated extended flow field covers a spatial range larger than the original distribution range of the key data hubs. For example, the input key data hubs may only be distributed in a portion of an industrial park, while the generated extended flow field can cover the entire monitoring grid of the industrial park.

[0043] In practical implementation, the construction steps of the flow field generation model include collecting historical fire-fighting data streams and performing multi-scale spatial mapping on these data streams to form a historical three-dimensional data volume. The data gradient distribution within the historical three-dimensional data volume is analyzed to construct the historical spatial flow field. Historical flow field singularities are identified within the historical spatial flow field, and connection channels between these singularities are established to form a historical singularity network. Historical key data hubs that satisfy intensity and connectivity thresholds are extracted from the historical singularity network. A deep neural network with an encoder-decoder architecture is constructed as the flow field generation model. The encoder consists of multiple convolutional and pooling layers for feature extraction and dimensionality reduction of the input historical key data hubs, while the decoder consists of multiple deconvolutional and upsampling layers for reconstructing the extended flow field based on the feature vectors output by the encoder.

[0044] In some embodiments, the collected historical fire data streams cover normal operating conditions and various typical fire risk scenarios, such as localized overheating, smoldering, and open flame. The aforementioned mapping, parsing, identification, and extraction processes are independently executed for each historical fire data stream to obtain multiple sets of historical key data hubs and their corresponding historical spatial flow fields. A set of historical key data hubs serves as a training input sample for the model, and its corresponding historical spatial flow field covering the entire monitoring area serves as the true label for that sample. See Table 1, which shows a set of historical key data hubs: Table 1: Historical Key Data Hub Coordinates and Attributes

[0045] In practice, historical key data hubs are used as training samples, and historical spatial flow fields are used as real labels to supervise the training of the flow field generation model. The training objective is to minimize the difference between the extended flow field generated by the model and the real historical spatial flow field. The model parameters are optimized by minimizing the difference between the generated flow field and the real flow field. The difference is usually quantified using a loss function; a commonly used loss function is a weighted sum of mean squared error and gradient difference loss.

[0046] In some embodiments, the loss function for supervised training It can be represented as: in: This represents the total training loss. This represents the total number of spatial flow field grid points in a training batch. The flow field generation model is represented in the first... Vector predictions of the extended flow field generated at each grid point. Indicates the corresponding first The true vector value of the historical spatial flow field at each grid point. The generated extended flow field is represented in the first... The predicted gradient vector at each grid point. The historical spatial flow field is represented in the first... The true gradient vector at each grid point. and It is a hyperparameter that balances the weights of the two loss terms, and its value is a positive real number. This represents the square of the L2 norm of the calculated vector. The first term is the mean squared error loss, which helps the generated flow field to numerically approximate the real flow field. The second term is the gradient difference loss, which helps the generated flow field to structurally approximate the real flow field.

[0047] In practice, the gradient descent algorithm is used to iteratively update the model weights during training. The gradient descent algorithm is based on the loss function. The gradient directions of all model parameters are used to adjust the parameter values ​​in the reverse direction with a set learning rate. Each complete traversal of the training dataset is called a training epoch, and the training process continues for multiple epochs until the loss function is satisfied. When the value of d drops to a stable low value region and no longer decreases significantly, or reaches the preset maximum number of training cycles, the model is considered to have converged, and a well-trained flow field generation model is obtained.

[0048] Optionally, the specific structure of the deep neural network in the encoder-decoder architecture can be adjusted. The encoder part can contain three convolutional layers, each followed by a pooling layer, progressively downsampling the set of key data hubs in the input into a compact latent spatial feature vector. The decoder part can contain three deconvolutional layers, each preceded by an upsampling layer, progressively upsampling the latent spatial feature vector and reconstructing it into an extended flow field consistent with the size of the target spatial flow field. Skip connections can be used between the encoder and decoder to fuse features at different scales. Optional hyperparameters during training, such as learning rate, batch size, and weights in the loss function, can also be adjusted. and This can be determined by performing a grid search on an independent validation dataset. When applied to new, unseen key data hubs, the trained flow field generation model can generate a reasonable, broader-coverage extended flow field prediction based on their spatial distribution and attribute characteristics.

[0049] It's understandable that a well-trained flow field generation model essentially learns the complex mapping relationship between the sparse distribution of key data hubs and the complete spatial flow field. This mapping relationship encompasses the physical laws and statistical patterns of the spatial propagation and evolution of fire data, enabling the model to infer the flow field situation of the entire monitoring area with only key node information. It's also understandable that the construction and training of the flow field generation model is an offline preparation process. Once the model is trained, it can be deployed online, receiving real-time extracted key data hubs and quickly generating extended flow fields, providing input for subsequent dynamic fusion and data augmentation. The model's performance depends on the quality of historical data, the scope of covered scenarios, and the appropriateness of the model architecture and training strategy.

[0050] See Figure 4 In the flow field analysis stage of fire data processing, this figure visually presents the spatial pattern of data flow within the monitoring area: the direction of the vector arrows in the figure corresponds to the gradient direction of the spatial flow field, reflecting the flow trend of data in the XY coordinate plane; the arrow colors correspond to the two color scales on the right, where red represents the gradient magnitude (gradient intensity value) and blue represents the flow field intensity value, with darker colors indicating higher intensity. The "+" symbols marked in the figure represent key data hubs, whose spatial distribution is highly correlated with the convergence / divergence characteristics of the flow field. Key data hubs are mostly located in the turning areas of the flow field gradient direction or in areas of high intensity, which conforms to the processing logic of "extracting core singularities from the singularity network as key data hubs". The spatial scale of this figure is 500m × 500m, covering the complete fire monitoring area grid, and its flow field distribution reflects the propagation law of fire data in the spatial dimension, providing a basic reference for the subsequent generation of extended flow fields.

[0051] Example 5: Component decomposition is performed on the 3D data volume to obtain low-frequency trend components, mid-frequency modal components, and high-frequency detail components. The low-frequency trend components are used as the fundamental components of the 3D data volume. In the spatiotemporal domain, the vector field of the extended flow field is superimposed point-by-point with the scalar field of the fundamental components. During the superposition process, the superposition weight is dynamically adjusted according to the gradient direction of the extended flow field at each point to generate an enhanced data field. Field decomposition is performed on the synthesized data field to separate the slowly changing background field and the rapidly changing fluctuation field. The rapidly changing fluctuation field is used as the enhanced fluctuation component. The same field decomposition is performed on the original fire monitoring data stream to separate the original high-frequency components. The fluctuation patterns of the enhanced fluctuation components are aligned and fused with the detail features of the original high-frequency components in the time and frequency domain to complete the reconstruction process and output the enhanced fire monitoring data.

[0052] In practice, the three-dimensional data volume is decomposed into components, yielding low-frequency trend components, mid-frequency modal components, and high-frequency detail components. The low-frequency trend components capture the slowly changing background or tone of the data in the spatiotemporal dimension; the mid-frequency modal components reflect fluctuation patterns with a certain periodicity or regularity; and the high-frequency detail components contain rapid, localized fluctuations and noise. The decomposed low-frequency trend components are used as the foundational components of the three-dimensional data volume. These foundational components represent the most stable and core evolutionary trends in the fire protection data field, such as the diurnal slow-changing baseline of the overall temperature field in an industrial park.

[0053] In practice, in the spatiotemporal domain, the vector field of the extended flow field is superimposed point-by-point with the scalar field of the fundamental component. The extended flow field is a vector field containing both direction and intensity information, while the fundamental component is a scalar field containing only intensity information. The superposition operation is performed at each spatiotemporal grid point, algebraically adding the vector value of the extended flow field at that point to the scalar value of the fundamental component at that point. The scalar value of the fundamental component is considered a vector in the zero direction. During the superposition process, the superposition weights are dynamically adjusted according to the gradient direction of the extended flow field at each point to generate an enhanced data field. The gradient direction of the extended flow field at a point indicates the main trend of data flow at that point, and the superposition weights are used to adjust the intensity of the final contribution of the extended flow field to that point.

[0054] In some embodiments, the adjustment of the dynamically superimposed weights can be based on the following relationship: in: Represents in spacetime coordinates The dynamic superposition weighting coefficient at the location has a value between 0 and 1. Represents in spacetime coordinates The gradient direction angle of the extended flow field is expressed in radians. This represents a preset reference direction angle used to indicate the dominant direction of data flow that is to be enhanced, such as in fire risk analysis. It can be set to point in the direction of the nearest evacuation exit or main ventilation path. Calculate the angle between the gradient direction of the extended flow field and the reference direction. The cosine value is 1 when the two directions are in the same direction, and the weight is... The cosine value reaches a maximum of 1; when the two directions are opposite, the cosine value is -1, and the weight... The value is 0. This weighting coefficient is then used to scale the extended flow field vector, which is then superimposed on the fundamental components, ultimately resulting in a value at point... The generated enhanced data field vector It can be represented as ,in These are the basic component scalar values. It is the unit vector in the zero direction. It is the extended flow field vector.

[0055] In practice, the synthetic data field is decomposed to separate the slowly changing background field and the rapidly changing fluctuation field. This field decomposition method is mathematically consistent with the aforementioned component decomposition method for 3D data volumes, but the application is to the stable synthetic data field obtained after iterative evolution. The rapidly changing fluctuation field obtained from the decomposition is used as the enhanced fluctuation component, which inherits the rapidly changing characteristics of the synthetic data field after model inference and iterative evolution. The original fire data stream is subjected to the same field decomposition to separate the original high-frequency components. Here, "the same field decomposition" refers to using the exact same decomposition algorithm and parameter settings as when processing the synthetic data field, directly applying it to the original fire data stream, which has not undergone the aforementioned multi-step processing, to extract its high-frequency components. The original high-frequency components contain the sensor's original rapid response details and noise.

[0056] In practice, the wave pattern of the enhanced wave component is aligned and fused with the detailed features of the original high-frequency component in the time and frequency domains to complete the reconstruction process and output enhanced fire monitoring data. Time-frequency domain alignment ensures that the enhanced wave component and the original high-frequency component correspond completely in terms of timestamps and spatial locations. The fusion operation is not a simple addition, but rather a selective combination of the signal energy from the enhanced wave component and the signal energy from the original high-frequency component within the time and frequency domain, for example, in the time-frequency representation obtained through wavelet transform or short-time Fourier transform.

[0057] In some embodiments, one fusion method involves comparing the amplitude spectra of the enhanced wave component and the original high-frequency component within each time-frequency unit, retaining the value of the component with the larger amplitude in that unit. Another fusion method involves linearly weighting the two components in the time-frequency domain, with the weights dynamically allocated based on frequency or the estimated local signal-to-noise ratio. The fused time-frequency representation is then reconstructed back into the spatiotemporal domain signal through an inverse transform. This reconstructed signal is the final enhanced fire monitoring data, which includes both the enhanced wave pattern obtained from the synthetic data field and retains the high-frequency detail features of the original data.

[0058] Optionally, the same time-series decomposition techniques can be used for component decomposition of the 3D data volume and field decomposition of the synthesized data field, such as empirical mode decomposition, variational mode decomposition, or wavelet multiresolution analysis. These techniques can adaptively decompose a complex signal into several eigenmode components from low to high frequencies. The lowest frequency component can be used as the background field or trend component, and the combination of several high-frequency components can be used as the fluctuation field or detail component. Optionally, the reference orientation angle... The settings can be based on prior knowledge, such as the main passageway direction in a building structure drawing, or on real-time analysis, such as using the principal direction of the spatial gradient of the base components at the current moment as... The specific form of the dynamic weighting function is not limited to the cosine function; it can also be other monotonic functions based on the direction difference.

[0059] It is understandable that the process of dynamically fusing the extended flow field with the fundamental components essentially combines the broader flow field information predicted by the model with the most stable trend components in the observation data. Dynamic weights allow the fusion process to adaptively adjust based on the reliability of the predicted flow field, thus reinforcing the data field and providing a high-quality starting point for subsequent iterative evolution. It is also understandable that the process of separating the enhanced fluctuation components from the synthetic data field and recombining them with the original high-frequency components achieves a kind of "feature injection." Synthetic data fields typically smooth out some high-frequency details, while the original high-frequency components contain these details but also noise. Through selective fusion in the time and frequency domains, beneficial fluctuation patterns evolved in the synthetic data field can be recombined with the true details of the original data while suppressing noise. The final enhanced data output, while maintaining authenticity, highlights the evolved risk characteristics.

[0060] See Figure 5 In the dynamic fusion process of the extended flow field and the basic components of the 3D data volume, the calculation of the dynamic superposition weight coefficient depends on the matching degree between the gradient direction angle of the extended flow field and the preset reference direction angle. Specifically, the dynamic superposition weight coefficient is expressed as the cosine function of the angle between the gradient direction angle of the extended flow field and the reference direction angle (range 0~1). The weight reaches its maximum value of 1 when the two directions are consistent, and the weight is 0 when the directions are opposite. The figure shows the weight curves corresponding to four different reference direction angles (0.00rad, 0.79rad, 1.57rad, 2.36rad): as the gradient direction angle (in radians) changes, each curve exhibits periodic fluctuation characteristics. The curves for different reference direction angles show phase shifts on the gradient direction angle axis, reflecting the adaptive adjustment characteristics of the weight coefficient to the reference direction. In the parameter configuration, the reference direction angle can be set based on prior knowledge (such as the direction of fire evacuation routes) or real-time analysis (such as the main direction of the spatial gradient of the basic components). The form of the dynamic weight function can be a cosine function or other monotonic functions based on the direction difference.

[0061] 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.

[0062] 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 protection data processing method based on big data, characterized in that, Includes the following steps: Receive raw fire data streams from a cluster of sensors deployed in the fire monitoring area; The original fire data stream is subjected to multi-scale spatial mapping to form a three-dimensional data volume containing spatial nodes, event channels, and time scales; Analyze the data gradient distribution inside the three-dimensional data volume to construct a spatial flow field describing the direction and intensity of data flow; In the spatial flow field, flow field singularities with convergence or divergence characteristics are identified, and connection channels between the flow field singularities are established to form a singularity network. Extract the core singularities in the singularity network that satisfy the strength threshold and connectivity threshold, and use them as key data hubs; The key data hub is input into the trained flow field generation model, which generates an extended flow field through a deep neural network structure. The extended flow field and the basic components of the three-dimensional data volume are dynamically fused to form an enhanced data field; The enhanced data field is iteratively evolved to obtain a stable synthetic data field; The enhanced fluctuation component is separated from the synthetic data field, and the enhanced fluctuation component is recombined with the high-frequency component of the original fire data stream to output the enhanced fire monitoring data.

2. The fire data processing method based on big data according to claim 1, characterized in that, The multi-scale spatial mapping of the original fire data stream includes: Define spatial scale parameters, which include macro-regional scale, meso-unit scale, and micro-viewpoint scale; For each spatial scale, extract the corresponding sensor readings, event labels, and time-stamped sequences from the raw fire data stream; The sensor readings at each spatial scale are mapped to spatial nodes, event tags are mapped to event channels, and time-scale sequences are mapped to time-scale scales. Spatial nodes, event channels, and time scales obtained by mapping at the macro-regional scale, meso-unit scale, and micro-viewpoint scale are spatially registered and hierarchically nested to form a three-dimensional data volume containing spatial nodes, event channels, and time scales.

3. The fire data processing method based on big data according to claim 1, characterized in that, The process of analyzing the data gradient distribution within the three-dimensional data volume includes: Along the spatial node dimensions of the three-dimensional data volume, the rate of change of data values ​​between adjacent spatial nodes is calculated to obtain the spatial gradient; Along the event channel dimension of the three-dimensional data volume, calculate the rate of change of correlation strength between event channels to obtain the correlation gradient; The time gradient is obtained by calculating the rate of change of the data value over time along the time scale dimension of the three-dimensional data volume. The spatial gradient, correlated gradient, and temporal gradient are vector-synthesized to construct the gradient direction and gradient magnitude of each data point, forming a spatial flow field that describes the direction and intensity of data flow.

4. The fire data processing method based on big data according to claim 3, characterized in that, The identification of flow field singularities with converging or diverging characteristics in the spatial flow field includes: Calculate the gradient vector divergence at each location point in the spatial flow field; Points with negative gradient vector divergence and absolute values ​​exceeding a preset convergence threshold are identified and marked as flow field singularities with convergence characteristics. Points with positive gradient vector divergence exceeding a preset divergence threshold are identified and marked as flow field singularities with divergence characteristics. All flow field singularities with converging characteristics and those with diverging characteristics are collectively referred to as flow field singularities.

5. The fire data processing method based on big data according to claim 4, characterized in that, Establishing the connection channels between the flow field singularities includes: In the spatial flow field, streamline tracing is performed along the gradient direction starting from each flow field singularity with divergent characteristics; Tracing the streamlines until reaching a flow field singularity with converging characteristics, the streamline path between the flow field singularity with diverging characteristics and the flow field singularity with converging characteristics is marked as a connecting channel; Repeat the streamline tracing process starting from each flow field singularity with divergent characteristics until streamline tracing is completed for all flow field singularities with divergent characteristics. All the marked connection channels together with the flow field singularity form a singularity network.

6. The fire data processing method based on big data according to claim 1, characterized in that, The extraction of core singularities in the singularity network that satisfy the strength threshold and connectivity threshold includes: Calculate the connection strength of each flow field singularity in the singularity network, whereby the connection strength is defined as the integral of the gradient magnitude of all connection channels connected to the flow field singularity; Calculate the network connectivity of each flow field singularity in the singularity network, where the network connectivity is defined as the number of other flow field singularities directly connected to the flow field singularity. Set intensity and connectivity thresholds; Singularities in the flow field with connection strength greater than the strength threshold and network connectivity greater than the connectivity threshold are selected as key data hubs.

7. The fire data processing method based on big data according to claim 1, characterized in that, The step of inputting the key data hub into the trained flow field generation model includes: The flow field generation model employs a deep neural network with an encoder-decoder architecture; The encoder of the flow field generation model performs feature encoding on the key data hubs of the input to generate a latent spatial feature vector. The decoder of the flow field generation model decodes the potential spatial feature vector to generate an extended flow field that is associated with the key data hub in terms of spatial distribution and intensity characteristics. The spatial range covered by the extended flow field is greater than the spatial range of the original distribution of the key data hub; The steps for constructing the flow field generation model include: Collect historical fire data streams and perform multi-scale spatial mapping on the historical fire data streams to form a historical three-dimensional data volume; Analyze the data gradient distribution within the historical three-dimensional data volume to construct the historical spatial flow field; In the historical spatial flow field, identify historical flow field singularities and establish connection channels between historical flow field singularities to form a historical singularity network; Extract historical key data hubs that satisfy the strength threshold and connectivity threshold from the historical singularity network; A deep neural network with an encoder-decoder architecture is constructed as a flow field generation model. The encoder consists of multiple convolutional and pooling layers, which are used to extract features and reduce dimensionality from the historical key data hubs of the input. The decoder consists of multiple deconvolutional and upsampling layers, which are used to reconstruct the extended flow field based on the feature vectors output by the encoder. Using the historical key data hub as training samples and the historical spatial flow field as the real label, the flow field generation model is trained under supervision. The model parameters are optimized by minimizing the difference between the generated flow field and the real flow field. During training, the gradient descent algorithm is used to iteratively update the model weights until the model converges, resulting in a trained flow field generation model.

8. The fire data processing method based on big data according to claim 1, characterized in that, The dynamic fusion of the extended flow field and the basic components of the three-dimensional data volume includes: The three-dimensional data volume is decomposed into low-frequency trend components, mid-frequency mode components, and high-frequency detail components. The low-frequency trend component is used as the basic component of the three-dimensional data volume; In the spatiotemporal domain, the vector field of the extended flow field is superimposed point by point with the scalar field of the fundamental component; During the superposition process, the superposition weights are dynamically adjusted according to the gradient direction of the extended flow field at each point to generate an enhanced data field.

9. The fire data processing method based on big data according to claim 1, characterized in that, The separation of the enhanced fluctuation component from the synthetic data field includes: The synthetic data field is decomposed to separate the slowly changing background field and the rapidly changing fluctuation field. The rapidly changing wave field is used as an enhanced wave component; The same field decomposition was performed on the original fire data stream to separate the original high-frequency components; The wave pattern of the enhanced wave component is aligned and fused with the detailed features of the original high-frequency component in the time and frequency domain to complete the reconstruction process and output the enhanced fire monitoring data.

10. A fire data processing system based on big data, comprising a memory, a processor, and a computer program stored in the memory and running on the processor, characterized in that, When the processor executes the computer program, it implements the steps of the fire data processing method based on big data as described in any one of claims 1 to 9.

Images