Fire-fighting event intelligent analysis method based on multi-modal fusion

Abstract

The invention discloses a fire-fighting event intelligent analysis method based on multi-modal fusion. The method comprises the following steps: collecting fire-fighting data to generate a multi-modal input sequence; pre-processing to generate a multi-modal pre-processing sequence; according to the physical evolution mechanism of the fire-fighting event, asynchronous modal self-adaptive time calibration is executed, and a calibration feature sequence is generated; modal degradation detection is executed, and a modal degradation judgment result is generated; according to a modal degradation determination result, executing inverse reconstruction to obtain a fusion input feature; creating event nodes based on the fused input features, and generating a topological time sequence coupling graph structure; inputting the topological time sequence coupling graph structure into a graph neural network to execute recursive reasoning, and generating a fire-fighting event state recognition result and an event development trend prediction result; and generating fire-fighting event risk output data according to the result. According to the method, multi-modal fusion and graph neural network reasoning are adopted, accurate recognition and trend prediction of fire-fighting events are achieved, and the method has the advantages of being high in synchronism, high in robustness and reliable in early warning.

Description

Technical Field

[0001] This invention relates to the field of fire incident analysis technology, and in particular to an intelligent analysis method for fire incidents based on multimodal fusion. Background Technology

[0002] Existing fire monitoring technologies typically rely on single or limited types of sensor data, such as video surveillance, infrared thermal imaging, or smoke concentration detection, to identify fires using fixed thresholds, static rules, or simple classification models. These monitoring methods are prone to misjudgments when faced with smoke obstruction, sudden temperature changes, illumination variations, uneven gas diffusion, or complex on-site structures. Furthermore, significant differences in time response between different sensors make it difficult to maintain the synchronization of multimodal data throughout the event's evolution, resulting in problems such as identification delays, high false alarm rates, and untimely status updates.

[0003] Furthermore, existing technologies generally lack the ability to automatically identify and correct multimodal data degradation. When there is sensor interference, data acquisition interruption, data loss, or increased modal noise, the monitoring system cannot self-recover the degraded modes, resulting in the overall monitoring results relying on a single mode and thus reducing accuracy. Existing methods also typically cannot construct a time-series structure that reflects the causal relationship between changes in flames, smoke, hot spots, gases, and acoustics. It is difficult to achieve event state identification and development trend prediction through a unified model, and thus cannot meet the real-time high-reliability analysis requirements in complex fire-fighting scenarios. Summary of the Invention

[0004] One objective of this invention is to propose an intelligent analysis method for fire incidents based on multimodal fusion. This invention employs multimodal fusion and graph neural network inference to achieve accurate identification and trend prediction of fire incidents, and has the advantages of strong synchronization, high robustness and reliable early warning.

[0005] A fire incident intelligent analysis method based on multimodal fusion according to an embodiment of the present invention includes the following steps:

[0006] Acquire video images, infrared thermal imaging, smoke concentration, temperature changes, combustible gas concentration, and acoustic shock wave data to construct a multimodal input sequence;

[0007] Noise filtering and missing segment imputation are performed on the multimodal input sequence to generate a multimodal preprocessed sequence;

[0008] Based on the physical evolution mechanism of fire incidents, asynchronous modal adaptive time calibration is performed on the multimodal preprocessing sequence. The flame brightness change rate, smoke diffusion gradient, heat conduction delay and gas concentration change slope are calculated. A modal change rate function is constructed. Based on this function, a nonlinear time calibration window is determined. The multimodal features are corrected on a time scale and a calibration feature sequence is generated.

[0009] Modal degradation detection is performed on each mode in the calibration feature sequence, and the modal signal stability, spatial structure preservation and noise level are calculated to generate modal degradation judgment results;

[0010] Based on the modal degradation determination results, the degraded modal features are reconstructed in reverse, and the recovered features are merged with the non-degraded modal features to obtain the fused input features;

[0011] Event nodes describing changes in flame, smoke, hot spots, gas, and acoustics are created based on fused input features, and a topological temporal coupling graph structure is generated based on physical causal relationships and temporal dependencies.

[0012] The topological temporally coupled graph structure is input into the graph neural network to perform recursive inference, generating fire incident status identification results and incident development trend prediction results.

[0013] Based on the results, fire incident risk output data is generated.

[0014] Optionally, the generation of the multimodal input sequence specifically includes:

[0015] Video image data, infrared thermal imaging data, smoke concentration data, temperature change data, combustible gas concentration data, and acoustic shock wave data are acquired according to the preset sampling frequency.

[0016] The acquired sampling data is stored in chronological order according to the corresponding acquisition time.

[0017] Based on the data collection time, various types of data are indexed and marked with a time index;

[0018] The six types of data at the same sampling time are combined according to the time index, and a multimodal data unit is generated for each sampling time.

[0019] The multimodal data units formed in chronological order are arranged sequentially to construct a multimodal input sequence.

[0020] Optionally, the generation of the multimodal preprocessing sequence specifically includes:

[0021] The multimodal input sequence is initially screened, and noise reduction processing is performed on the six types of data in the sequence to remove interference information generated by the acquisition device, thereby generating a noise-reduced multimodal dataset.

[0022] The continuity of the noise-reduced dataset is checked one by one to identify data gaps caused by acquisition interruptions or abnormal readings.

[0023] For data gaps, while maintaining the chronological order, the corresponding interpolation content is added to generate a data sequence without interruption on a unified time reference axis;

[0024] Various types of data are reorganized according to a unified time reference axis to generate structured multimodal time segments;

[0025] All corresponding multimodal time segments are integrated according to their time sequence to generate a multimodal preprocessing sequence.

[0026] Optionally, the generation of the calibration feature sequence specifically includes:

[0027] Dynamic features of video image data, infrared thermal imaging data, smoke concentration data, temperature change data, combustible gas concentration data and acoustic shock wave data over time are extracted from the multimodal preprocessing sequence.

[0028] The corresponding flame brightness change rate, smoke diffusion gradient, heat conduction delay, and gas concentration change slope are calculated based on the dynamic characteristics of each mode.

[0029] A modal change rate function is constructed based on the flame brightness change rate, smoke diffusion gradient, heat conduction delay, and gas concentration change slope.

[0030] A nonlinear time calibration window is generated based on the modal change rate function to calibrate the time position of video image data, infrared thermal imaging data, smoke concentration data, temperature change data, combustible gas concentration data, and acoustic shock wave data.

[0031] The six types of data calibrated by nonlinear time are arranged in the order of calibration time, so that the six types of modes form a consistent time series structure under a unified time reference;

[0032] Under the calibration time reference, the features of each mode at the same time position are merged to generate a calibration feature unit that includes video image features, infrared thermal imaging features, smoke concentration features, temperature change features, combustible gas concentration features and acoustic shock wave features.

[0033] Generate a calibration feature sequence by arranging all calibration feature units in the order of calibration time.

[0034] Optionally, the generation of the modal degradation determination result specifically includes:

[0035] Under each calibration time index, video image features, infrared thermal imaging features, smoke concentration features, temperature change features, combustible gas concentration features, and acoustic shock wave features are extracted item by item from the calibration feature sequence, and the six types of features are recorded as corresponding modal feature values.

[0036] For each calibration time index, perform signal fluctuation detection operation on the six modal feature values ​​recorded to generate signal stability detection values ​​for the six modalities under that time index;

[0037] For each calibration time index, the structure retention test operation is performed on the six types of modal characteristic values ​​recorded, and the test results are recorded as the spatial structure retention test values ​​of the six types of modes under that time index;

[0038] For each calibration time index, perform noise intensity detection on the six modal characteristic values ​​recorded, and record the detection results as the noise level detection values ​​of the six modalities under that time index;

[0039] Under the same calibration time index, the signal stability test values, spatial structure retention test values, and noise level test values ​​corresponding to the six modes are combined into a test data group in a fixed order;

[0040] The detection and judgment operations are performed on each item of the detection data set to generate degradation mark values ​​for the six modes respectively. The generated degradation mark values ​​are recorded as video image mode degradation mark, infrared thermal imaging mode degradation mark, smoke mode degradation mark, temperature change mode degradation mark, combustible gas concentration mode degradation mark and acoustic shock wave mode degradation mark under the time index.

[0041] The six types of modal degradation markers generated under each calibration time index are recorded sequentially in chronological order to generate modal degradation judgment results.

[0042] Optionally, the generation of the fused input features specifically includes:

[0043] Based on the modal degradation determination results, the features corresponding to the degraded and non-degraded modes are extracted one by one according to the calibration time index and recorded as the degraded feature set and the non-degraded feature set, respectively.

[0044] The degenerate modal features under each time index in the degenerate feature set are read item by item to obtain the degenerate modal features;

[0045] Perform inverse feature deduction operation on the degenerate mode features of each time index to generate the recovery features under the corresponding time index;

[0046] The recovery features are recorded one by one according to the calibration time index, and accumulated in chronological order to generate a recovery feature sequence;

[0047] The set of non-degraded features is read one by one according to the calibration time index and accumulated and recorded in chronological order as a sequence of non-degraded features;

[0048] Under each calibration time index, the recovery feature at the corresponding time position in the recovery feature sequence is read, and the non-degraded modal feature at the corresponding time position in the non-degraded feature sequence is read. The two types of features are then concatenated in a fixed order to generate multimodal concatenated features.

[0049] The multimodal splicing features are written one by one according to the calibration time index to generate a splicing feature sequence arranged in chronological order;

[0050] The concatenated feature sequence is written sequentially from the first to the last into the fusion feature record structure, and all the concatenated features after writing are used as the fusion input features.

[0051] Optionally, the generation of the topological temporal coupling graph structure specifically includes:

[0052] Flame change features, smoke change features, hot spot change features, gas change features, and acoustic change features are extracted from the fused input features according to the calibration time index.

[0053] Under each calibration time index, the flame change characteristics are written to the flame event node, the smoke change characteristics are written to the smoke event node, the hot spot change characteristics are written to the hot spot event node, the gas change characteristics are written to the gas event node, and the acoustic change characteristics are written to the acoustic event node.

[0054] Under the same calibration time index, the five types of event nodes are connected to each other in a preset node order to generate parallel time-series edges;

[0055] Between adjacent calibration time indices, flame event nodes, smoke event nodes, hot spot event nodes, gas event nodes, and acoustic event nodes of the same type are connected to each other in ascending time order to generate time-progressive edges;

[0056] Under each calibration time index, the corresponding event nodes are connected sequentially according to the physical causal order of flame change, smoke change, hot spot change, gas change and acoustic change to generate cross-modal causal edges;

[0057] The temporal edges, time-progressive edges, and cross-modal causal edges are written into the graph structure record cells in the order of calibration time index to generate a topological temporal coupled graph structure.

[0058] Optionally, the recursive reasoning specifically includes:

[0059] Flame event node features, smoke event node features, hot spot event node features, gas event node features, and acoustic event node features are extracted from the topological time-series coupled graph structure according to the calibration time index. Node data loading operations are performed on the extracted node features to generate node input data.

[0060] Extract temporal edge connections, time-progressive edge connections, and cross-modal causal edge connections from the topological temporal coupling graph structure in the order of records. Perform edge data loading operations on the extracted connections to generate edge input data.

[0061] In graph neural networks, node aggregation processing is performed on node input data and edge input data. Neighborhood feature aggregation is performed on each event node to generate the first-stage node aggregation feature.

[0062] The first-stage node aggregation features are transformed by performing feature transformation processing, which transforms the aggregation features of each event node to generate node transformation features.

[0063] In the recursive inference processing area, the node transformation features are recursively calculated according to the calibration time index order, and the node transformation features of each time index are iteratively updated to generate a recursive node feature sequence.

[0064] Under the same calibration time index, recursive features of flame event nodes, smoke event nodes, hot spot event nodes, gas event nodes, and acoustic event nodes are extracted from the recursive node feature sequence, respectively. Vector concatenation is performed on the five types of recursive features to generate a state vector.

[0065] For each calibration time index, perform state output processing on the state vector, parse the fields in the state vector into an event category field and an event severity field, and generate fire event status identification results;

[0066] Trend calculation processing is performed on the state vectors generated under multiple continuous time indices, and sequential operations are performed on the state vector sequence to generate the prediction result of the development trend of fire incidents.

[0067] Optionally, the generation of the fire incident risk output data specifically includes:

[0068] Under each calibration time index, the event category field and the event severity field are extracted from the fire event status identification results. Field acquisition processing is performed on the two extracted fields to generate a status field set.

[0069] Under multiple continuous calibration time indices, the event development trend field is extracted from the fire event development trend prediction results, and the extracted trend field is processed to generate a trend field sequence.

[0070] Under the same calibration time index, perform field combination processing on the event category field and event severity field in the status field set, and extract the event development trend field under the corresponding time index from the trend field sequence, and combine the three types of fields to generate a risk basic field group;

[0071] Perform field formatting on the risk base field group, and generate the formatted field content into a risk structure field group;

[0072] In the risk structure field group, perform event category field positioning on the event category field, event severity field positioning on the event severity field, and event trend field positioning on the event development trend field. Generate risk field unit from the positioned field set.

[0073] Under each calibration time index, risk entry construction processing is performed on the risk field unit, and the risk field unit is arranged into risk data entries.

[0074] Under multiple calibration time indices, sequential splicing is performed on all risk data entries to generate a risk data sequence from the spliced ​​data entries;

[0075] After the risk data sequence is generated, data integration processing is performed on the risk data sequence, and the integrated data is used to generate fire incident risk output data.

[0076] The beneficial effects of this invention are:

[0077] This invention introduces a multimodal data collaborative processing mechanism to achieve real-time acquisition, unified organization, and structured representation of six types of data, including video images, infrared thermal imaging, smoke concentration, temperature changes, combustible gas concentration, and acoustic shock waves. It overcomes the technical limitations of existing fire monitoring technologies, such as large differences in the time response of multimodal data and weak modal coupling. By setting up noise filtering, time calibration, and missing segment interpolation processes, the multimodal data forms a preprocessed sequence that is temporally continuous, structurally consistent, and content-complete before entering the event analysis stage. This effectively avoids monitoring instability caused by acquisition noise, sensor delays, and data discontinuities. The asynchronous modal adaptive time calibration strategy, based on the physical evolution mechanism of fire events, realigns key dynamic features such as flame brightness changes, smoke diffusion, heat conduction, and gas concentration changes to a unified event rhythm, ensuring consistent temporal orientation of different modalities during fire development. This fundamentally solves the shortcomings of existing technologies in reliably synchronizing multimodal data in complex backgrounds and dynamic environments.

[0078] After completing time calibration, this invention further performs modal degradation detection and inverse reconstruction on each modality, enabling real-time recovery of modal features under conditions of data degradation, acquisition interruption, or increased noise. This achieves dynamic integrity assurance of multimodal data and significantly improves data fusion quality. Subsequently, based on the actual physical causal relationships between flames, smoke, hot spots, gases, and acoustic changes, this invention constructs a topological temporal coupling graph structure. This structure uniformly expresses the correlation between events in space, time, and modality, forming a structured event chain in a graph format. This provides a reliable foundation for global modeling of complex fires. By using graph neural networks to perform recursive reasoning, the system can continuously update event states based on the relationship features between nodes and generate event categories, event severity, and event development trend prediction results on multiple time indices, achieving real-time identification and forward-looking prediction of fires. The final generated fire incident risk output data covers three core information categories: incident category, incident severity, and development trend, enabling the system to have high adaptability, robustness, and early warning capabilities. This invention achieves breakthroughs in multimodal data synchronization, modal degradation recovery, event structured modeling, and intelligent reasoning, effectively improving the accuracy, timeliness, and reliability of fire incident monitoring, and meeting the actual needs for high-precision fire identification and risk prediction in complex scenarios. Attached Figure Description

[0079] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings:

[0080] Figure 1 This is a flowchart of a fire incident intelligent analysis method based on multimodal fusion proposed in this invention;

[0081] Figure 2 This is a schematic diagram of the asynchronous modal adaptive time calibration process of a fire incident intelligent analysis method based on multimodal fusion proposed in this invention. Detailed Implementation

[0082] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0083] refer to Figure 1 and Figure 2 A fire incident intelligent analysis method based on multimodal fusion includes the following steps:

[0084] Acquire video images, infrared thermal imaging, smoke concentration, temperature changes, combustible gas concentration, and acoustic shock wave data to construct a multimodal input sequence;

[0085] Noise filtering and missing segment imputation are performed on the multimodal input sequence to generate a multimodal preprocessed sequence;

[0086] Based on the physical evolution mechanism of fire incidents, asynchronous modal adaptive time calibration is performed on the multimodal preprocessing sequence. The flame brightness change rate, smoke diffusion gradient, heat conduction delay and gas concentration change slope are calculated. A modal change rate function is constructed. Based on this function, a nonlinear time calibration window is determined. The multimodal features are corrected on a time scale and a calibration feature sequence is generated.

[0087] Modal degradation detection is performed on each mode in the calibration feature sequence, and the modal signal stability, spatial structure preservation and noise level are calculated to generate modal degradation judgment results;

[0088] Based on the modal degradation determination results, the degraded modal features are reconstructed in reverse, and the recovered features are merged with the non-degraded modal features to obtain the fused input features;

[0089] Event nodes describing changes in flame, smoke, hot spots, gas, and acoustics are created based on fused input features, and a topological temporal coupling graph structure is generated based on physical causal relationships and temporal dependencies.

[0090] The topological temporally coupled graph structure is input into the graph neural network to perform recursive inference, generating fire incident status identification results and incident development trend prediction results.

[0091] Based on the results, fire incident risk output data is generated.

[0092] In this embodiment, the generation of the multimodal input sequence specifically includes:

[0093] Video image data, infrared thermal imaging data, smoke concentration data, temperature change data, combustible gas concentration data, and acoustic shock wave data are acquired according to the preset sampling frequency.

[0094] The acquired sampling data is stored in chronological order according to the corresponding acquisition time.

[0095] Based on the data collection time, various types of data are indexed and marked with a time index;

[0096] The six types of data at the same sampling time are combined according to the time index, and a multimodal data unit is generated for each sampling time.

[0097] The multimodal data units formed in chronological order are arranged sequentially to construct a multimodal input sequence containing video image data, infrared thermal imaging data, smoke concentration data, temperature change data, combustible gas concentration data, and acoustic shock wave data.

[0098] In this embodiment, the generation of the multimodal preprocessing sequence specifically includes:

[0099] The multimodal input sequence is initially screened, and noise reduction processing is performed on the six types of data in the sequence to remove interference information generated by the acquisition device, thereby generating a noise-reduced multimodal dataset.

[0100] The continuity of the noise-reduced dataset is checked one by one to identify data gaps caused by acquisition interruptions or abnormal readings.

[0101] For data gaps, while maintaining the chronological order, the corresponding interpolation content is added to generate a data sequence without interruption on a unified time reference axis;

[0102] All types of data are reorganized according to a unified time reference axis, so that each time point contains video image data, infrared thermal imaging data, smoke concentration data, temperature change data, combustible gas concentration data and acoustic shock wave data, generating structured multimodal time segments.

[0103] All corresponding multimodal time segments are integrated according to their time sequence to generate a multimodal preprocessing sequence.

[0104] In this embodiment, the generation of the calibration feature sequence specifically includes:

[0105] Dynamic features of video image data, infrared thermal imaging data, smoke concentration data, temperature change data, combustible gas concentration data and acoustic shock wave data over time are extracted from the multimodal preprocessing sequence to characterize the temporal behavior of flame brightness change, smoke diffusion change, heat conduction change and gas concentration change.

[0106] The dynamic feature extraction specifically includes: reading the original data of the six types of data one by one according to the time index to form time segments; calculating the numerical difference between adjacent time segments to form the change amplitude value; and arranging the change amplitude values ​​in the order of the time index to generate dynamic features;

[0107] Based on the dynamic characteristics of each mode, the corresponding flame brightness change rate, smoke diffusion gradient, heat conduction delay and gas concentration change slope are calculated, so that the change rate of each mode in the fire event evolution process can be represented in a time sequence.

[0108] The calculation of the flame brightness change rate, smoke diffusion gradient, heat conduction delay, and gas concentration change slope specifically includes: reading adjacent brightness values ​​one by one according to the time index of the dynamic features of video image data, calculating the brightness change, and dividing the change by the corresponding time interval to obtain the flame brightness change rate; reading adjacent concentration values ​​one by one according to the time index of the dynamic features of smoke concentration data, calculating the concentration difference, and dividing the difference by the corresponding time interval to obtain the smoke diffusion gradient; reading adjacent temperature values ​​one by one according to the time index of the dynamic features of infrared thermal imaging data, calculating the temperature change, determining the offset position of the change on the time index, and using this offset position as the time lag of the thermal response to obtain the heat conduction delay; and reading adjacent concentration change values ​​one by one according to the time index of the dynamic features of combustible gas concentration data, calculating the change amplitude, and dividing the change amplitude by the corresponding time interval to obtain the gas concentration change slope.

[0109] Based on the flame brightness change rate, smoke diffusion gradient, heat conduction delay and gas concentration change slope, a modal change rate function is constructed to provide a unified description of the response speed of the six modes in the time dimension.

[0110] The construction of the modal change rate function specifically includes: reading the brightness change amount by time indexing the flame brightness change rate, and performing a fixed differential calculation step for each brightness change amount, and writing the differential calculation step into the flame change processing flow; reading the diffusion change amount by time indexing the smoke diffusion gradient, and performing a differential calculation step in the same order as the flame change processing flow for each diffusion change amount, and writing this step into the smoke change processing flow; reading the delay amount by time indexing the heat conduction delay, and performing a fixed delay comparison step for each delay amount, and writing this step into the heat conduction change processing flow; reading the concentration change amount by time indexing the gas concentration change slope, and performing a fixed change amount comparison step for each concentration change amount, and writing this step into the gas change processing flow; combining the flame change processing flow, smoke change processing flow, heat conduction change processing flow, and gas change processing flow in a predetermined execution order to form a unified processing flow that can perform fixed differential calculation, fixed delay comparison, and fixed change amount comparison on any input change amount, and defining this unified processing flow as the modal change rate function.

[0111] A nonlinear time calibration window is generated based on the modal change rate function to calibrate the time position of video image data, infrared thermal imaging data, smoke concentration data, temperature change data, combustible gas concentration data, and acoustic shock wave data, so that different modes are synchronized with the same event evolution reference point on the time axis.

[0112] The generation of the nonlinear time calibration window specifically includes: generating time displacement records as a calibration displacement value sequence based on the modal change rate value index by index; arranging the calibration displacement value sequence in time index order to generate a nonlinear calibration window value; and using the nonlinear calibration window value as a calibration reference.

[0113] The six types of data calibrated by nonlinear time are arranged in the order of calibration time, so that the six types of modes form a consistent time series structure under a unified time reference;

[0114] Under the calibration time reference, the features of each mode at the same time position are merged to generate a calibration feature unit that includes video image features, infrared thermal imaging features, smoke concentration features, temperature change features, combustible gas concentration features and acoustic shock wave features.

[0115] Generate a calibration feature sequence by arranging all calibration feature units in the order of calibration time.

[0116] In this embodiment, the generation of the modal degradation determination result specifically includes:

[0117] Under each calibration time index, video image features, infrared thermal imaging features, smoke concentration features, temperature change features, combustible gas concentration features, and acoustic shock wave features are extracted item by item from the calibration feature sequence, and the six types of features are recorded as corresponding modal feature values.

[0118] For each calibration time index, perform signal fluctuation detection operation on the six modal feature values ​​recorded to generate signal stability detection values ​​for the six modalities under that time index;

[0119] For each calibration time index, the structure retention test operation is performed on the six types of modal characteristic values ​​recorded, and the test results are recorded as the spatial structure retention test values ​​of the six types of modes under that time index;

[0120] For each calibration time index, perform noise intensity detection on the six modal characteristic values ​​recorded, and record the detection results as the noise level detection values ​​of the six modalities under that time index;

[0121] Under the same calibration time index, the signal stability test values, spatial structure retention test values, and noise level test values ​​corresponding to the six modes are combined into a test data group in a fixed order;

[0122] The detection and judgment operations are performed on each item of the detection data set to generate degradation mark values ​​for the six modes respectively. The generated degradation mark values ​​are recorded as video image mode degradation mark, infrared thermal imaging mode degradation mark, smoke mode degradation mark, temperature change mode degradation mark, combustible gas concentration mode degradation mark and acoustic shock wave mode degradation mark under the time index.

[0123] The six types of modal degradation markers generated under each calibration time index are recorded sequentially in chronological order to generate modal degradation judgment results.

[0124] In this embodiment, the generation of the fused input features specifically includes:

[0125] Based on the modal degradation determination results, the features corresponding to the degraded and non-degraded modes are extracted one by one according to the calibration time index and recorded as the degraded feature set and the non-degraded feature set, respectively.

[0126] The degenerate modal features under each time index in the degenerate feature set are read item by item to obtain the degenerate modal features;

[0127] The degradation feature reading specifically includes: locating the feature records of the degradation mode at the corresponding time position one by one according to the calibration time index; performing a single data extraction operation on the located feature records to generate degradation feature values; and generating index-by-index reading results by ordering the generated degradation feature values ​​according to the time index.

[0128] Perform inverse feature deduction operation on the degenerate mode features of each time index to generate the recovery features under the corresponding time index;

[0129] The reverse feature estimation specifically includes: reading the degenerate modal features one by one according to the time index and performing a reverse calculation step to generate the reverse calculation result; performing a correction comparison operation between the reverse calculation result and the degenerate modal features under the same time index to generate the correction result; and recording the correction result as the recovery feature at the corresponding time position according to the time index.

[0130] The recovery features are recorded one by one according to the calibration time index, and accumulated in chronological order to generate a recovery feature sequence;

[0131] The set of non-degraded features is read one by one according to the calibration time index and accumulated and recorded in chronological order as a sequence of non-degraded features;

[0132] Under each calibration time index, the recovery feature at the corresponding time position in the recovery feature sequence is read, and the non-degraded modal feature at the corresponding time position in the non-degraded feature sequence is read. The two types of features are then concatenated in a fixed order to generate multimodal concatenated features.

[0133] The multimodal splicing features are written one by one according to the calibration time index to generate a splicing feature sequence arranged in chronological order;

[0134] The concatenated feature sequence is written sequentially from the first to the last into the fusion feature record structure, and all the concatenated features after writing are used as the fusion input features.

[0135] In this embodiment, the generation of the topological temporal coupling graph structure specifically includes:

[0136] Flame change features, smoke change features, hot spot change features, gas change features, and acoustic change features are extracted from the fused input features according to the calibration time index.

[0137] Under each calibration time index, the flame change characteristics are written to the flame event node, the smoke change characteristics are written to the smoke event node, the hot spot change characteristics are written to the hot spot event node, the gas change characteristics are written to the gas event node, and the acoustic change characteristics are written to the acoustic event node.

[0138] Under the same calibration time index, the five types of event nodes are connected to each other in a preset node order to generate parallel time-series edges;

[0139] The generation of the time-series edges of the parallel relationships specifically includes: performing node connection operations on flame event nodes, smoke event nodes, hot spot event nodes, gas event nodes, and acoustic event nodes in a preset node order under the same calibration time index; recording the generated node connections as time-series edge groups; and generating time-series edges of the parallel relationships by recording the time-series edge groups according to the time index.

[0140] Between adjacent calibration time indices, flame event nodes, smoke event nodes, hot spot event nodes, gas event nodes, and acoustic event nodes of the same type are connected to each other in ascending time order to generate time-progressive edges;

[0141] Under each calibration time index, the corresponding event nodes are connected sequentially according to the physical causal order of flame change, smoke change, hot spot change, gas change and acoustic change to generate cross-modal causal edges;

[0142] The generation of the cross-modal causal edge specifically includes: performing node connection operations sequentially in a fixed order of flame change, smoke change, hot spot change, gas change, and acoustic change under the same calibration time index; recording each node connection as a causal connection record; and forming a cross-modal causal edge by arranging all causal connection records in index order.

[0143] The temporal edges, time-progressive edges, and cross-modal causal edges are written into the graph structure record cells in the order of calibration time index to generate a topological temporal coupled graph structure.

[0144] In this embodiment, the recursive reasoning specifically includes:

[0145] Flame event node features, smoke event node features, hot spot event node features, gas event node features, and acoustic event node features are extracted from the topological time-series coupled graph structure according to the calibration time index. Node data loading operations are performed on the extracted node features to generate node input data.

[0146] Extract temporal edge connections, time-progressive edge connections, and cross-modal causal edge connections from the topological temporal coupling graph structure in the order of records. Perform edge data loading operations on the extracted connections to generate edge input data.

[0147] In graph neural networks, node aggregation processing is performed on node input data and edge input data. Neighborhood feature aggregation is performed on each event node to generate the first-stage node aggregation feature.

[0148] The generation of node aggregation features in the first stage specifically includes: reading node features from the node input data by node index; determining the set of adjacent nodes for each node based on the edge input data; performing a node-by-node accumulation operation on the features of the adjacent node set to generate aggregation feature values; and recording the aggregation feature values ​​by node index as the node aggregation features of the first stage.

[0149] The first-stage node aggregation features are transformed by performing feature transformation processing, which transforms the aggregation features of each event node to generate node transformation features.

[0150] In the recursive inference processing area, the node transformation features are recursively calculated according to the calibration time index order, and the node transformation features of each time index are iteratively updated to generate a recursive node feature sequence.

[0151] The generation of the recursive node feature sequence specifically includes: reading the node transformation features index by index according to the calibration time index to generate time step features; performing iterative updates on the time step features step by step to generate recursive update values; and recording the recursive update values ​​as a recursive node feature sequence according to the time index.

[0152] Under the same calibration time index, recursive features of flame event nodes, smoke event nodes, hot spot event nodes, gas event nodes, and acoustic event nodes are extracted from the recursive node feature sequence, respectively. Vector concatenation is performed on the five types of recursive features to generate a state vector.

[0153] For each calibration time index, perform state output processing on the state vector, parse the fields in the state vector into an event category field and an event severity field, and generate fire event status identification results;

[0154] Trend calculation processing is performed on the state vectors generated under multiple continuous time indices, and sequential operations are performed on the state vector sequence to generate the prediction result of the development trend of fire incidents.

[0155] In this embodiment, the generation of the fire incident risk output data specifically includes:

[0156] Under each calibration time index, the event category field and the event severity field are extracted from the fire event status identification results. Field acquisition processing is performed on the two extracted fields to generate a status field set.

[0157] Under multiple continuous calibration time indices, the event development trend field is extracted from the fire event development trend prediction results, and the extracted trend field is processed to generate a trend field sequence.

[0158] Under the same calibration time index, perform field combination processing on the event category field and event severity field in the status field set, and extract the event development trend field under the corresponding time index from the trend field sequence, and combine the three types of fields to generate a risk basic field group;

[0159] The generation of the risk basic field group specifically includes: reading the event category field and the event severity field from the status field set to form a status field pair; reading the event development trend field under the corresponding time index from the trend field sequence to form a trend field value; and combining the status field pair and the trend field value in a fixed field order to form the risk basic field group.

[0160] Perform field formatting on the risk base field group, and generate the formatted field content into a risk structure field group;

[0161] In the risk structure field group, perform event category field positioning on the event category field, event severity field positioning on the event severity field, and event trend field positioning on the event development trend field. Generate risk field unit from the positioned field set.

[0162] Under each calibration time index, risk entry construction processing is performed on the risk field unit, and the risk field unit is arranged into risk data entries.

[0163] Under multiple calibration time indices, sequential splicing is performed on all risk data entries to generate a risk data sequence from the spliced ​​data entries;

[0164] After the risk data sequence is generated, data integration processing is performed on the risk data sequence, and the integrated data is used to generate fire incident risk output data.

[0165] Example 1:

[0166] To verify the feasibility of this invention in practice, it was applied to a fire monitoring scenario located within an urban commercial complex. This area includes shops, corridors, storage rooms, air conditioning rooms, and some internal pipeline passages, exhibiting a complex structure, high personnel mobility, and multiple potential fire sources. Video acquisition devices, infrared thermal imaging devices, smoke sensors, temperature sensors, combustible gas concentration sensors, and acoustic shock wave sensors were deployed simultaneously in this scenario to achieve real-time monitoring of various risk factors such as flames, smoke, heat, combustible gas leaks, and acoustic disturbances. Due to uneven airflow, shelving obstructions, high equipment noise, and significant day-night light variations in this scenario, traditional multimodal monitoring methods are prone to problems such as data asynchrony, severe modal degradation, and difficulty in analyzing causal relationships, making it impossible to reliably judge the development process of complex fires. This invention was applied in this scenario to verify its performance in terms of multimodal synchronization consistency, modal degradation recovery capability, event structure representation capability, and intelligent reasoning capability.

[0167] In this scenario, the invention first utilizes six types of installed sensors to continuously collect multi-source data, including video images, infrared thermal imaging, smoke concentration, temperature changes, combustible gas concentration, and acoustic shock waves. All data types are stored within a unified time frame, forming a multimodal data foundation with true business continuity. As data continuously enters the monitoring system, the system performs noise filtering, time base alignment, and missing segment imputation, integrating data from different devices into a structurally consistent multimodal time segment. This presents a complete information structure of the scene at different time locations. The pre-processed data achieves continuity in the time dimension, with each time location containing all six types of collected data, thus avoiding information loss caused by temporary sensor failures or acquisition delays.

[0168] As personnel movement, air conditioning wind patterns, and equipment obstruction constantly change in the monitoring scenario, some modal data exhibit significant asynchronous response phenomena. For example, the video brightness changes faster than the smoke concentration changes during flame diffusion, while the heat conduction response lags behind the infrared temperature rise. This invention dynamically extracts the trends of flame brightness changes, smoke diffusion patterns, heat conduction changes, and gas concentration changes, and generates modal change rate values ​​based on these characteristics, enabling a quantitative expression of the response speed of different modalities during fire development. The system further generates a nonlinear time calibration window based on the differences in change rates, recalibrating the positions of each modality on the time axis. This ensures that the moments of change perceived by different modalities during fire development are remapped to the same event reference point. After calibration, the multimodal data maintains consistency in temporal directionality, allowing the system to observe the changes of each modality at different stages of the event at a uniform pace, effectively solving the common time misalignment problem in traditional multimodal monitoring methods.

[0169] During scenario operation, the movement of goods in the warehouse obstructs the viewpoints of some sensors, resulting in significant degradation of video image modalities. Noise from the air conditioning room degrades acoustic data quality at certain times, and changes in ambient humidity cause shifts in gas concentration readings. This invention identifies degraded modal data by performing modal degradation detection on each modality in the calibration feature sequence, comparing signal stability, spatial structure preservation, and noise levels. After identification, the degraded modal features are reverse-processed by performing reverse calculations on the original degraded features and correcting and comparing the calculation results, thus restoring the degraded modal features. The restored data is then combined with the non-degraded modal data, ensuring the system maintains the integrity of multimodal information throughout the monitoring period, thereby avoiding false alarms and missed alarms caused by severe modal degradation.

[0170] After completing modal recovery, this invention constructs event nodes describing changes in flames, smoke, hot spots, gases, and acoustics based on the fused multimodal features. A topological temporal coupling graph structure is established according to the causal relationships between events in space, time, and modalities, allowing the entire process of a fire from its initiation and spread to its development to be presented coherently through the graph structure. Based on this structure, this invention inputs the graph into a graph neural network to perform recursive reasoning, enabling the system to continuously generate event categories and severity levels during scenario operation. The node states are continuously updated as events change at different time locations, thus demonstrating a dynamic understanding of fire situations in real-world monitoring. The state vectors from multiple time locations are further used to construct event development trends, allowing the system to predict the future direction of a fire in its early stages, providing on-site management personnel with a basis for early intervention.

[0171] In actual operation, this invention has maintained stable operation within complex, noisy commercial complexes with significant lighting variations. Even when faced with multimodal degradation, brief sensor outages, and environmental interference, it can still output complete, continuous, and structured monitoring results. The risk output data includes three core fields: event category, event severity, and event development trend. This allows managers to intuitively grasp the current status and future trends through a unified interface, providing a reliable basis for timely and effective measures. This embodiment demonstrates that this invention has achieved significant results in multimodal data synchronization, degradation repair, structured event modeling, and intelligent fire reasoning. It can effectively solve the problems of time asynchrony, severe modal degradation, and inability to predict event trends in complex environments of existing technologies in real-world scenarios.

[0172] Table 1 Performance Evaluation Table of Multimodal Temporal Coordination and Structured Inference Model

[0173]

[0174] As shown in Table 1, the overall performance of this invention in multimodal data processing is improved to a certain extent compared with traditional methods. First, in terms of synchronization consistency, the modal time alignment accuracy of this invention is improved from 82.3% of the traditional method to 91.4%, an improvement of about 9.1%. This improvement is mainly related to the asynchronous modal adaptive time calibration strategy proposed in this invention. This strategy constructs a time reference based on dynamic features such as flame brightness change, smoke diffusion, heat conduction and gas change, so that the responses of different modalities can be closer to the same event node, thereby reducing the error caused by time offset. In the effective fusion ratio under asynchronous conditions, this invention improves by 14.8% compared with the traditional method, reflecting that the participation of multimodalities under different time-series conditions is enhanced.

[0175] In terms of robustness, the modal degradation recognition accuracy of the present invention is improved by about 15.7% compared with the traditional method, indicating that the modal recognition capability of the system is more stable under sensor noise, occlusion or data fluctuation. The recovery availability after modal degradation is improved from 54.7% to 78.4%, an increase of 23.7%, indicating that the degraded modes can re-participate in the subsequent analysis process after the reverse processing of the present invention, so that the overall information structure remains more complete. This recovery capability also helps to reduce the system judgment bias caused by local modal failure.

[0176] In terms of recognition and prediction performance, the accuracy of fire event status recognition is improved by about 5.2%, and the consistency rate of trend prediction is improved by 11.4%. This is related to the topological time-series coupled graph structure constructed by the present invention and the recursive reasoning based on graph neural network. By continuously updating the relationship between event nodes, the system can better track changes in the fire situation and form continuous judgments between multiple time slices, thereby improving the matching degree between the prediction results and the actual trend.

[0177] In terms of comprehensive indicators, the present invention increases the early warning capability for severe fires by about 18%, indicating that the present invention can capture risk signals more promptly in the early stages of a fire. At the same time, in terms of false alarm rate, the present invention reduces the false alarm rate by about 17% compared with traditional methods. This is related to the improved time consistency between multiple modes and the enhanced modal degradation recovery capability, enabling the system to maintain more stable judgment performance under various environmental disturbances.

[0178] Overall, the data in the table reflects that the present invention has improved in terms of synchronization, data integrity, reasoning continuity and prediction stability. These improvements are due to the combined effect of the multimodal adaptive time calibration mechanism, modality degradation processing flow, topological time-series coupling graph structure construction and recursive reasoning ability based on graph neural network, which enables the present invention to maintain relatively reliable analysis performance in complex monitoring environments.

[0179] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A fire incident intelligent analysis method based on multimodal fusion, characterized in that, Includes the following steps: Acquire video images, infrared thermal imaging, smoke concentration, temperature changes, combustible gas concentration, and acoustic shock wave data to construct a multimodal input sequence; Noise filtering and missing segment imputation are performed on the multimodal input sequence to generate a multimodal preprocessed sequence; Based on the physical evolution mechanism of fire incidents, asynchronous modal adaptive time calibration is performed on the multimodal preprocessing sequence. The flame brightness change rate, smoke diffusion gradient, heat conduction delay and gas concentration change slope are calculated. A modal change rate function is constructed. Based on this function, a nonlinear time calibration window is determined. The multimodal features are corrected on a time scale and a calibration feature sequence is generated. Modal degradation detection is performed on each mode in the calibration feature sequence, and the modal signal stability, spatial structure preservation and noise level are calculated to generate modal degradation judgment results; Based on the modal degradation determination results, the degraded modal features are reconstructed in reverse, and the recovered features are merged with the non-degraded modal features to obtain the fused input features; Event nodes describing changes in flame, smoke, hot spots, gas, and acoustics are created based on fused input features, and a topological temporal coupling graph structure is generated based on physical causal relationships and temporal dependencies. The topological temporally coupled graph structure is input into the graph neural network to perform recursive inference, generating fire incident status identification results and incident development trend prediction results. Based on the results, fire incident risk output data is generated.

2. The intelligent analysis method for fire incidents based on multimodal fusion according to claim 1, characterized in that, The generation of the multimodal input sequence specifically includes: Video image data, infrared thermal imaging data, smoke concentration data, temperature change data, combustible gas concentration data, and acoustic shock wave data are acquired according to the preset sampling frequency. The acquired sampling data is stored in chronological order according to the corresponding acquisition time. Based on the data collection time, various types of data are indexed and marked with a time index; The six types of data at the same sampling time are combined according to the time index, and a multimodal data unit is generated for each sampling time. The multimodal data units formed in chronological order are arranged sequentially to construct a multimodal input sequence.

3. The intelligent analysis method for fire incidents based on multimodal fusion according to claim 1, characterized in that, The generation of the multimodal preprocessing sequence specifically includes: The multimodal input sequence is initially screened, and noise reduction processing is performed on the six types of data in the sequence to remove interference information generated by the acquisition device, thereby generating a noise-reduced multimodal dataset. The continuity of the noise-reduced dataset is checked one by one to identify data gaps caused by acquisition interruptions or abnormal readings. For data gaps, while maintaining the chronological order, the corresponding interpolation content is added to generate a data sequence without interruption on a unified time reference axis; Various types of data are reorganized according to a unified time reference axis to generate structured multimodal time segments; All corresponding multimodal time segments are integrated according to their time sequence to generate a multimodal preprocessed sequence.

4. The intelligent analysis method for fire incidents based on multimodal fusion according to claim 1, characterized in that, The generation of the calibration feature sequence specifically includes: Dynamic features of video image data, infrared thermal imaging data, smoke concentration data, temperature change data, combustible gas concentration data and acoustic shock wave data over time are extracted from the multimodal preprocessing sequence. The corresponding flame brightness change rate, smoke diffusion gradient, heat conduction delay, and gas concentration change slope are calculated based on the dynamic characteristics of each mode. A modal change rate function is constructed based on the flame brightness change rate, smoke diffusion gradient, heat conduction delay, and gas concentration change slope. A nonlinear time calibration window is generated based on the modal change rate function to calibrate the time position of video image data, infrared thermal imaging data, smoke concentration data, temperature change data, combustible gas concentration data, and acoustic shock wave data. The six types of data calibrated by nonlinear time are arranged in the order of calibration time, so that the six types of modes form a consistent time series structure under a unified time reference; Under the calibration time reference, the features of each mode at the same time position are merged to generate a calibration feature unit that includes video image features, infrared thermal imaging features, smoke concentration features, temperature change features, combustible gas concentration features and acoustic shock wave features. Generate a calibration feature sequence by arranging all calibration feature units in the order of calibration time.

5. The intelligent analysis method for fire incidents based on multimodal fusion according to claim 1, characterized in that, The generation of the modal degradation determination result specifically includes: Under each calibration time index, video image features, infrared thermal imaging features, smoke concentration features, temperature change features, combustible gas concentration features, and acoustic shock wave features are extracted item by item from the calibration feature sequence, and the six types of features are recorded as corresponding modal feature values. For each calibration time index, perform signal fluctuation detection operation on the six modal feature values ​​recorded to generate signal stability detection values ​​for the six modes under that time index; For each calibration time index, the structure retention test operation is performed on the six types of modal characteristic values ​​recorded, and the test results are recorded as the spatial structure retention test values ​​of the six types of modes under that time index; For each calibration time index, perform noise intensity detection on the six modal characteristic values ​​recorded, and record the detection results as the noise level detection values ​​of the six modalities under that time index; Under the same calibration time index, the signal stability test values, spatial structure retention test values, and noise level test values ​​corresponding to the six modes are combined into a test data group in a fixed order; The detection and judgment operations are performed on each item of the detection data set to generate degradation mark values ​​for the six modes respectively. The generated degradation mark values ​​are recorded as video image mode degradation mark, infrared thermal imaging mode degradation mark, smoke mode degradation mark, temperature change mode degradation mark, combustible gas concentration mode degradation mark and acoustic shock wave mode degradation mark under the time index. The six types of modal degradation markers generated under each calibration time index are recorded sequentially in chronological order to generate modal degradation judgment results.

6. The intelligent analysis method for fire incidents based on multimodal fusion according to claim 1, characterized in that, The generation of the fused input features specifically includes: Based on the modal degradation determination results, the features corresponding to the degraded and non-degraded modes are extracted one by one according to the calibration time index and recorded as the degraded feature set and the non-degraded feature set, respectively. The degenerate modal features under each time index in the degenerate feature set are read item by item to obtain the degenerate modal features; Perform inverse feature deduction operation on the degenerate mode features of each time index to generate the recovery features under the corresponding time index; The recovery features are recorded one by one according to the calibration time index, and accumulated in chronological order to generate a recovery feature sequence; The set of non-degraded features is read one by one according to the calibration time index and accumulated and recorded in chronological order as a sequence of non-degraded features; Under each calibration time index, the recovery feature at the corresponding time position in the recovery feature sequence is read, and the non-degraded modal feature at the corresponding time position in the non-degraded feature sequence is read. The two types of features are then concatenated in a fixed order to generate multimodal concatenated features. The multimodal splicing features are written one by one according to the calibration time index to generate a splicing feature sequence arranged in chronological order; The concatenated feature sequence is written sequentially from the first to the last into the fusion feature record structure, and all the concatenated features after writing are used as the fusion input features.

7. The intelligent analysis method for fire incidents based on multimodal fusion according to claim 1, characterized in that, The generation of the topological temporal coupling graph structure specifically includes: Flame change features, smoke change features, hot spot change features, gas change features, and acoustic change features are extracted from the fused input features according to the calibration time index. Under each calibration time index, the flame change characteristics are written to the flame event node, the smoke change characteristics are written to the smoke event node, the hot spot change characteristics are written to the hot spot event node, the gas change characteristics are written to the gas event node, and the acoustic change characteristics are written to the acoustic event node. Under the same calibration time index, the five types of event nodes are connected to each other in a preset node order to generate parallel time-series edges; Between adjacent calibration time indices, flame event nodes, smoke event nodes, hot spot event nodes, gas event nodes, and acoustic event nodes of the same type are connected to each other in ascending time order to generate time-progressive edges; Under each calibration time index, the corresponding event nodes are connected sequentially according to the physical causal order of flame change, smoke change, hot spot change, gas change and acoustic change to generate cross-modal causal edges; The temporal edges, time-progressive edges, and cross-modal causal edges are written into the graph structure record cells in the order of calibration time index to generate a topological temporal coupled graph structure.

8. The intelligent analysis method for fire incidents based on multimodal fusion according to claim 1, characterized in that, The recursive reasoning specifically includes: Flame event node features, smoke event node features, hot spot event node features, gas event node features, and acoustic event node features are extracted from the topological time-series coupled graph structure according to the calibration time index. Node data loading operations are performed on the extracted node features to generate node input data. Extract temporal edge connections, time-progressive edge connections, and cross-modal causal edge connections from the topological temporal coupling graph structure in the order of records. Perform edge data loading operations on the extracted connections to generate edge input data. In graph neural networks, node aggregation processing is performed on node input data and edge input data. Neighborhood feature aggregation is performed on each event node to generate the first-stage node aggregation feature. The first-stage node aggregation features are transformed by performing feature transformation processing, which transforms the aggregation features of each event node to generate node transformation features. In the recursive inference processing area, the node transformation features are recursively calculated according to the calibration time index order, and the node transformation features of each time index are iteratively updated to generate a recursive node feature sequence. Under the same calibration time index, recursive features of flame event nodes, smoke event nodes, hot spot event nodes, gas event nodes, and acoustic event nodes are extracted from the recursive node feature sequence, respectively. Vector concatenation is performed on the five types of recursive features to generate a state vector. For each calibration time index, perform state output processing on the state vector, parse the fields in the state vector into an event category field and an event severity field, and generate fire event status identification results; Trend calculation processing is performed on the state vectors generated under multiple continuous time indices, and sequential operations are performed on the state vector sequence to generate the prediction result of the development trend of fire incidents.

9. The intelligent analysis method for fire incidents based on multimodal fusion according to claim 1, characterized in that, The generation of the fire incident risk output data specifically includes: Under each calibration time index, the event category field and the event severity field are extracted from the fire event status identification results. Field acquisition processing is performed on the two extracted fields to generate a status field set. Under multiple continuous calibration time indices, the event development trend field is extracted from the fire event development trend prediction results, and the extracted trend field is processed to generate a trend field sequence. Under the same calibration time index, perform field combination processing on the event category field and event severity field in the status field set, and extract the event development trend field under the corresponding time index from the trend field sequence, and combine the three types of fields to generate a risk basic field group; Perform field formatting on the risk base field group, and generate the formatted field content into a risk structure field group; In the risk structure field group, perform event category field positioning on the event category field, event severity field positioning on the event severity field, and event trend field positioning on the event development trend field. Generate risk field unit from the positioned field set. Under each calibration time index, risk entry construction processing is performed on the risk field unit, and the risk field unit is arranged into risk data entries. Under multiple calibration time indices, sequential splicing is performed on all risk data entries to generate a risk data sequence from the spliced ​​data entries; After the risk data sequence is generated, data integration processing is performed on the risk data sequence, and the integrated data is used to generate fire incident risk output data.

Images