Fire passage blockage risk prediction method and system based on space-time feature modeling

Abstract

The present application relates to the technical field of security alarm and risk monitoring, and specifically discloses a fire-fighting passage blockage risk prediction method and system based on space-time feature modeling. The present application generates passage alignment observation data by fusing multi-source perception data, processing time synchronization and coordinate unification. By extracting the time evolution features and spatial correlation features of the occupation event sequence, a space-time feature vector is constructed and input into a risk prediction model for reasoning to generate segmented blockage risk results. According to the risk results and passage importance information, graded early warning instructions are generated, and combined with the characteristics of on-site disposal resources, executable disposal content is output, realizing dynamic risk early warning and efficient resource scheduling. In addition, a closed-loop calibration mechanism is introduced to continuously optimize the model prediction accuracy using actual disposal feedback data, ensuring stable operation in various complex building environments, effectively improving the intelligent level of fire-fighting passage management, and ensuring the smoothness of emergency rescue passages.

Description

Technical Field

[0001] This invention belongs to the field of security alarm and risk monitoring technology, and relates to a method and system for predicting the risk of fire lane blockage based on spatiotemporal feature modeling. Background Technology

[0002] Fire lanes, as dedicated lifelines for rescue vehicles and personnel during emergencies such as fires, are crucial to the success of emergency rescue operations. For the daily management and monitoring of fire lanes, a security alarm system uses various on-site detection devices to collect real-time information on the occupancy status of the lane area. When illegally parked vehicles or accumulated debris are detected in the lane, the system sends an alert signal to the management center, enabling security personnel to promptly clear the debris and ensure the lane's emergency passage capability.

[0003] Existing fire lane monitoring and alarm systems typically rely on conventional detection equipment for status assessment. Common practices include installing cameras above the lane and using image recognition algorithms to detect stationary obstacles, or laying ground magnetic sensors to detect parked vehicles. When the occupancy signal from the detection equipment continuously exceeds a preset time threshold, the monitoring system directly triggers a fixed-level alarm and pushes a notification with lane location information to the operator's receiving terminal.

[0004] Existing technologies have the following shortcomings: 1) They have a single sensing dimension and poor robustness, typically relying on a single video image or specific physical sensor for passive state determination. They lack cross-validation and coordinate unification of multi-source sensing data, making them highly susceptible to false alarms or missed alarms when faced with sudden changes in lighting, object occlusion, or complex environmental interference; 2) They lack spatiotemporal correlation analysis and dynamic prediction capabilities, treating the occupancy of each channel area as an isolated, static event, relying on fixed area ratios or time thresholds for determination. They fail to analyze the spatial spread trend and temporal evolution of congestion between adjacent segments, and cannot achieve static monitoring. 3) The alarm mechanism is rigid and lacks enforceability. When an occupation occurs, it can only send homogeneous warning signals or notifications in batches. It fails to prioritize and schedule the alarm information based on the importance of the passage and the available personnel, equipment and other disposal resources on site, which makes it impossible to transform the alarm information into efficient on-site cleanup operations. 4) It lacks a system self-adaptation and closed-loop calibration mechanism. It relies entirely on manually set static rules and thresholds. It has failed to establish a model self-learning and correction framework based on the actual on-site disposal feedback results. It cannot automatically correct prediction deviations in long-term operation to adapt to the complex and ever-changing building environment. Summary of the Invention

[0005] In view of this, in order to solve the problems mentioned in the background technology, a method and system for predicting the risk of fire lane blockage based on spatiotemporal feature modeling is proposed.

[0006] The objective of this invention can be achieved through the following technical solution: The first aspect of this invention provides a method for predicting the risk of fire lane blockage based on spatiotemporal feature modeling, including: S1, acquiring multi-source sensing data of fire lanes and performing time synchronization and coordinate unification to generate lane-aligned observation data.

[0007] S2. Parse the channel-aligned observation data and perform channel segmentation mapping to generate segmented occupancy observation sequences.

[0008] S3. Identify the start and end changes of occupancy in the segmented occupancy observation sequence and perform event-based aggregation to generate an occupancy event sequence.

[0009] S4. Extract the temporal evolution features of the occupied event sequence and fuse them with the influence relationship of adjacent segments to generate a spatiotemporal feature vector.

[0010] S5. Input the spatiotemporal feature vector into the risk prediction model and perform inference to generate segmented congestion risk results.

[0011] S6. Generate early warning output based on the segmented congestion risk results and send it to the alarm terminal to generate risk warning information.

[0012] The second aspect of the present invention provides a fire lane blockage risk prediction system based on spatiotemporal feature modeling, comprising: a lane alignment observation data generation module, which acquires multi-source sensing data of fire lanes and performs time synchronization and coordinate unification to generate lane alignment observation data.

[0013] The segmented occupancy observation sequence generation module parses the channel-aligned observation data and performs channel segmentation mapping to generate segmented occupancy observation sequences.

[0014] The occupancy event sequence generation module identifies the start and end changes of occupancy in the segmented occupancy observation sequence and performs event-based aggregation to generate an occupancy event sequence.

[0015] The spatiotemporal feature vector generation module extracts the temporal evolution features of the occupied event sequence and fuses them with the influence relationship of adjacent segments to generate spatiotemporal feature vectors.

[0016] The segmented congestion risk result generation module inputs spatiotemporal feature vectors into the risk prediction model and performs inference to generate segmented congestion risk results.

[0017] The risk warning information generation module generates warning outputs based on the segmented congestion risk results and sends them to the alarm terminal, thus generating risk warning information.

[0018] Compared with the prior art, the embodiments of the present invention have at least the following advantages or beneficial effects: (1) The present invention overcomes the limitations of conventional detection equipment in complex environments by fusing multi-source sensing data and generating channel-aligned observation data through time synchronization and coordinate unification. By utilizing the cross-verification of visual observation and physical sensor status, combined with the reliability measurement of data missingness and conflict degree, the actual occupancy situation on site is objectively restored, thereby providing a solid data foundation for subsequent risk prediction and avoiding false alarms or missed alarms due to errors in the underlying data.

[0019] (2) This invention extracts the temporal evolution characteristics of the occupancy event sequence and constructs a spatiotemporal feature vector by combining the influence relationship between adjacent segments, fully considering the continuous evolution law of channel congestion in time and its spread and transmission effect in space. By inputting multi-dimensional spatiotemporal correlation information into the risk prediction model, the system dynamically evaluates the congestion probability and uncertainty of each channel segment, realizing the leap from static state monitoring to dynamic risk early warning, enabling security management personnel to grasp potential channel paralysis crises in advance.

[0020] (3) This invention generates graded early warning instructions based on the segmented congestion risk results and channel importance information, and further outputs enhanced early warning instructions by combining the characteristics of available personnel and equipment on site. The mechanism that deeply binds risk assessment with actual rescue resource scheduling not only sends simple warning signals to the alarm terminal, but also simultaneously provides executable operation content including disposal path suggestions, so that on-site personnel can quickly carry out obstacle clearing work according to the plan after receiving the alarm, thus shortening the emergency response cycle.

[0021] (4) This invention introduces a closed-loop calibration mechanism for the risk prediction model, which constructs a calibration sample set by collecting feedback data such as actual disposal completion time and disposal method. The model mapping weights and alarm triggering conditions are continuously updated using the actual on-site fading observation results, so that the prediction system can adapt to the changes in the operating rules under different building environments, and continuously correct the prediction deviation during long-term operation, ensuring the reliable operation of the alarm system in various complex scenarios. Attached Figure Description

[0022] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0023] Figure 1 This is a schematic diagram of the method steps of the present invention.

[0024] Figure 2This is a schematic diagram of the system structure connection of the present invention. Detailed Implementation

[0025] 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.

[0026] Please see Figure 1 The first aspect of the present invention provides a method for predicting the risk of fire lane blockage based on spatiotemporal feature modeling, including: S1, acquiring multi-source sensing data of fire lanes and performing time synchronization and coordinate unification to generate lane-aligned observation data.

[0027] In a specific embodiment of the present invention, acquiring multi-source sensing data of fire lanes and performing time synchronization and coordinate unification to generate lane alignment observation data includes: acquiring image streams from camera equipment and extracting target observation information of the lane area to generate visual occupancy observations.

[0028] Acquire status data from geomagnetic sensors, access control systems, or vehicle gates, extract traffic change information, and generate traffic occupancy observations.

[0029] By fusing visual occupancy observations and traffic occupancy observations and performing time alignment and coordinate unification, multi-source aligned observations are generated and directly output as channel aligned observation data.

[0030] Specifically, the system acquires multi-source sensing data of fire lanes and performs time synchronization and coordinate unification to generate lane alignment observation data. Fire lanes are dedicated roads for fire trucks and personnel. Multi-source sensing data refers to on-site information from different types of detection equipment. Time synchronization adjusts the timestamps recorded by different devices to a unified reference, while coordinate unification maps the two-dimensional or three-dimensional spatial positions of different detection devices to the same reference frame. Lane alignment observation data is the comprehensive detection result after time and space alignment. Camera equipment, as optical instruments capturing on-site images, continuously outputs a sequence of continuously captured video frames to form an image stream. The lane area defines the pixel range occupied by the fire lane in the image, and target observation information reflects the position and outline data of objects appearing in the image. Visual occupancy observation reflects the situation where the lane is occupied by objects, as determined through image analysis. After receiving the image stream, the system analyzes the object outline through a multi-layer image feature extractor to identify target observation information. Visual occupancy observation is calculated based on the proportion of target observation information in the lane area, and the calculation relationship satisfies the formula:

[0031]

[0032] in the formula Represents visual occupancy observation, This represents the number of pixels corresponding to the target observation information, with the unit being pixels. Represents the total number of pixels in the channel region, measured in pixels. Represents the visual confidence weight, dimensionless. Based on 500 test records under different lighting conditions, Set to 0.85.

[0033] It should be noted that the specific underlying execution method for achieving "time synchronization" is as follows: First, the system uses the Network Time Protocol (NTP) to forcibly align the host clock of the camera device with the hardware clock of edge IoT gateways such as geomagnetic sensors and access control systems, ensuring that the absolute time reference error of all original data sources is less than 5 milliseconds. Second, since the camera device outputs a high-frequency periodic image stream, while the geomagnetic / access control system outputs a low-frequency asynchronous event stream, the system introduces a "timestamp sliding window buffer mechanism" for dynamic alignment. The system uses the image frame timestamp of the camera device as the high-frequency sampling axis and sets a time tolerance window (e.g., 200 milliseconds before and after). When a "visual occupancy observation" of a certain frame is extracted, the system searches for data reported by the physical detection device within this time tolerance window; if it exists, the corresponding "access occupancy observation" is extracted; if the physical device does not report new data within this window, the system executes a "zero-order hold" strategy, that is, directly inherits and reuses the steady-state value of the last valid report from the physical device. By combining this high-frequency spindle with zero-order hold mechanism, the time discontinuity problem caused by inconsistent sampling rates of heterogeneous sensors is solved, ensuring reliable fusion of multi-source data.

[0034] Geomagnetic sensors, access control systems, or vehicle gates are physical devices installed on the ground or at entrances / exits to detect the passage of vehicles or people. Their status data records the switching outputs or changes in magnetic field strength. Passage change information is derived from this status data, reflecting the dynamic record of objects entering or leaving the passageway. Passage occupancy observation reflects the occupancy status of the passageway based on the status of the physical devices. The system collects status data, analyzes the numerical fluctuation patterns of the status data to extract passage change information, and calculates passage occupancy observation based on this information. The calculation relationship satisfies the following formula:

[0035]

[0036] in the formula Represents traffic occupancy observation, Representing the Traffic change information during round sampling is represented by a value of 1 when an object passes by and a value of 0 when no object passes by. It is dimensionless. This represents the duration of the state, measured in seconds. This represents the time conversion factor, with the unit being the reciprocal of a second. Represents the total number of sampling rounds within the time window, dimensionless. Based on 200 actual vehicle traffic measurement records, The time was set to the reciprocal of 0.1 seconds to eliminate the influence of the time dimension. The multi-source aligned observations fused visual and transit observations, forming a combination within the same spatiotemporal dimension.

[0037] The system matches visual occupancy observations and traffic occupancy observations at the same timestamp, and performs weighted calculations based on spatial deviations during coordinate unification to obtain multi-source aligned observations. These multi-source aligned observations are directly output as channel aligned observation data. The calculation relationship satisfies the formula:

[0038]

[0039] in the formula Represents multi-source aligned observations, Represents visual weight, dimensionless. Represents the pass weight, dimensionless. Represents the spatial deviation distance during the coordinate unification process, measured in meters. The representative reference is the normalized distance, measured in meters. Based on 1000 historical fusion records, Set to 0.6, Set to 0.4. Based on the standard width specification for fire lanes, It is set to 4.0 meters.

[0040] It should be noted that the spatial deviation distance is obtained during the coordinate unification process as follows: Since the camera equipment acquires two-dimensional pixel coordinates, while physical devices such as geomagnetic sensors and access control systems acquire fixed spatial coordinates from the real physical world, they exist in different coordinate systems. Therefore, obtaining the spatial deviation distance specifically includes the following steps: First, using pre-calibrated intrinsic and extrinsic parameter matrices of the camera equipment (including the camera's installation height, pitch angle, etc.), the pixel center point coordinates of the target vehicle / object extracted from visual occupancy observations are converted into two-dimensional or three-dimensional plane coordinates on the ground in the real physical world through an inverse perspective transformation algorithm; second, the system retrieves the pre-recorded absolute geographic coordinates (such as unified latitude and longitude or local relative coordinates XY of the park) of the geomagnetic sensor or access control gate; finally, the Euclidean distance between the visually transformed coordinates and the absolute coordinates of the physical device is calculated using the spatial geometric distance formula. This calculation result is the spatial deviation distance. This step, through rigorous mathematical spatial mapping, aligns sensor data from different dimensions to the same physical reference plane, providing a metric for the reliable fusion of multi-source data.

[0041] For example, in a fire lane monitoring scenario, the system acquires multi-source sensing data and performs time synchronization and coordinate unification to generate lane-aligned observation data. The system acquires the image stream from the camera equipment and extracts target observation information for the lane area. At this point, the number of pixels corresponding to the target observation information is 15,000, the total number of pixels in the lane area is 100,000, the visual confidence weight is 0.85, and the visual occupancy observation is calculated to be 0.1275 according to the formula. The system acquires the status data of geomagnetic sensors, access control systems, or vehicle gates and extracts passage change information. Within a time window of 10 sampling rounds, only one round has an object passing through; in this case, the passage change information value is 1, and the corresponding state duration is 5 seconds. In other rounds, the passage change information value is 0, the time conversion coefficient is 0.1, and the passage occupancy observation is calculated to be 0.5 according to the formula. The system integrates the visual occupancy observation and the passage occupancy observation and performs time alignment and coordinate unification. The visual weight is 0.6, the passage weight is 0.4, the spatial deviation distance during coordinate unification is 0.8 meters, and the reference normalization distance is 4.0 meters. Based on the formula, the multi-source aligned observation is calculated to be 0.2402. The system stores and outputs the calculated multi-source aligned observation as channel aligned observation data.

[0042] S2. Parse the channel-aligned observation data and perform channel segmentation mapping to generate segmented occupancy observation sequences.

[0043] In a specific embodiment of the present invention, parsing the channel alignment observation data and performing channel segment mapping to generate a segmented occupancy observation sequence includes: acquiring the channel structure data of the fire lane and constructing the segmented topology relationship to generate the channel segment topology.

[0044] The channel-aligned observation data is mapped to the channel segment topology and segment assignment is determined to generate segment-assigned observations.

[0045] The segmented observations are continuously organized and sorted by time to form a segmented status, generating a segmented occupancy observation sequence.

[0046] Specifically, the system parses and aligns observation data of fire lanes and performs segmented mapping to generate segmented occupancy observation sequences. The system acquires the structural data of fire lanes and constructs segmented topological relationships, generating a segmented topology. The structural data refers to the spatial geometric information describing the physical boundaries, length, and width of the fire lane. The segmented topology reflects the spatial adjacency and connectivity between different sub-regions after dividing the complete lane into different sub-regions according to a fixed length. The segmented topology is a spatial network structure composed of nodes in different sub-regions and the lines connecting them. The system calculates the segmented topology relationships based on the structural data, and the calculated relationships satisfy the formula:

[0047]

[0048] in the formula Representing the Segmentation and the first The segmented topological relationships between segments are dimensionless. This represents the standard width of the channel recorded in the channel structure data, measured in meters. Representing the Segment center point and the first The actual physical distance between the center points of the segments, in meters. Represents the connectivity attribute coefficient; direct adjacency is valued at 1, non-adjacent adjacency is valued at 0, and it is dimensionless. Based on 300 building fire lane design codes, It is set to 4.0 meters.

[0049] The system maps channel-aligned observation data to channel segmentation topology and performs segmentation attribution determination, generating segmentation-attributed observations. Channel-aligned observation data is the data generated in step S1, containing multi-source aligned observations and their spatial coordinates. Channel segmentation mapping is the process of comparing and matching the spatial coordinates of the observation data with the boundaries of each segment in the channel segmentation topology. Segmentation attribution determination is the logical judgment process that determines which specific observation data belongs to a corresponding segment. Segmentation-attributed observations are the observation values ​​whose specific segment locations are clearly defined. The system calculates the segmentation-attributed observations based on the channel segmentation mapping results, and the calculation relationship satisfies the formula:

[0050]

[0051] in the formula Representing the Segmented segmental attribution observations, dimensionless. Represents multi-source aligned observations in channel-aligned observation data, dimensionless. The coordinate position of the channel-aligned observation data is related to the first... The spatial distance between the center points of the segments, measured in meters. The reference distance, measured in meters, represents the determination of segment affiliation. Based on 500 channel segment mapping test records, It is set to 2.0 meters.

[0052] It should be noted that the coordinate positions of the channel-aligned observation data are related to the first... The spatial distance between the center points of each segment is obtained as follows: When establishing the channel segment topology, the system has already divided the continuous fire lane into several logical sub-grids (i.e., segments) of fixed length based on the channel structure data (such as high-precision vector maps or CAD drawings), and has pre-calculated and fixed the coordinates of the geometric center point of the physical geometric region of each segment (i.e., the first segment). (Segment center point coordinates). When generating channel-aligned observation data, this data already carries the global physical coordinates after coordinate unification in step S1. Therefore, the process of obtaining spatial distance is as follows: the system extracts the global physical coordinates of the current channel-aligned observation data and queries the database to obtain the target's first... The coordinates of the center point of each segment are used, and then the Euclidean distance between these two points is directly calculated. This spatial distance... It directly reflects the physical proximity between the location of the occupancy event and the core area of ​​each channel segment, and is the core geometric input for realizing the allocation of occupancy status by distance attenuation.

[0053] The system continuously organizes the segmented observations and forms segmented states ordered by time, generating a segmented occupancy observation sequence. Continuous organization involves a smoothing process that removes isolated anomalous fluctuations on the time axis and fills in transiently missing data. The time-ordered segmented states are the occupancy status of each segment arranged in chronological order after smoothing. The segmented occupancy observation sequence is time-series data composed of time-ordered segmented states within a continuous time step. The system calculates the time-ordered segmented states using exponential smoothing, and the calculation relationship satisfies the formula:

[0054]

[0055] in the formula Representing the The segmentation state at the current time step, ordered by time, is dimensionless. Representing the The segmented observations, after continuous processing at the current time step, are dimensionless. Representing the The segmented state, ordered by time in the preceding time step, is dimensionless. Represents the time smoothing coefficient, dimensionless. Based on the fluctuation analysis of 400 historical time series, Set to 0.7. The system combines the time-ordered segmented states within consecutive time steps to output a segmented occupancy observation sequence.

[0056] For example, in a fire lane monitoring scenario, the system parses the lane alignment observation data and performs lane segmentation mapping, ultimately generating a segmented occupancy observation sequence. The system acquires the lane structure data and constructs the segmented topology. At this point, the standard lane width recorded in the lane structure data is 4.0 meters, and the... Segment center point and the first The actual physical distance between the center points of the segments is 5.0 meters. The corresponding segments are directly adjacent, resulting in a connectivity coefficient of 1. Based on the formula, the segment topology is calculated to be 0.8. The system generates the channel segment topology accordingly. The system maps the channel aligned observation data to the channel segment topology and determines segment affiliation. At this point, the multi-source aligned observation in the channel aligned observation data is 0.2402, and the coordinate position of the channel aligned observation data is related to the first... The spatial distance between the segment center points is 1.0 meter, and the reference distance for segment assignment determination is 2.0 meters. Based on the formula, the segment assignment observation is calculated to be 0.1457. The system continuously organizes the segment assignment observations and forms a time-sorted segment status. At the current time step, the segment assignment observation after continuous organization is 0.1457, the time-sorted segment status of the previous time step is 0.1000, and the time smoothing coefficient is 0.7. Based on the formula, the time-sorted segment status of the current time step is calculated to be 0.1320. The system combines the time-sorted segment statuses within consecutive time steps to generate a segment occupancy observation sequence.

[0057] S3. Identify the start and end changes of occupancy in the segmented occupancy observation sequence and perform event-based aggregation to generate an occupancy event sequence.

[0058] In a specific embodiment of the present invention, identifying the start and end changes of occupancy in the segmented occupancy observation sequence and performing event-based aggregation to generate an occupancy event sequence includes: detecting state transitions in the segmented occupancy observation sequence and marking the start and end of occupancy, and generating occupancy boundary markers.

[0059] Based on the occupancy boundary marker, continuous occupancy segments are merged and intermittent segments are connected to generate candidate occupancy events.

[0060] The program determines the persistence and decay of candidate occupancy events and outputs the event attributes to generate an occupancy event sequence.

[0061] Specifically, the system identifies and aggregates the start and end changes of occupancy in the segmented occupancy observation sequence to generate an occupancy event sequence. The segmented occupancy observation sequence, generated in step S2, reflects the occupancy status data of each segment of the channel over time. The start and end changes of occupancy refer to the dynamic process of segmented status values ​​crossing preset judgment criteria. Event aggregation combines discrete occupancy segments on the time axis into a logical whole with a complete time span. The final aggregated data set containing various independent occupancy events is the occupancy event sequence. The system detects state transitions in the segmented occupancy observation sequence and marks the start and end of occupancy, generating occupancy boundary markers. State transitions are manifested as the change of observed values ​​from below the judgment criteria to above the judgment criteria or vice versa. The time node when the state value changes from low to high and crosses the judgment criteria is defined as the start of occupancy, and the time node when the state value changes from high to low and crosses the judgment criteria is defined as the end of occupancy. The data combination recording the start and end times of occupancy constitutes the occupancy boundary marker. The system calculates the occupancy boundary marker based on the state transition logic, and the calculation relationship satisfies the formula:

[0062]

[0063] in the formula Representing the The segment's state transition value at the current time step is dimensionless. This represents a step function that outputs 1 when the input value is greater than or equal to zero, and 0 otherwise. It is dimensionless. Representing the The segmentation state at the current time step, ordered by time, is dimensionless. This represents the occupancy threshold, which is dimensionless. Representing the The segmented state, ordered by time in the preceding time step, is dimensionless. When When the value is 1, the system records the current time step as the start of occupation; when... When the value is negative 1, the system records the current time step as the end of occupancy. Based on 600 measured channel occupancy data points, Set to 0.12.

[0064] The system merges consecutive occupancy segments based on occupancy boundary markers and performs connection determination on intermittent segments to generate candidate occupancy events. A consecutive occupancy segment is a time interval consisting of the start and end of adjacent occupancy periods, while an intermittent segment is a time interval without occupancy between adjacent consecutive occupancy segments. Connection determination is completed through a logical process that judges whether adjacent consecutive occupancy segments belong to the same event. The event set initially aggregated after connection determination is the candidate occupancy event. The system calculates the connection determination index of adjacent segments, and the calculation relationship satisfies the formula:

[0065]

[0066] in the formula Representing the Item continuously occupied fragment and the first The connectivity index between consecutively occupied segments is dimensionless. Representing the The start time of occupancy of a consecutive segment, measured in seconds. Representing the The end time of occupancy of a continuously occupied segment, measured in seconds. This represents the tolerance interval time, measured in seconds. When the connection determination index threshold is greater than or equal to the threshold value, the system will merge the corresponding consecutively occupied segments. Based on 300 temporary transport activity records, Set to 30 seconds.

[0067] It should be noted that the connection determination index threshold is set to... The physical meaning and basis for their values ​​are as follows: in the negative exponential decay formula In, if and only if the actual interval time Exactly equal to the system tolerance interval time When the value of the exponential function is... And the natural constant reciprocal The approximate value is (or approximately 36.8%). In the fields of engineering physics and signal processing, the natural decay constant... (36.8%) is generally recognized as the critical time constant boundary for the decay of a system's state "memory" or "inertia." In this invention, it represents the critical point at which the influence of a previous occupancy event on the subsequent spatial state decays. If the calculated connectivity determination index... Mathematically, this is strictly equivalent to the actual interval being less than or equal to the set maximum tolerable interval. (For example, 30 seconds). In a physical scenario, this means that a car moves from its original position and is detected again in the same area within 30 seconds. This brief interval is typical of "moving or reversing" or "returning after a short period of yielding," where the "event inertia" of its spatial occupancy has not yet completely dissipated. Therefore, the system uses this natural threshold value... Using this as a boundary, these high-frequency intermittent segments are forcibly merged into a single continuous "occupancy event," thereby effectively filtering out fragmented "pseudo-state switching" caused by slow vehicle movement or temporary obstruction, and significantly improving the anti-interference capability and accuracy of subsequent congestion duration statistics and evolution feature extraction.

[0068] The system assesses candidate occupancy events based on their persistence and fading probability, outputting event attributes and generating an occupancy event sequence. Persistence measures the total time span of the candidate occupancy events, while fading probability assesses the likelihood of the event ending spontaneously in the future. The combined characteristics of event duration and fading probability constitute the event attributes. The system calculates event attribute scores, with the calculation relationship satisfying the formula:

[0069]

[0070] in the formula The score represents the attribute score of the event; it is dimensionless. Represents persistent weights, dimensionless. The total duration of the candidate occupancy events after merging, measured in seconds. Represents the maximum tolerable duration, measured in seconds. Represents the degeneracy weight, dimensionless. Represents the probability of decay, dimensionless. Based on 400 records of illegal parking in fire lanes, Set to 300 seconds. Based on 200 incident hazard assessments, Set to 0.6, The value is set to 0.4. The system will arrange the candidate occupancy events with event attribute scores in chronological order to generate an occupancy event sequence.

[0071] For example, in a fire lane monitoring scenario, the system identifies the start and end changes of occupancy in a segmented occupancy observation sequence and performs event aggregation to ultimately generate an occupancy event sequence. The system detects state transitions in the segmented occupancy observation sequence and marks the start and end of occupancy. At this point, the [number missing]th [event missing]... The segmentation state sorted by time at the current time step is 0.1320, the segmentation state sorted by time at the previous time step is 0.1000, and the occupancy determination threshold is 0.12. According to the formula, subtracting the threshold from the current time step results in a step function output of 1, while subtracting the threshold from the previous time step results in a step function output of 0. The state transition value is 1, and the system marks the current time step as the start of occupancy, generating an occupancy boundary marker. The system merges consecutive occupancy segments and performs connection determination on intermittent segments based on the occupancy boundary marker. At this point, the... The occupancy of the continuously occupied segment begins at the 150th second. The occupancy end time of the consecutive occupancy segment is 135 seconds, and the tolerance interval is 30 seconds. The connection determination index calculated using the formula is 0.6065. Since this value is greater than 0.368, the system merges the corresponding consecutive occupancy segments to generate candidate occupancy events. The system performs persistence and decay determination on the candidate occupancy events and outputs event attributes. At this point, the total duration of the merged candidate occupancy events is 150 seconds, the maximum tolerance duration is 300 seconds, the persistence weight is 0.6, and the decay weight is 0.4. Based on the formula, the decay probability is 0.6065, and the event attribute score is 0.4574. The system adds events containing the corresponding event attribute scores to the set, generating an occupancy event sequence.

[0072] S4. Extract the temporal evolution features of the occupied event sequence and fuse them with the influence relationship of adjacent segments to generate a spatiotemporal feature vector.

[0073] In a specific embodiment of the present invention, the temporal evolution features of the occupancy event sequence are extracted and fused with the influence relationship of adjacent segments to generate a spatiotemporal feature vector, including: calculating the duration, repetition frequency and intermittent pattern of the occupancy event sequence and forming time statistics to generate temporal evolution features.

[0074] Based on the channel segmentation topology, the occupancy transmission intensity and diffusion direction of adjacent segments are calculated and spatial correlation quantities are formed, generating spatial correlation features.

[0075] By integrating temporal evolution features and spatial correlation features and performing normalization, a spatiotemporal feature vector is generated.

[0076] Specifically, the temporal evolution features of the occupancy event sequence are extracted and fused with the influence relationships of adjacent segments to generate a spatiotemporal feature vector. The occupancy event sequence is the dataset generated in step S3, containing event attribute scores and time spans. The temporal evolution features reflect the development pattern of channel occupancy status over time. The influence relationships of adjacent segments reflect the interaction of occupancy status between different spatial regions. The spatiotemporal feature vector is a comprehensive data structure formed by combining temporal and spatial dimension features. The system calculates the duration, repetition frequency, and intermittent pattern of the occupancy event sequence and forms time statistics, generating temporal evolution features. The duration records the total time occupied by the event within the statistical time window. The repetition frequency counts the total number of times the event occurs within the same time window. The intermittent pattern reflects the average interval between two adjacent events. The time statistics are the numerical results of the above three time dimension indicators. The system calculates the temporal evolution features based on the time statistics, and the calculation relationship satisfies the formula:

[0077]

[0078] in the formula Represents the characteristics of temporal evolution, dimensionless. Represents duration, measured in seconds. Represents a statistical time window, measured in seconds. Represents the frequency of repetition, measured in units of "times". Represents the highest reference frequency, with the dimension being "times". This represents the average interval time corresponding to the intermittent mode, measured in seconds. This represents the reference interval time, measured in seconds. Represents duration weight, dimensionless. Represents the repetition frequency weight, dimensionless. Represents the weight of the intermittent pattern, dimensionless. Based on the time series analysis of 300 historical channel occupancy data, Set it to 0.5, Set to 0.3, Set it to 0.2, Set to 3600 seconds, Set to 20 times, Set to 300 seconds.

[0079] The system calculates the occupancy transmission intensity and diffusion direction of adjacent segments based on the channel segment topology, forming spatial correlation quantities and generating spatial correlation features. The channel segment topology is the data generated in step S2 reflecting the spatial connectivity structure of each sub-region of the channel. Adjacent segments are sub-regions directly connected to the target segment in the channel segment topology. Occupancy transmission intensity measures the severity of the spread of occupancy status from adjacent segments to the target segment. Diffusion direction reflects the spatial angular relationship of occupancy status spreading from adjacent segments to the target segment. Spatial correlation quantity is a value calculated by combining occupancy transmission intensity and diffusion direction. Spatial correlation features are characteristic values ​​reflecting the degree of mutual spatial influence of channel occupancy. The system calculates spatial correlation features based on the spatial correlation quantity, and the calculation relationship satisfies the formula:

[0080]

[0081] in the formula Representing the Segmented spatial correlation characteristics, dimensionless Representing the Segmentation and the first The segmented topological relationships between segments are dimensionless. Representing the Segmented occupancy transfer intensity, dimensionless. This represents the included angle corresponding to the direction of diffusion, with the dimension of radians. is the diffusion direction coefficient, which is dimensionless.

[0082] It should be noted that further explanation is needed regarding the core parameters in the above spatial correlation feature calculation formula. , and Specific acquisition methods, physical meanings, and their technical role in prediction models: 1. Regarding "the first Segmented occupancy transfer strength Detailed explanation of " It is not a fixed constant, but rather represents the adjacent first... The dynamic quantitative indicators of the severity of congestion within a segment and its potential to "overflow" to surrounding areas. Acquisition method: The system extracts the... The "event attribute score" or "segment status sorted by time" of each segment at the current time step is divided by a preset local saturation limit value for normalization, resulting in a... to The dimensionless floating-point number between these segments, where the typical local saturation limit value is 0.40. Physical meaning: This parameter directly reflects the relationship between adjacent segments. The volume of internal obstacles (such as illegally parked cars or piles of debris). The closer the value is to... This indicates adjacent segments. The physical space is already extremely saturated. Any new occupancy will have nowhere to go, which will inevitably create a strong space compression effect and easily "overflow" outwards, affecting the surrounding target segments.

[0083] 2. Regarding the "angle corresponding to the diffusion direction" Detailed explanation of " This is a geometric parameter, measured in radians, used to characterize the "direction of spread" within a narrow fire lane. It is obtained as follows: First, the system extracts the main central axis of the fire lane based on lane structure data (such as a high-precision map), defining it as the "lane reference vector" (usually consistent with the designated entry and exit direction for fire trucks); second, the system connects adjacent segments... Geometric center point and target segment The geometric center point is used to generate a "spatial topological connecting vector"; finally, by calculating the angle between these two spatial vectors on the horizontal two-dimensional projection plane, the result can be obtained. Physical meaning: From a geometric and topological perspective, it precisely describes adjacent segments. Segmentation with target The spatial relationship is whether it is along the direction of the passage, against the direction of the passage, or at a right-angle turn.

[0084] 3. Regarding the "diffusion direction coefficient" Detailed explanation of " This is a spatial weight adjustment operator that uses trigonometric functions to nonlinearly map the aforementioned angle. Its value is based on the "vector projection effect" in physics. In narrow and directional fire lanes, the spread of obstacles typically occurs along the lane's axis. When segmented... Completely positive towards the target segmentation Extend (i.e., along the main axis of the channel), angle ,at this time This means that the spillover risk of adjacent segments will... The signal is transmitted to the target segment without loss. When the fire lane is at an "L" or "T" bend, the segment... With segmentation They are arranged in diagonal or right-angled adjacent patterns (e.g., at an angle). or ),at this time This aligns with actual physical laws: at corners, the spread of obstacles is hindered and constrained by physical walls, causing them to segment towards the target. The "impact" of diffusion decays geometrically. Synergistic effect: By introducing... The system breaks away from the crude model of relying solely on "distance" to determine influence in traditional models, and gives spatial correlation features a true "sense of direction". This enables the model to accurately distinguish between different blockage evolution trends of "downstream spread" and "angle attenuation", improving the accuracy of the model in predicting the blockage risk of fire lanes in complex shapes (such as bends and intersections).

[0085] The system integrates temporal evolution features and spatial correlation features and performs normalization processing to generate a spatiotemporal feature vector. Normalization is a mathematical operation that maps feature values ​​of different dimensions to a unified interval. The system calculates the feature elements of the spatiotemporal feature vector based on the normalization processing logic, and the calculation relationship satisfies the formula:

[0086]

[0087] in the formula The feature elements representing the spatiotemporal eigenvectors are dimensionless. Represents the weight of time features, dimensionless. Represents spatial feature weights, dimensionless. Represents a normalized reference constant, dimensionless. Based on a fusion test of 500 spatiotemporal features, Set to 0.6, Set to 0.4, Set to 1.0. The system will combine the calculated feature elements to generate a spatiotemporal feature vector.

[0088] For example, in a fire lane monitoring scenario, the system extracts the temporal evolution features of the occupancy event sequence and fuses them with the influence relationships of adjacent segments to ultimately generate a spatiotemporal feature vector. The system calculates the duration, repetition frequency, and intermittent pattern of the occupancy event sequence and forms temporal statistics. In this case, the duration is 1800 seconds, the statistical time window is 3600 seconds, the repetition frequency is 10 times, the maximum reference frequency is 20 times, the average interval time corresponding to the intermittent pattern is 300 seconds, and the reference intermittent time is 300 seconds. The weight of duration is 0.5, the weight of repetition frequency is 0.3, and the weight of intermittent pattern is 0.2. The calculated temporal evolution feature is 0.4736. Based on the channel segment topology, the system calculates the occupancy transmission intensity and diffusion direction of adjacent segments and forms a spatial correlation quantity. At this point... Segmentation and the first The segmentation topology between segments is 0.8, the first segment... The segmented occupancy transmission strength is 0.5, the included angle corresponding to the diffusion direction is 0 radians, and the diffusion direction coefficient is set to 1. Based on the formula, the spatial correlation feature is calculated to be 0.4. The system integrates the temporal evolution feature and the spatial correlation feature and performs normalization processing. At this point, the temporal feature weight is 0.6, the spatial feature weight is 0.4, and the normalization reference constant is 1.0. Based on the formula, the feature element of the spatiotemporal feature vector is calculated to be 0.4442. The system incorporates this feature element into the comprehensive data structure to generate the spatiotemporal feature vector.

[0089] S5. Input the spatiotemporal feature vector into the risk prediction model and perform inference to generate segmented congestion risk results.

[0090] In a specific embodiment of the present invention, the spatiotemporal feature vector is input into the risk prediction model and inference is performed to generate segmented congestion risk results, including: assessing the data missingness, occlusion degree or conflict degree of channel-aligned observation data and forming a reliability metric, and generating observation confidence.

[0091] The observation confidence level and the spatiotemporal feature vector are weighted and combined to suppress the influence of low-reliability observations, thus generating a confidence-weighted feature vector.

[0092] Input the confidence-weighted feature vector into the risk prediction model and output the risk score and uncertainty measure to generate segmented congestion risk results.

[0093] Specifically, the spatiotemporal feature vector is input into the risk prediction model and inference is performed to generate segmented congestion risk results. The spatiotemporal feature vector is a comprehensive data structure combining time and space dimensions generated in step S4. The risk prediction model is an algorithm program that analyzes historical patterns through a multi-layer data feature mapper and outputs prediction results. The segmented congestion risk results are comprehensive evaluation data reflecting the probability of congestion in a specific channel area and the reliability of prediction. The system evaluates the data missingness, occlusion degree, or conflict degree of the channel alignment observation data and forms a reliability metric, generating observation confidence. The channel alignment observation data is the data fused with multi-source sensing information generated in step S1. Data missingness reflects the proportion of valid data not returned by the detection device on time. Occlusion degree reflects the proportion of pixels in the visual image where the target is occluded by other objects. Conflict degree measures the numerical difference between the output results of the visual sensor and the physical sensor. The reliability metric is a basic evaluation value derived from the above three anomaly indicators. The system calculates the reliability metric based on each anomaly indicator, and the calculation relationship satisfies the formula:

[0094]

[0095] in the formula Represents a reliability metric, dimensionless. This represents data that is missing or dimensionless. Represents the degree of occlusion, dimensionless. Represents the degree of conflict, dimensionless. Represents missing weights, dimensionless. Represents occlusion weight, dimensionless. Represents conflict weights, dimensionless. Based on 300 multi-source sensor fault test records, Set to 0.4, Set to 0.3, Set to 0.3. Observation confidence is an indicator that reflects the reliability of the data after inverse mapping.

[0096] The system calculates the observation confidence level based on reliability metrics, and the calculation relationship satisfies the formula:

[0097]

[0098] in the formula Representing observation confidence, this value is dimensionless. The system weights and combines the observation confidence with the spatiotemporal feature vector to suppress the influence of low-reliability observations, generating a confidence-weighted feature vector. This confidence-weighted feature vector is the feature data adjusted for confidence decay. The system adjusts the feature elements of the spatiotemporal feature vector based on the observation confidence, and the calculation relationship satisfies the formula:

[0099]

[0100] in the formula The feature elements representing the confidence-weighted eigenvectors are dimensionless. The feature elements representing the spatiotemporal feature vectors are dimensionless.

[0101] The system inputs a confidence-weighted feature vector into the risk prediction model and outputs a risk score and an uncertainty measure, generating segmented congestion risk results. The risk score represents the probability of substantial congestion occurring in a channel segment. The uncertainty measure reflects the dispersion of the model's prediction results. The system calculates the risk score using the mapping weights of the risk prediction model, and the calculation relationship satisfies the formula:

[0102]

[0103] in the formula Represents a risk score, dimensionless. The mapping weights, representing the risk prediction model, are dimensionless. Based on the model training records of 500 historical congestion events, [the following will be used]. The value is set to 2.5. The system calculates the uncertainty measure based on the risk score, and the calculation relationship satisfies the formula:

[0104]

[0105] in the formula This represents a dimensionless measure of uncertainty. The system combines risk scores with this uncertainty measure to generate segmented congestion risk results.

[0106] It should be noted that the "risk prediction model" in this embodiment is constructed using a deep feedforward neural network architecture. The number of nodes in the input layer of this model strictly corresponds to the dimension of the "confidence-weighted feature vector" generated in step S4, used to receive normalized features of temporal evolution and spatial correlation. The model contains at least two hidden layers, and the nodes in the hidden layers are fully connected through the aforementioned "mapping weights" and bias terms, and the ReLU function is uniformly used as the activation function to solve the gradient vanishing problem in deep networks, while accelerating the fitting of nonlinear spatiotemporal features. The output layer of the model consists of a single node, the key of which is the use of the Sigmoid function as the terminal activation operator. The Sigmoid function can force the high-dimensional feature values ​​transmitted from the hidden layer to be mapped and compressed into a continuous interval of (0, 1). The value generated by this output node is the aforementioned "risk score", which physically represents the probability of a specific channel segment experiencing substantial congestion within a set time window in the future. In the subsequent closed-loop calibration step, "updating the mapping weights" refers to calculating the cross-entropy loss using the actual labels and the Sigmoid output value, and then updating the weight parameters of each layer of the network in reverse using the Adam optimization algorithm.

[0107] For example, in a fire lane monitoring scenario, the system inputs a spatiotemporal feature vector into a risk prediction model and performs inference, ultimately generating segmented blockage risk results. The system evaluates the data missingness, occlusion, or conflict degree of the aligned observation data for the lane and forms a reliability metric. Here, data missing is 0.1, occlusion is 0.2, and conflict is 0.15. The weights for missing data are 0.4, occlusion is 0.3, and conflict is 0.3. The reliability metric is calculated to be 0.145. The system calculates the observation confidence level to be 0.855 based on the reliability metric. The system then weights the observation confidence level with the spatiotemporal feature vector. Here, the feature element of the spatiotemporal feature vector is 0.4442, the observation confidence level is 0.855, and the feature element of the confidence-weighted feature vector is calculated to be 0.3798. The system inputs the confidence-weighted feature vector into the risk prediction model and outputs a risk score and uncertainty metric. Here, the mapping weight of the risk prediction model is 2.5, and the risk score is calculated to be 0.7210. The system calculates an uncertainty measure based on the risk score, and the calculated uncertainty measure is 0.5920 according to the formula. The system combines the calculated risk score and uncertainty measure to generate segmented congestion risk results.

[0108] S6. Generate early warning output based on the segmented congestion risk results and send it to the alarm terminal to generate risk warning information.

[0109] In a specific embodiment of the present invention, a warning output is generated based on the segmented congestion risk results and sent to the alarm terminal to generate risk warning information, including: fusing the segmented congestion risk results with channel importance information to form a risk priority and generating a segmented warning priority.

[0110] Based on the segmented early warning priority, a tiered alarm command is generated and associated with the channel location identifier to generate a tiered early warning command.

[0111] In a specific embodiment of the present invention, a graded alarm instruction is generated based on the segmented warning priority and associated with the channel location identifier. The generation of the graded warning instruction includes: obtaining the status of disposal resources and extracting information on available personnel and available equipment, and generating disposal resource characteristics.

[0112] The system matches the characteristics of the resources to be disposed of with the priority of the segmented early warning system and generates disposal path suggestions, thus generating disposal suggestion information.

[0113] The system combines handling suggestions with tiered alarm instructions and outputs executable handling content to generate tiered early warning instructions.

[0114] Send tiered early warning commands to the alarm terminal and record the sending status to generate risk warning information.

[0115] Specifically, based on the segmented congestion risk results, an early warning output is generated and sent to the alarm terminal, generating risk warning information. The segmented congestion risk result is the data generated in step S5, including risk scores and uncertainty measures. The early warning output is the warning signal issued by the system. The alarm terminal is the receiving device held by security or management personnel. The risk warning information is the complete alarm data that is finally recorded and displayed. The system integrates the segmented congestion risk results with the channel importance information to form a risk priority, generating segmented early warning priorities. The channel importance information reflects the criticality of a specific channel in the overall fire and rescue network. The risk priority is a ranking index derived by comprehensively considering the congestion probability and the channel's status. The segmented early warning priority is the final priority value of the corresponding channel segment. The system calculates the segmented early warning priority based on the risk score and channel importance information, and the calculation relationship satisfies the formula:

[0116]

[0117] in the formula Represents segmented early warning priority, dimensionless. This represents the risk score in the segmented congestion risk results, and is dimensionless. Represents a dimensionless measure of uncertainty in the outcome of segmented congestion risk. Represents the importance of access routes, dimensionless. Based on 200 fire access route classification standards, main fire access routes are classified... Set to 1.5.

[0118] The system generates tiered alarm commands based on segmented warning priorities and associates them with channel location identifiers. Tiered alarm commands are action requirements corresponding to the urgency level, categorized by priority. Channel location identifiers are spatial coordinates or area numbers indicating the specific physical location of a channel. Tiered warning commands are comprehensive commands combining action requirements and specific locations. The system acquires the status of disposal resources and extracts information on available personnel and equipment, generating disposal resource characteristics. Disposal resource status reflects the current logistical support available for handling congestion. Available personnel refers to the number of manpower currently idle and capable of participating in clearing. Available equipment information refers to the number of machines available for moving obstacles. Disposal resource characteristics are quantitative assessments of comprehensive manpower and material resources. The system calculates disposal resource characteristics based on available personnel and equipment information, with the calculation relationship satisfying the formula:

[0119]

[0120] in the formula Represents the characteristics of resource disposal, dimensionless. This represents the quantity of available personnel that can be extracted, measured in human units. This represents the quantity of available device information that can be extracted, measured in units. Representative conversion factor, with the dimension of the reciprocal of human. This represents the equipment conversion coefficient, with the dimension being the reciprocal of the unit. Based on 100 property scheduling records, Set the countdown to 0.5 people, and... Set to the countdown of 0.8 units.

[0121] The system matches resource characteristics with segmented early warning priorities and generates suggested disposal routes, thus generating disposal suggestion information. The suggested disposal routes are recommended routes planned by the system to the congestion location. The disposal suggestion information is a comprehensive guidance plan including route planning and resource allocation. The system calculates the path urgency score in the disposal suggestion information, and the calculation relationship satisfies the formula:

[0122]

[0123] in the formula The urgency score represents the path's priority; it is dimensionless. This represents a compensation constant to prevent the denominator from being zero. It is dimensionless and set to 0.01.

[0124] The system combines suggested actions with tiered alarm instructions and outputs executable actions, generating tiered warning instructions. Executable actions consist of specific steps that security personnel can directly follow. The system sends the tiered warning instructions to the alarm terminal and records the sending status, generating risk warning information. The sending status record indicates whether the instruction was successfully delivered and the delivery time. The system then sends the final tiered warning instruction, combining it with the sending status to form complete risk warning information.

[0125] For example, in a fire lane monitoring scenario, the system generates early warning outputs based on segmented blockage risk results and sends them to the alarm terminal, ultimately generating risk warning information. The system integrates the segmented blockage risk results with the lane importance information to form a risk priority. At this point, the risk score in the segmented blockage risk results is 0.7210, the uncertainty measure is 0.5920, and the lane importance information is 1.5. The segmented warning priority is calculated to be 1.7217 based on the formula. The system generates tiered alarm commands based on the segmented warning priority and associates them with lane location identifiers. The system obtains the status of disposal resources and extracts information on available personnel and equipment. At this point, the number of available personnel is 4, the number of available equipment is 2, the personnel conversion coefficient is the reciprocal of 0.5 people, and the equipment conversion coefficient is the reciprocal of 0.8 units. The disposal resource characteristic is calculated to be 3.6 based on the formula. The system matches the disposal resource characteristics with the segmented warning priority and generates disposal path suggestions. At this point, the segmented warning priority is 1.7217, the compensation constant is 0.01, and the path urgency score in the handling suggestion information is calculated to be 0.4769 based on the formula. The system combines the handling suggestion information with the tiered alarm instructions and outputs executable handling content, generating tiered warning instructions. The system sends the tiered warning instructions to the alarm terminal and records the sending status. At this point, the sending status is "successfully delivered." The system combines the above instructions and status data to generate risk warning information.

[0126] In a specific embodiment of the present invention, the method further includes closed-loop calibration of the risk prediction model, wherein the closed-loop calibration includes: obtaining the handling results corresponding to the risk warning information and extracting the handling completion time and handling method, and generating handling feedback data.

[0127] Obtain the segmented occupancy observation sequence after treatment and extract the occupancy decay change to generate post-treatment observation results.

[0128] The feedback data from the treatment process is correlated with the observation results after the treatment to form training samples, and a calibration sample set is generated.

[0129] Update the risk prediction model or update the triggering conditions of the graded alarm command using the calibration sample set to generate the calibrated prediction configuration.

[0130] Specifically, a closed-loop calibration is performed on the risk prediction model. Closed-loop calibration is an optimization process that feeds actual results back to the system to correct prediction biases. The system acquires the handling results corresponding to risk warning information and extracts the handling completion time and handling method, thereby generating handling feedback data. The risk warning information inherits the complete alarm data generated in the previous steps, including alarm commands and sending status. The handling result represents the actual on-site feedback information after security personnel execute the warning command. The handling completion time records the time span from receiving the alarm to the clearing of obstacles on-site, while the handling method reflects the specific clearing methods used on-site. The combined quantified feedback values ​​of handling completion time and handling method constitute the handling feedback data. The system calculates the handling feedback data based on the handling completion time and handling method, and the calculation relationship satisfies the formula:

[0131]

[0132] in the formula Representative feedback data, dimensionless. This represents the time it takes to complete the process, measured in seconds. This represents the reference processing time, measured in seconds. The efficiency coefficient corresponding to the disposal method is dimensionless. Represents time-feedback weights, dimensionless. The representative method provides feedback weights, which are dimensionless. Based on 200 standard fire drill records, Set to 600 seconds. Based on the analysis of 300 historical feedback points, Set to 0.6, Set to 0.4.

[0133] It should be noted that, in further detail, the efficiency coefficients corresponding to the disposal methods in the above-mentioned disposal feedback data calculation formulas require further explanation. The acquisition mechanism, specific mapping rules, and physical technical significance:

[0134] 1. Acquisition Mechanism and Physical Significance: In real fire lane management scenarios, the clearing methods adopted by security personnel after arriving at the scene vary, which directly reflects the "physical stubbornness" and "resource consumption" of the lane blockage. This system transforms unstructured qualitative on-site handling actions (such as "moving things" or "towing") into quantitative evaluation indicators that can be used for machine learning. The acquisition mechanism is as follows: After completing on-site cleanup, security personnel use a handheld alarm terminal or mobile app to select the primary handling method used in the event from a preset drop-down menu. The system backend then automatically looks up the corresponding method and assigns it to the appropriate efficiency coefficient based on a pre-set "handling method-efficiency coefficient mapping matrix". Numerical value.

[0135] 2. Specific Quantitative Mapping Rules and Typical Values: The logic behind the value of this coefficient is as follows: the higher the degree of mechanization and the stronger the enforcement required by the disposal methods, the greater the difficulty in clearing the blockage and the deeper the potential fire-fighting delay hazard. Therefore, the corresponding... The larger the assigned value, the better. In a typical embodiment, the specific mapping rules are set as follows: "Vehicle owner cooperates to leave" or "Simple persuasion to leave": In such cases, the vehicle owner is usually inside or nearby, and verbal intervention can quickly clear the way, consuming very few resources. A typical value is set to 0.3. "Manual handling / cleaning": When dealing with unattended non-motorized vehicles, cardboard boxes, debris, or trash cans, security personnel need to expend physical effort to move them, resulting in a certain time cost for cleanup. The typical value is set to 0.8. "Mechanical Vehicle Relocation / Forced Towing": When faced with illegally parked vehicles with locked doors and unreachable owners, it is necessary to use vehicle relocation equipment or even professional towing equipment. These situations are extremely time-consuming, require large equipment, and pose a severe threat to fire truck access. The typical value is set to the highest level, 1.5.

[0136] 3. The technical role in the closed-loop model is achieved by introducing the aforementioned gradient-clear... The system provides high-dimensional "penalty / reward" labels for the closed-loop calibration of the risk prediction model. For example, if the predicted risk in a certain segment was initially low, but feedback data shows that "mandatory towing" was ultimately used... The calculated feedback data will surge dramatically. This high feedback label will force the backpropagation algorithm to strongly modify the model mapping weights of this segment, allowing the model to "learn from its mistakes" and immediately identify as a high-risk blockage that is difficult to clear when encountering similar spatiotemporal characteristics in the future, thus triggering a high-level warning in advance. This enables the alarm system to intelligently adapt to complex on-site characteristics and learn deeply.

[0137] The system acquires the segmented occupancy observation sequence after processing and extracts the occupancy decay change to generate the post-processing observation results. The segmented occupancy observation sequence continuously reflects the occupancy status data of each segment of the channel over time. The occupancy decay change reflects the rate at which the channel occupancy status value decreases over time after processing. The quantified value after the occupancy decay change is the post-processing observation result. The system calculates the post-processing observation results based on the occupancy decay change, and the calculation relationship satisfies the formula:

[0138]

[0139] in the formula This represents the observation results after treatment, with the dimension of the reciprocal of a second. Represents the segmented states prior to treatment, ordered by time; dimensionless. Represents the segmented states after treatment, ordered by time; dimensionless. The time window for observing the decline of the representative state is measured in seconds. Based on 100 monitoring records of the decline process, Set to 60 seconds.

[0140] The system correlates the feedback data from the treatment process with the post-treatment observations to form training samples, and then aggregates these to generate a calibration sample set. The training samples pair the input features from the prediction phase with the actual feedback results to form a specific data structure; multiple training samples are combined to constitute the calibration sample set. The system calculates the target label values ​​of the training samples, and the calculation relationship satisfies the formula:

[0141]

[0142] in the formula The target label value represents the training sample; it is dimensionless. This represents the time dimension elimination coefficient, with the dimension in seconds. Represents the weight of the feedback data, dimensionless. Represents the weight of the observed results, dimensionless. Based on 500 model calibration tests, Set it to 0.5, Set it to 0.5, Set to 1 second to eliminate the influence of the time dimension.

[0143] The system uses a calibration sample set to update the risk prediction model or the triggering conditions of tiered alarm commands, ultimately generating a calibrated prediction configuration. The triggering conditions of tiered alarm commands act as threshold parameters determining when and what level of alarm to send. The updated model weights and the set of triggering threshold parameters together form the calibrated prediction configuration. The system updates the mapping weights of the risk prediction model through error backpropagation logic, and the calculation relationship satisfies the formula:

[0144]

[0145] in the formula The mapping weights, representing the updated risk prediction model, are dimensionless. The mapping weights representing the original risk prediction model are dimensionless. The learning rate parameter is dimensionless. Represents a risk score, dimensionless. The feature elements representing the confidence-weighted eigenvectors are dimensionless. Based on model optimization experience, Set it to 0.01. The system will save the updated parameters and generate the calibrated prediction configuration.

[0146] For example, in a fire escape monitoring scenario, the system performs closed-loop calibration on the risk prediction model. The system acquires the handling results corresponding to the risk warning information and extracts the handling completion time and handling method. At this point, the handling completion time is 400 seconds, the reference handling time is 600 seconds, the efficiency coefficient for manual handling is 0.8, the time feedback weight is 0.6, and the method feedback weight is 0.4. The handling feedback data is calculated to be 1.22 based on the formula. The system acquires the segmented occupancy observation sequence after handling and extracts the occupancy decay change. At this point, the segmented state before handling, sorted by time, is 0.1320, and the segmented state after handling, sorted by time, is 0.0120. The state decay observation time window is 60 seconds. The post-handling observation result is calculated to be the reciprocal of 0.002 seconds based on the formula. The system correlates the handling feedback data with the post-handling observation results to form training samples. At this point, the feedback data weight is 0.5, the observation result weight is 0.5, and the time dimension elimination coefficient is 1 second. The target label value of the training sample is calculated to be 0.611 based on the formula. The system adds training samples containing the target label value to the set, generating a calibration sample set. The system uses this calibration sample set to update the risk prediction model or the triggering conditions of the tiered alarm command. At this point, the original risk prediction model has a mapping weight of 2.5, a learning rate of 0.01, a risk score of 0.7210, and 0.3798 feature elements in the confidence-weighted feature vector. The updated risk prediction model's mapping weight is calculated to be 2.4996 according to the formula. The system saves this weight parameter and finally generates the calibrated prediction configuration.

[0147] Reference Figure 2 The second aspect of the present invention provides a fire lane blockage risk prediction system based on spatiotemporal feature modeling, comprising: a lane alignment observation data generation module, a segmented occupancy observation sequence generation module, an occupancy event sequence generation module, a spatiotemporal feature vector generation module, a segmented blockage risk result generation module, and a risk warning information generation module.

[0148] The channel alignment observation data generation module is connected to the segmented occupancy observation sequence generation module, the segmented occupancy observation sequence generation module is connected to the occupancy event sequence generation module, the occupancy event sequence generation module is connected to the spatiotemporal feature vector generation module, the spatiotemporal feature vector generation module is connected to the segmented congestion risk result generation module, and the segmented congestion risk result generation module is connected to the risk warning information generation module.

[0149] The channel alignment observation data generation module acquires multi-source sensing data of fire lanes and performs time synchronization and coordinate unification to generate channel alignment observation data.

[0150] The segmented occupancy observation sequence generation module parses the channel-aligned observation data and performs channel segmentation mapping to generate segmented occupancy observation sequences.

[0151] The occupancy event sequence generation module identifies the start and end changes of occupancy in the segmented occupancy observation sequence and performs event-based aggregation to generate an occupancy event sequence.

[0152] The spatiotemporal feature vector generation module extracts the temporal evolution features of the occupied event sequence and fuses them with the influence relationship of adjacent segments to generate spatiotemporal feature vectors.

[0153] The segmented congestion risk result generation module inputs spatiotemporal feature vectors into the risk prediction model and performs inference to generate segmented congestion risk results.

[0154] The risk warning information generation module generates warning outputs based on the segmented congestion risk results and sends them to the alarm terminal, thus generating risk warning information.

[0155] The above content is merely an example and illustration of the concept of the present invention. Those skilled in the art can make various modifications or additions to the specific embodiments described, or use similar methods to replace them, as long as they do not deviate from the concept of the invention or exceed the scope defined by the present invention, and all such modifications and additions should fall within the protection scope of the present invention.

Claims

1. A method for predicting the risk of fire lane blockage based on spatiotemporal feature modeling, characterized in that, include: S1. Acquire multi-source sensing data of fire lanes and synchronize and unify time and coordinates to generate lane alignment observation data; S2. Parse the channel-aligned observation data and perform channel segmentation mapping to generate segmented occupancy observation sequences; S3. Identify the start and end changes of occupancy in the segmented occupancy observation sequence and perform event-based aggregation to generate an occupancy event sequence; S4. Extract the temporal evolution features of the occupied event sequence and fuse them with the influence relationship of adjacent segments to generate a spatiotemporal feature vector; S5. Input the spatiotemporal feature vector into the risk prediction model and perform inference to generate segmented congestion risk results; S6. Generate early warning output based on the segmented congestion risk results and send it to the alarm terminal to generate risk warning information.

2. The method for predicting the risk of fire lane blockage based on spatiotemporal feature modeling according to claim 1, characterized in that, The process of acquiring multi-source sensing data of fire lanes and synchronizing and unifying time and coordinates to generate lane-aligned observation data includes: Acquire the image stream from the camera device and extract target observation information for the channel region to generate visual occupancy observations; Acquire status data from geomagnetism, access control, or vehicle gates, extract passage change information, and generate passage occupancy observations; By fusing visual occupancy observations and traffic occupancy observations and performing time alignment and coordinate unification, multi-source aligned observations are generated and directly output as channel aligned observation data.

3. The method for predicting the risk of fire lane blockage based on spatiotemporal feature modeling according to claim 1, characterized in that, The analytical channel aligns the observation data and performs channel segmentation mapping to generate a segmented occupancy observation sequence, including: Obtain the structural data of fire lanes and construct segmented topology relationships to generate segmented topology of the lanes; The channel-aligned observation data is mapped to the channel segment topology and segment assignment is determined to generate segment-assigned observations. The segmented observations are continuously organized and sorted by time to form a segmented status, generating a segmented occupancy observation sequence.

4. The method for predicting the risk of fire lane blockage based on spatiotemporal feature modeling according to claim 1, characterized in that, The process of identifying and aggregating the start and end changes of occupancy in the segmented occupancy observation sequence to generate an occupancy event sequence includes: Detect state transitions in the segmented occupancy observation sequence and mark the start and end of occupancy, generating occupancy boundary markers; Based on the occupancy boundary marker, continuous occupancy segments are merged and the connection of intermittent segments is determined to generate candidate occupancy events; The program determines the persistence and decay of candidate occupancy events and outputs the event attributes to generate an occupancy event sequence.

5. The method for predicting the risk of fire lane blockage based on spatiotemporal feature modeling according to claim 1, characterized in that, The extraction of the temporal evolution features of the occupancy event sequence and their fusion with the influence relationships of adjacent segments to generate a spatiotemporal feature vector include: Calculate the duration, repetition frequency, and intermittent patterns of the occupied event sequence and generate temporal statistics to produce temporal evolution characteristics; Based on the channel segmentation topology, the occupancy transmission intensity and diffusion direction of adjacent segments are calculated and spatial correlation quantities are formed to generate spatial correlation features; By integrating temporal evolution features and spatial correlation features and performing normalization, a spatiotemporal feature vector is generated.

6. The method for predicting the risk of fire lane blockage based on spatiotemporal feature modeling according to claim 1, characterized in that, The input spatiotemporal feature vector is fed into the risk prediction model and inferred to generate segmented congestion risk results, including: Evaluate the data missingness, occlusion, or conflict of channel-aligned observation data and form a reliability metric to generate observation confidence. The observation confidence level and the spatiotemporal feature vector are weighted and combined to suppress the influence of low-reliability observations, thereby generating a confidence-weighted feature vector; Input the confidence-weighted feature vector into the risk prediction model and output the risk score and uncertainty measure to generate segmented congestion risk results.

7. The method for predicting the risk of fire lane blockage based on spatiotemporal feature modeling according to claim 1, characterized in that, The step of generating a warning output based on the segmented congestion risk results and sending it to the alarm terminal, generating risk warning information, includes: The segmented congestion risk results are integrated with channel importance information to form risk priorities, and segmented early warning priorities are generated. Based on the segmented warning priority, generate hierarchical alarm instructions and associate them with channel location identifiers to generate hierarchical warning instructions; Send tiered early warning commands to the alarm terminal and record the sending status to generate risk warning information.

8. The method for predicting the risk of fire lane blockage based on spatiotemporal feature modeling according to claim 7, characterized in that, The step of generating tiered alarm instructions based on segmented warning priorities and associating them with channel location identifiers includes: Acquire the status of disposal resources and extract information on available personnel and equipment to generate disposal resource characteristics; The system matches the characteristics of the resources to be disposed of with the priority of the segmented early warning and generates disposal path suggestions, thus generating disposal suggestion information; The system combines handling suggestions with tiered alarm instructions and outputs executable handling content to generate tiered early warning instructions.

9. The method for predicting the risk of fire lane blockage based on spatiotemporal feature modeling according to claim 8, characterized in that, It also includes closed-loop calibration of the risk prediction model, wherein closed-loop calibration includes: Obtain the handling results corresponding to risk warning information, extract the handling completion time and handling method, and generate handling feedback data; Obtain the segmented occupancy observation sequence after treatment and extract the occupancy decay change to generate post-treatment observation results; The feedback data from the treatment process is correlated with the observation results after the treatment to form training samples, and a calibration sample set is generated. Update the risk prediction model or update the triggering conditions of the graded alarm command using the calibration sample set to generate the calibrated prediction configuration.

10. A fire lane blockage risk prediction system based on spatiotemporal feature modeling, characterized in that, include: The channel alignment observation data generation module acquires multi-source sensing data of fire lanes and performs time synchronization and coordinate unification to generate channel alignment observation data. The segmented occupancy observation sequence generation module parses the channel-aligned observation data and performs channel segmentation mapping to generate segmented occupancy observation sequences; The occupancy event sequence generation module identifies the start and end changes of occupancy in the segmented occupancy observation sequence and performs event-based aggregation to generate an occupancy event sequence; The spatiotemporal feature vector generation module extracts the temporal evolution features of the occupied event sequence and fuses them with the influence relationship of adjacent segments to generate spatiotemporal feature vectors. The segmented congestion risk result generation module inputs spatiotemporal feature vectors into the risk prediction model and performs inference to generate segmented congestion risk results. The risk warning information generation module generates warning outputs based on the segmented congestion risk results and sends them to the alarm terminal, thus generating risk warning information.

Images