A safety supervision method and system for urban intelligent fire fighting

Abstract

The application discloses a kind of safety supervision method and system for urban wisdom fire, it is related to wisdom city public safety supervision technical field;Including: obtaining the record of fire hazard that has been sold, generates and sells the hidden danger anchor point of hazard, including regulatory object identification, original hidden danger space unit, original hidden danger risk consequence label, sale confirmation time and sale before historical abnormal time sequence;Determine the distribution of historical period in the same position supervision range;According to regulatory object identification, obtain the fire supervision data after sale, extract the low-intensity abnormal state set that occurs in the same position supervision range and is related to original hidden danger risk consequence;Calculate weak recurrence rebound index of same period and determine rebound state;In suspected rebound state, select regulatory action based on reinforcement learning model, and update the state of fire hazard that has been sold according to action execution feedback.The application can identify risk rebound trend before hidden danger re-formal fire hazard, improve the authenticity and timeliness of wisdom fire supervision closed loop.

Description

Technical Field

[0001] This invention relates to the field of smart city public safety supervision technology, and in particular to a safety supervision method and system for smart fire protection in cities. Background Technology

[0002] With the increasing density of urban buildings, the number of densely populated places, and multi-functional complexes, urban fire safety supervision targets are characterized by a large number, complex types, and rapidly changing risks. Existing urban smart fire supervision platforms can typically manage key units, buildings, fire protection facilities, inspection records, hazard rectification records, and maintenance records through information technology. They achieve closed-loop management of fire hazards through processes such as hazard reporting, rectification assignment, rectification feedback, review and confirmation, and closure and archiving. Such platforms play a positive role in improving fire supervision efficiency, standardizing hazard handling procedures, and promoting the implementation of rectification by responsible units.

[0003] Existing smart fire safety supervision methods typically treat the removal of a hazard as the end point of a rectification process. Once a fire hazard is marked as removed, the system often stops continuously assessing its subsequent risk status. In actual supervision, hazards such as obstructed evacuation routes, blocked safety exits, short-term failure of fire-fighting facilities, illegal use of electricity, and illegal charging of electric vehicles are recurring and time-sensitive, and may reappear in a low-intensity, short-term, and repeated manner after rectification and removal. Because such anomalies do not reach the level of being re-identified as formal fire hazards, and are easily treated as isolated anomalies or ordinary supervision records, the platform struggles to promptly identify the weak recurrence and rebound trends of removed fire hazards in the same time period, resulting in a distorted supervision problem where the process is closed but the risk recurs. Summary of the Invention

[0004] The purpose of this invention is to address the shortcomings of existing technologies, such as the lack of continuous monitoring of fire hazards that have been delisted, the difficulty in timely identification of the rebound trend of hazards caused by the repeated occurrence of low-intensity anomalies in the same spatial range, with the same risk consequences, and during historically high-incidence periods. Therefore, this invention proposes a safety monitoring method and system for smart fire protection in cities.

[0005] To address the problems existing in the prior art, the present invention adopts the following technical solution:

[0006] A safety supervision method for smart fire protection in cities includes:

[0007] S1. Obtain the records of fire hazards that have been cancelled and generate anchor points for cancelled hazards. The anchor points for cancelled hazards include the regulatory object identifier, the original hazard spatial unit, the original hazard risk consequence label, the cancellation confirmation time, and the historical abnormal time series before cancellation.

[0008] S2. Determine the scope of supervision based on the original spatial unit of the hidden danger, and determine the distribution of historical time periods based on the historical abnormal time series before the cancellation.

[0009] S3. From the time of confirmation of cancellation, obtain the fire supervision data of the supervised object after cancellation based on the identifier of the supervised object, screen the abnormal status that occurs within the same supervision scope and is consistent with the original hidden danger risk consequence label or belongs to the risk consequence category corresponding to the original hidden danger risk consequence label, and extract the set of low intensity abnormal status based on the stable upper bound determined by the stable operation sample after cancellation and the formal hidden danger judgment boundary determined by the fire supervision rules, platform hidden danger judgment rules or manual confirmation sample.

[0010] S4. Calculate the weak recurrence rebound index for the same time period based on the set of low-intensity abnormal states, the scope of supervision at the same time, and the distribution of historical time periods, and determine the rebound status based on the weak recurrence rebound index for the same time period.

[0011] S5. When the rebound status is a suspected rebound status, the weak recurrence rebound index and rebound status in the same time period are used as the input basis for the reinforcement learning model to select regulatory actions.

[0012] S6. Generate regulatory tasks based on regulatory actions, receive action execution feedback, and update the status of closed fire hazards based on the action execution feedback.

[0013] Preferably, the original hazard risk consequence label is used to characterize the fire safety impact caused by the fire hazard that has been eliminated, including: classifying the obstruction of evacuation routes, locked safety exits and closed escape windows as evacuation obstruction risk consequences; classifying insufficient fire hydrant pressure, abnormal sprinkler system and alarm host failure as fire facility failure risk consequences; and classifying electric vehicle charging in homes, private wiring and illegal connection of high-power electrical appliances as illegal electricity use-induced risk consequences.

[0014] Preferably, the scope of supervision is determined based on the original potential hazard spatial unit, including:

[0015] The original hazardous space units will be included in the scope of same-location supervision.

[0016] Identify adjacent spatial units that have a direct fire risk connection with the original hazardous spatial unit;

[0017] The adjacent spatial units shall be included in the scope of co-location supervision;

[0018] The direct fire risk associations include adjacent evacuation routes, connected fire protection facility circuits, connected fire compartments on the same floor, connected electrical circuits, or managed by the same responsible entity.

[0019] Preferably, the distribution of historical time periods is determined based on the historical anomaly time series before the cancellation of the registration number, including:

[0020] Time slots are divided according to the regulatory cycle set by the regulatory platform;

[0021] Statistical analysis of the number of abnormal occurrences of fire hazards that have been removed from the list within each time period prior to their removal;

[0022] The historical occurrence probability for each time slice is determined by the proportion of the number of anomalies occurring in each time slice to the total number of anomalies occurring in all time slices.

[0023] The historical time period distribution is formed by combining each time slice with its corresponding historical occurrence probability.

[0024] Preferably, a set of low-intensity abnormal states is extracted based on the upper bound of stability and the threshold for determining formal hazards, including:

[0025] The upper bound of stability is determined based on stable operating samples after the cancellation of the number;

[0026] The formal hazard determination boundary shall be determined in accordance with fire safety supervision rules, platform hazard determination rules, or manually confirmed samples.

[0027] Determine the intensity of the abnormality in the abnormal state;

[0028] When the abnormal intensity is greater than the upper limit of stability but less than the threshold for formal hazard identification, the abnormal state is regarded as a low-intensity abnormal state.

[0029] Summarize low-intensity abnormal states to form a set of low-intensity abnormal states;

[0030] The abnormal intensity is calculated based on the patrol deduction points, the degree of deviation of facility status, the degree of passage occupancy, the maintenance abnormality level, or the existing abnormality level of the platform, according to the unified abnormal intensity scale preset by the urban smart fire supervision platform; the fire supervision data includes patrol records, maintenance records, rectification and review records, fire facility status records, passage occupancy records, electricity abnormality records, unit self-inspection records, and grid member verification records.

[0031] Preferably, the calculation of the weak recurrence rebound index in the same time period includes:

[0032] The degree of weak anomaly is determined by the position of the anomalous intensity of the low-intensity anomaly state between the upper limit of stability and the boundary for formal hazard assessment.

[0033] The spatial isomorphism is determined based on the degree of correlation between the spatial unit where the low-intensity abnormal state is located and the original hidden danger spatial unit in the fire protection spatial topology.

[0034] The overlap of time periods is determined by the historical probability of occurrence of time slices of low-intensity abnormal states in the historical time period distribution.

[0035] The similarity of risk consequences is determined based on the consistency between the risk consequence label of the low-intensity abnormal state and the original risk consequence label of the hidden danger, or whether the two belong to the same risk consequence category.

[0036] The rebound contribution of a single low-intensity anomaly is determined by multiplying the degree of weak anomaly, spatial co-location, temporal overlap, and similarity of risk consequences. The rebound contributions of each low-intensity anomaly in the set of low-intensity anomalies are then summed to obtain the weak recurrence rebound index in the same time period.

[0037] Preferably, determining the rebound state includes:

[0038] The criteria for determining a rebound are determined based on historical samples of previously identified hidden risks.

[0039] When the weak recurrence rebound index in the same time period meets the rebound judgment conditions, the fire hazards that have been cancelled will be marked as suspected rebound status;

[0040] When the weak rebound index in the same time period does not meet the rebound judgment conditions, the normal cancellation status of the cancelled fire hazards shall be maintained.

[0041] Preferably, the selection of regulatory actions based on a reinforcement learning model includes:

[0042] The reinforcement learning model consists of the following: the weak recurrence rebound index, rebound status, abnormal persistence in the set of low-intensity abnormal statuses, the duration from the confirmation time of cancellation to the current time, the historical rectification feedback of responsible units obtained from the cancelled fire hazard records or fire supervision data, and the current status of supervision resource occupancy determined from the supervision task records of the city's smart fire supervision platform.

[0043] The actions of the reinforcement learning model will include sending self-inspection reminders to responsible units, generating online review tasks, assigning grid members to conduct on-site verification tasks, triggering third-party maintenance review tasks, and submitting suggestions for fire supervision spot checks.

[0044] The reward for the reinforcement learning model is determined based on changes in the rebound index, the elimination of low-intensity anomalies, the results of verification and confirmation, and the consumption of regulatory resources.

[0045] Regulatory actions are determined based on the actions output by the reinforcement learning model.

[0046] Preferably, the status of closed fire hazards is updated based on action execution feedback, including:

[0047] The set of low-intensity abnormal states is updated based on the feedback from the action execution, and the weak recurrence rebound index in the same time period is redefined.

[0048] When the action execution feedback indicates that the low-intensity anomaly has been eliminated, and the re-determined weak recurrence rebound index in the same time period does not meet the rebound judgment condition, the normal cancellation status of the cancelled fire hazard shall be maintained.

[0049] When the action execution feedback indicates that the original hidden danger has re-formed into a formal fire hazard, or the newly determined weak recurrence rebound index in the same time period still meets the rebound judgment conditions, the fire hazard that has been cancelled will be updated from the suspected rebound status to the confirmed rebound status, and a rectification task will be generated.

[0050] When the action execution feedback indicates that the abnormal state is not related to the fire hazard that has been cleared, the corresponding abnormal state is removed from the set of low-intensity abnormal states, and the action execution feedback is used as the reward feedback for the reinforcement learning model.

[0051] Compared with the prior art, the beneficial effects of the present invention are:

[0052] 1. This invention acquires records of closed fire hazards and generates anchor points for these hazards. It binds the regulatory object identifier, the original hazard spatial unit, the original hazard risk consequence label, the closure confirmation time, and the historical abnormal time series before closure. This ensures that closed fire hazards are no longer merely archived as historical records, but become objects for continuous monitoring. Furthermore, this invention determines the corresponding regulatory scope based on the original hazard spatial unit, determines the historical time period distribution based on the historical abnormal time series before closure, and extracts low-intensity abnormal states that occur within the corresponding regulatory scope and are related to the risk consequences of the original hazard from the fire safety monitoring data after closure. This establishes a spatial, temporal, and risk consequence correlation between the abnormal states after closure and the original closed fire hazards, avoiding the existing system's treatment of post-closure anomalies as isolated events. This solves the problems of lack of continuous correlation monitoring for closed fire hazards and difficulty in timely detection of the diminishing effects of rectification.

[0053] 2. This invention further calculates the weak recurrence rebound index of the same time period based on the low-intensity abnormal state set, the same regulatory scope, and the distribution of historical time periods, and determines the rebound status based on this index. This enables the system to identify the cumulative trend of weak anomalies that repeatedly occur in the same spatial range, with the same risk consequences, and in historically high-incidence periods before the hidden danger reaches the level of formal fire hazard judgment again. For suspected rebound status, this invention uses the weak recurrence rebound index of the same time period and the rebound status as the input basis for the reinforcement learning model to select regulatory actions, and updates the status of the closed fire hazard and the feedback of the reinforcement learning model based on the action execution feedback. This realizes the transformation from a one-time process closed loop of discovery-rectification-closure to a dynamic regulatory closed loop of continuous identification after closure-intervention for suspected rebounds-feedback updates. This improves the ability of smart fire supervision to identify the rebound trend of hidden dangers in advance, the rationality of the allocation of regulatory tasks, and the authenticity and timeliness of the closed-loop management of fire hazards. Attached Figure Description

[0054] The accompanying drawings, which are included to provide a further understanding of the invention and form part of this application, illustrate exemplary embodiments of the invention and, together with their description, serve to explain the invention and do not constitute an undue limitation thereof. In the drawings:

[0055] Figure 1 This is a flowchart illustrating a safety supervision method for smart fire protection in cities, provided as an embodiment of the present invention. Detailed Implementation

[0056] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments.

[0057] This embodiment provides a safety supervision method and system for smart fire protection in cities. Its core lies in continuously tracking fire hazards that have been rectified and delisted, and identifying whether the delisted fire hazard shows a weak recurrence or rebound in the same time period after delisting. The weak recurrence or rebound in the same time period refers to the low-intensity abnormal state that repeatedly occurs in the original hazard spatial unit or adjacent spatial units that have a direct fire risk association with the original hazard spatial unit during the historical high-incidence period before delisting. This state is consistent with or belongs to the risk consequence category corresponding to the original hazard risk consequence label.

[0058] In this embodiment, the urban smart fire protection supervision platform is an existing fire safety supervision platform, fire hazard closed-loop management platform, fire facility operation status supervision platform, or a data processing system that includes the above functions. This embodiment does not require the addition of specific hardware equipment as a necessary condition. The fire protection supervision data used comes from the existing hazard ledgers, inspection records, maintenance records, rectification and review records, fire facility status records, passage occupancy records, abnormal electricity use records, unit self-inspection records, and grid member verification records in the urban smart fire protection supervision platform.

[0059] Example 1: As Figure 1 As shown in the figure, this embodiment provides a safety supervision method for smart fire protection in cities, including the following steps;

[0060] S1. Obtain the records of fire hazards that have been cancelled and generate anchor points for cancelled hazards;

[0061] The city's smart fire safety supervision platform acquires records of fire hazards that have been closed; these records refer to fire hazards that have undergone hazard discovery, rectification assignment, rectification feedback, and re-inspection confirmation, and have been marked as closed by the platform.

[0062] For each fire hazard record that has been closed, generate a closed hazard anchor point h;

[0063] In this embodiment, the fire hazards that have been delisted refer to fire hazards in the city's smart fire protection supervision platform that have been rectified, reviewed and confirmed, and marked as delisted. A delisted fire hazard does not mean that the fire risk has permanently disappeared in the subsequent supervision cycle, but rather that the hazard has been closed-loop processed in the current rectification process. This invention uses delisted fire hazards as continuous tracking objects to determine whether there are signs of risk rebound after they are delisted.

[0064] In this embodiment, the hazard anchor point h refers to a continuously tracking data unit established around a hazard that has been delisted. The hazard anchor point h is used to bind the delisted fire hazard to its original hazard spatial unit, original hazard risk consequence label, delisting confirmation time, and historical abnormal time series before delisting, so that the delisted fire hazard can still be used as the basis for subsequent abnormal screening, rebound index calculation and regulatory action selection after delisting.

[0065] The anchor point h for the cancellation of hidden dangers includes the regulatory object identifier, the original hidden danger spatial unit, the original hidden danger risk consequence label, the cancellation confirmation time, and the historical abnormal time series before cancellation.

[0066] The regulatory object identifier is used to uniquely identify the fire safety object being regulated; the regulatory objects are commercial complexes, shopping malls, hospitals, schools, elderly care institutions, warehousing and logistics sites, rental houses, catering establishments, industrial parks, or other objects included in the city's smart fire safety supervision platform; the regulatory object identifier is the unit number, building number, site number, or regulatory object file number in the supervision platform.

[0067] The original hidden danger spatial unit is used to represent the spatial location of the fire hazard that has been delisted before it was delisted;

[0068] In this embodiment, the original hidden danger space unit refers to the location of a fire risk that was discovered, recorded, or confirmed before the fire hazard that has been delisted was delisted; the original hidden danger space unit is not simply a geographical location name, but a spatial object that can participate in the fire risk association judgment, which corresponds to other fire supervision space units divided by evacuation routes, safety exits, fire-fighting facility locations, floors, fire compartments, electrical circuit areas, or platforms; the original hidden danger space unit is used to determine the corresponding supervision scope in subsequent steps;

[0069] The original hazard risk consequence label was used to characterize the potential fire safety impact of fire hazards that have been delisted, rather than simply recording the hazard name;

[0070] In this embodiment, the original hazard risk consequence label refers to a label used to describe the fire safety consequences that may be caused by a fire hazard that has been delisted. The original hazard risk consequence label is different from the hazard name, and the same risk consequence label corresponds to multiple specific hazard names. For example, although the specific manifestations of evacuation route obstruction, safety exit lock, and escape window closure are different, they may all lead to obstruction of evacuation and escape, and are therefore classified as evacuation obstruction risk consequences. By setting the original hazard risk consequence label, the system can identify low-intensity abnormal states that have the same risk consequences or the same risk consequence category as the original hazard after it has been delisted.

[0071] Specifically, obstructed evacuation routes, locked safety exits, and closed escape windows are classified as risks and consequences of obstructed evacuation; insufficient fire hydrant pressure, abnormal sprinkler systems, and malfunctioning alarm control panels are classified as risks and consequences of fire protection facility failure; and charging electric vehicles in homes, illegally connecting electrical wires, and illegally connecting high-power appliances are classified as risks and consequences of illegal electricity use.

[0072] The confirmation time of cancellation is used to determine the start time of subsequent observation; the historical anomaly time series before cancellation is used to record the time information of each anomaly that occurred before the cancellation of the fire hazard, which is then used to determine the distribution of historical time periods.

[0073] S2. Determine the scope of supervision based on the original spatial unit of the hidden danger, and determine the distribution of historical time periods based on the historical abnormal time series before the cancellation.

[0074] The city's smart fire safety supervision platform determines the scope of supervision based on the original spatial unit of the hazard in the hazard anchor point h after it has been delisted. ;

[0075] In this embodiment, the scope of peer supervision This refers to the fire risk associated area defined around the original hazardous spatial unit; this area is not an arbitrarily expanded geographical region, but rather consists of the original hazardous spatial unit and adjacent spatial units that have a direct fire risk association with the original hazardous spatial unit; [This refers to] the scope of supervision at the same location. This is used to limit the screening scope of abnormal states after the closure of the hazard, so that the subsequent rebound identification is focused on the original hazard location and its fire risk-related location, and avoids mistaking the abnormal state in the unrelated space unit as a rebound sign of the original closure of the fire hazard.

[0076] The scope of peer supervision This includes the original hazardous space unit and adjacent space units that have a direct fire risk association with the original hazardous space unit; the direct fire risk association includes adjacent evacuation routes, connected fire protection facility circuits, connected fire compartments on the same floor, connected electrical circuits, or managed by the same responsible entity;

[0077] Specifically, when the original hazardous space unit is an evacuation route, the corresponding regulatory scope... This includes the evacuation route, the safety exit connected to the evacuation route, the entrance to the adjacent stairwell, and adjacent areas on the same evacuation route; when the original hazardous space unit is a fire hydrant location, the corresponding supervisory area... This includes the fire hydrant location, adjacent fire hydrant locations on the same fire water supply branch, and associated fire water pressure monitoring points; when the original hazardous space unit is an electrical circuit area, the same area is included in the monitoring scope. This includes the area of ​​the power circuit and adjacent power areas that are on the same power circuit or managed by the same responsible entity;

[0078] The city's smart fire safety monitoring platform determines the distribution of historical time periods based on the historical anomaly time series prior to the cancellation of hazard anchor points h. ;

[0079] In this embodiment, the historical anomaly time sequence before the cancellation refers to a data sequence formed by the chronological order of the occurrence times of each anomaly before the cancellation confirmation time of the cancelled fire hazard; the historical time period distribution This refers to the distribution of the probability of a closed fire hazard occurring at different time points, obtained from statistical analysis of historical anomalies before its closure; historical time period distribution. This is used to determine whether the low-intensity abnormal state after the hazard has been cleared occurred during the historical high-incidence period of the original hazard.

[0080] In this embodiment, the regulatory period set by the regulatory platform is divided into multiple time slices. The statistics show the number of abnormal occurrences of fire hazards that have been removed from the list within each time period before their removal. The historical time period distribution is determined according to the following formula:

[0081]

[0082] in, The fire hazard has been removed from the list. The probability of occurrence in history, This indicates that the fire hazard has been removed from the list. (The last part, "h", appears to be a fragment and doesn't translate directly. It likely refers to a time period before the hazard was removed.) The number of times it appears in the text This indicates the total number of abnormal occurrences of the now-closed fire hazard within the entire time frame prior to its closure.

[0083] The above formula is a frequency statistics formula, used to characterize the historical distribution of revoked fire hazards across different time periods; historical time period distribution. In subsequent step S4, the overlap in time periods between low-intensity abnormal states and fire hazards that have been cleared is determined.

[0084] S3. Starting from the time of confirmation of cancellation, obtain the fire supervision data of the supervised object after cancellation based on the supervised object identifier, and extract the set of low-intensity abnormal states.

[0085] The city's smart fire safety supervision platform takes the confirmation time of the cancellation of the hazard anchor point h as the starting point and obtains the fire safety supervision data after the corresponding regulated object is cancelled according to the regulated object identifier.

[0086] The fire safety supervision data includes inspection records, maintenance records, rectification and review records, fire protection facility status records, passageway occupancy records, abnormal electricity usage records, unit self-inspection records, and grid member verification records; the city's smart fire safety supervision platform does not perform indiscriminate analysis on all fire safety supervision data, but rather performs targeted screening around the identified hazard anchor point h.

[0087] Specifically, the city's smart fire safety supervision platform filters fire incidents occurring within the same supervisory area from fire safety supervision data. Within the scope of the original hazard risk consequence label, an abnormal state is defined as either consistent with or belonging to the risk consequence category corresponding to the original hazard risk consequence label. For any abnormal state e, the spatial unit of occurrence, time slice of occurrence, risk consequence label, and abnormal intensity of the abnormal state e are determined. ;

[0088] In this embodiment, the abnormal state e refers to a status record identified by the urban smart fire protection supervision platform in the fire protection supervision data after the status has been deactivated, which deviates from the normal fire protection supervision status. Abnormal state e does not necessarily constitute a formal fire hazard; it originates from inspection deductions, facility status deviations, passageway obstruction, maintenance anomaly levels, electricity usage anomaly records, or existing anomaly levels on the platform. Abnormal state e only occurs when it occurs within the scope of the same level of supervision. Only when the condition is consistent with the original hazard risk consequence label or belongs to the risk consequence category corresponding to the original hazard risk consequence label will it be included in the subsequent low-intensity abnormal state judgment.

[0089] Abnormal intensity Used to characterize the degree of deviation of abnormal state e from normal fire supervision state; the abnormal intensity The abnormal intensity is calculated based on the inspection deduction, the degree of deviation of facility status, the degree of passage occupancy, the level of maintenance abnormality, or the existing abnormality level of the platform, according to the unified abnormality intensity scale preset by the urban smart fire protection supervision platform. The unified abnormality intensity scale is used to convert abnormal data from different sources into comparable abnormality intensity values. The unified abnormality intensity scale is determined by the urban smart fire protection supervision platform based on existing business rules, historical samples, and manually confirmed samples.

[0090] In one implementation, the unified anomaly intensity scale is a normalized intensity scale, and the anomaly intensity... The value range is 0 to 1; for patrol deduction type data, the ratio of a single patrol deduction value to the maximum deduction value corresponding to this type of anomaly is used as the anomaly intensity. For facility status deviation data, the deviation between the measured facility status value and the normal status benchmark value is divided by the allowable deviation of the facility to determine the anomaly intensity. For channel occupancy data, the ratio of the occupied channel width to the effective passage width of the channel, or the ratio of the occupied channel area to the effective passage area of ​​the channel, is used as the anomaly intensity. For maintenance anomaly levels or existing anomaly levels on the platform, the ratio of the level number corresponding to the current anomaly level to the highest anomaly level number is used as the anomaly intensity. ;

[0091] When the same abnormal state corresponds to an anomaly intensity from multiple sources, the maximum value among the multiple anomaly intensities is taken as the anomaly intensity of that abnormal state. Through the above methods, fire supervision data from different sources can be converted to the same anomaly intensity scale for comparison, thereby enabling inspection anomalies, facility status anomalies, passageway occupancy anomalies, and maintenance anomalies to be uniformly included in the identification of low-intensity anomaly states.

[0092] The city's smart fire protection monitoring platform determines the upper limit of stability based on stable operation samples after the case is closed. The "stable operating sample after cancellation" refers to a sample that, after the cancellation of a fire hazard, has been confirmed by the responsible unit through self-inspection, online review, on-site verification, or platform operation status confirmation, has not shown any risk of rebound; (Upper limit of stability) This indicates the upper limit of abnormal intensity that is allowed to occur when the fire hazard that has been removed from the list is in a stable state after rectification.

[0093] In one implementation, the urban smart fire protection monitoring platform calculates the anomaly intensity of abnormal states in stable operating samples that are consistent with or belong to the risk consequence category corresponding to the original hazard risk consequence label, thus obtaining a stable sample anomaly intensity set; the maximum anomaly intensity in the stable sample anomaly intensity set is used as the upper bound of stability. ;

[0094] When the number of stable operating samples corresponding to a closed fire hazard is insufficient, the city's smart fire protection supervision platform selects stable samples of closed fire hazards with the same risk consequences from similar supervised objects, calculates the corresponding set of stable sample anomalies, and uses the maximum anomaly intensity in this set of stable sample anomalies as the upper bound of stability. The aforementioned similar regulatory objects refer to regulatory objects that are consistent with the current regulatory objects in terms of regulatory object type, building use, or fire risk consequences category.

[0095] In another implementation, when there are samples that have been manually confirmed to be in a stable state after rectification, the highest anomaly intensity that has been manually confirmed to be in a stable state is used as the upper bound of stability. The upper bound is stabilized through the above methods. The samples are derived from stable operation samples or manually confirmed samples after the fire hazard has been delisted, and can reflect the upper limit of the abnormal intensity of the fire hazard in the stable state after rectification.

[0096] The city's smart fire safety supervision platform determines the formal hazard determination boundaries based on fire safety supervision rules, platform hazard determination rules, or manually confirmed samples. The formal hazard determination threshold Used to indicate that an abnormal state has reached the threshold for re-identifying as a formal fire hazard.

[0097] In one implementation, when the fire safety supervision rules or the platform's hazard assessment rules have already provided quantifiable assessment criteria for a certain type of hazard, the city's smart fire safety supervision platform will use the lowest abnormal intensity corresponding to meeting those criteria as the formal hazard assessment threshold. For example, for channel occupancy anomalies, when the platform's hazard assessment rules stipulate that channel occupancy reaches the formal hazard assessment criteria, the intensity of the channel occupancy anomaly corresponding to those criteria will be used as the threshold for formal hazard assessment. For facility status anomalies, when the platform's hazard assessment rules stipulate that a facility status deviation meets the criteria for formal hazard assessment, the intensity of the facility status anomaly corresponding to that criterion will be used as the threshold for formal hazard assessment. ;

[0098] When fire safety supervision rules or platform hazard assessment rules do not provide directly quantifiable assessment criteria, the city's smart fire safety supervision platform selects samples that have been manually confirmed to have re-formed formal fire hazards in the past, calculates the set of abnormal intensities corresponding to these samples, and uses the minimum abnormal intensity in this set as the threshold for formal hazard assessment. ;

[0099] Through the above methods, the formal hazard assessment boundary is established. Based on fire safety supervision rules, platform hazard assessment rules, or manually confirmed samples, low-intensity abnormal states can be defined as intermediate abnormal states that are higher than stable states but lower than formal hazard assessment states.

[0100] In this embodiment, the stability upper bound This refers to the upper limit of abnormal intensity that a fire hazard h that has been rectified and is now in a stable state is allowed to occur; the formal hazard determination boundary... This refers to the abnormal intensity limit corresponding to when an abnormal state reaches the level required to be re-identified as a formal fire hazard; upper limit of stability. Boundary between formal hazard assessment and Together, we define the intensity range of the low-intensity abnormal state, so that the low-intensity abnormal state is defined as an intermediate state that is higher than the stable state after rectification but lower than the formal fire hazard state.

[0101] When the abnormal intensity of abnormal state e satisfy:

[0102]

[0103] The city's smart fire protection monitoring platform then identifies this abnormal state e as a low-intensity abnormal state.

[0104] In this embodiment, the low-intensity abnormal state refers to an abnormality intensity that has exceeded the upper limit of stability. However, it has not yet reached the threshold for formal hazard assessment. The abnormal state; the low-intensity abnormal state is different from normal fluctuations and also different from the abnormal state that has re-formed a formal fire hazard. It is used to characterize the early risk resurgence that may occur after a fire hazard that has been cleared; multiple low-intensity abnormal states are summarized to form a low-intensity abnormal state set. This is used for subsequent calculation of the weak recurrence rebound index in the same time period;

[0105] This condition indicates that the abnormal state e is higher than the stable state after rectification, but has not yet reached the level of being re-identified as a formal fire hazard. Therefore, a low-intensity abnormal state cannot be directly treated as a new formal fire hazard, but it can indicate that the rectification effect of a fire hazard that has been cleared may be weakened or the risk may rebound.

[0106] The city's smart fire protection monitoring platform aggregates low-intensity abnormal states that meet the above conditions, forming a set of low-intensity abnormal states. .

[0107] S4. Calculate the weak recurrence rebound index for the same time period based on the set of low-intensity abnormal states, the scope of supervision in the same location, and the distribution of historical time periods, and determine the rebound status.

[0108] The city's smart fire safety monitoring platform targets fire hazards that have been resolved, based on a set of low-intensity abnormal states. Scope of supervision at the same level Distribution of historical periods Calculate the weak recurrence index in the same time period ;

[0109] In this embodiment, the weak recurrence rebound index in the same time period This refers to an indicator used to quantify whether a fire hazard h that has been cleared shows signs of risk rebound after its clearance. This indicator is not a generalized score of the overall fire risk of the regulated entity, but rather a rebound identification result for a specific cleared fire hazard, considering the degree of weak anomaly in its low-intensity abnormal state after clearance, spatial co-location, temporal overlap, and similarity of risk consequences. It is identified through the weak recurrence rebound index in the same time period. The system can determine whether fire hazards that have been removed from the list show a weak recurrence trend within the original spatial context, the original historical high-incidence period, and under the same risk consequences.

[0110] For low-intensity anomalous state sets For any low-intensity anomaly state 'e' in the data, the urban smart fire protection monitoring platform first calculates the degree of weak anomaly. ;

[0111] In this embodiment, the degree of weak anomaly , used to represent a low-intensity anomalous state e at a stable upper bound. Boundary between formal hazard assessment and The relative positions between them; the degree of weak anomaly The larger the value, the closer the low-intensity anomaly is to re-emerging as a formal fire hazard; the weaker the anomaly... The smaller the value, the closer the low-intensity abnormal state is to the stable state after rectification;

[0112] Weak anomaly Used to characterize the upper bound of a low-intensity anomalous state e in stability. Boundary between formal hazard assessment and The positions between them are determined according to the following formula:

[0113]

[0114] in, This indicates the degree of weak anomalousness of the low-intensity anomalous state e. This indicates the anomalous intensity of the low-intensity anomalous state e. This indicates the stable upper limit of the number of fire hazards that have been removed from the list. This represents the formal hazard determination boundary for fire hazards that have been removed from the list; this formula is a linear normalization formula used to convert the position of low-intensity abnormal states between the upper limit of stability and the formal hazard determination boundary into dimensionless values.

[0115] The city's smart fire protection monitoring platform determines spatial co-location based on the degree of correlation between the spatial unit where the low-intensity abnormal state 'e' is located and the original hazardous spatial unit within the fire protection spatial topology. ;

[0116] In this embodiment, the spatial isomorphism Used to characterize the fire risk correlation strength between the location of the low-intensity abnormal state e and the original hidden spatial unit; spatial isomorphism It is not simply a matter of physical distance, but is determined based on the fire protection space topology, which includes the relationships of evacuation routes, fire protection facility circuits, fire compartments, electrical circuits, and the management relationships of responsible entities.

[0117] When a low-intensity abnormal state e occurs in the original hazardous spatial unit, its spatial co-location is higher than when it occurs in an adjacent spatial unit; when a low-intensity abnormal state e occurs in a spatial unit that is not directly related to the original hazardous spatial unit in terms of fire risk, this abnormal state is not included in the set of low-intensity abnormal states. The rebound index is calculated.

[0118] In one implementation, the urban smart fire protection supervision platform establishes a fire protection space topology map; the nodes in the fire protection space topology map are spatial units, and the edges indicate that there is a direct fire risk association between two spatial units; the direct fire risk association includes adjacent evacuation routes, connected fire protection facility circuits, connected fire compartments on the same floor, connected electrical circuits, or managed by the same responsible entity.

[0119] If the low-intensity abnormal state e occurs in the original hidden danger spatial unit, then the spatial isotope Take 1; if the low-intensity anomaly e occurs in an adjacent spatial unit directly connected to the original hidden danger spatial unit, then the spatial isomorphism is determined according to the edge type between the adjacent spatial unit and the original hidden danger spatial unit. If a low-intensity abnormal state e occurs within the scope of the same level of supervision. Otherwise, this abnormal state will not be included in the set of low-intensity abnormal states. Calculate the index for weak rebounds in the same time period.

[0120] In one specific implementation, adjacent evacuation routes, connected fire protection circuits, connected fire compartments on the same floor, connected electrical circuits, and managed by the same responsible entity correspond to different association weights. Within the same implementation, the same association weight is used for the same type of hazard. The association weights are determined by the city's smart fire safety supervision platform based on historically resolved hazard samples or manually configured rules. The spatial co-location of adjacent spatial units directly connected to the original hazard spatial unit is also considered. For values ​​greater than 0 and less than 1, the same side type corresponds to the same spatial coherence in the same embodiment; thus, spatial coherence... It can reflect the strength of the fire risk correlation between low-intensity abnormal conditions and the original hidden danger space units;

[0121] The city's smart fire safety monitoring platform analyzes the distribution of low-intensity abnormal conditions (e) across historical time periods. The probability of historical occurrences determines the overlap of time periods. ;

[0122] In this embodiment, the time period overlap This is used to characterize the degree of overlap between the occurrence time of a low-intensity abnormal state e and the historical high-incidence period before the elimination of a fire hazard h; the degree of overlap in time periods. The higher the value, the more likely the low-intensity abnormal state is to occur during the business period when the original hidden danger was formed or recurred, thus more likely to represent a weak recurrence or rebound of the fire hazard that has been cleared.

[0123] If the low-intensity abnormal state e occurs in time slice Then the degree of overlap in time periods Take this time slice Distribution in historical periods The probability of corresponding historical occurrences in the data;

[0124] The city's smart fire safety monitoring platform determines the similarity of risk consequences based on the consistency between the risk consequence label of a low-intensity abnormal state 'e' and the original risk consequence label of a hidden danger, or whether the two belong to the same risk consequence category. ;

[0125] In this embodiment, the similarity of risk consequences This is used to characterize the degree of consistency between low-intensity abnormal state e and closed fire hazard h in terms of fire safety impact; risk consequence similarity. This is to prevent the misidentification of abnormal states that are spatially close but have different risk consequences as a rebound of a previously eliminated fire hazard; for example, if the original hazard's risk consequence was obstruction of evacuation, an abnormal state occurring on the same floor but belonging to the category of equipment maintenance delay should not be directly identified as a rebound of that obstruction of evacuation hazard simply because of spatial proximity.

[0126] When the risk consequences labels of the two are consistent, the similarity of risk consequences is higher than when the two only belong to the same risk consequences category; when the two are inconsistent and do not belong to the same risk consequences category, the abnormal state is not included in the rebound index calculation as a weak recurrence anomaly of the fire hazard that has been cancelled.

[0127] In one implementation, if the risk consequence label of a low-intensity abnormal state e is consistent with the original hazard risk consequence label, then the risk consequence similarity is... Take 1; if the risk consequence label of the low-intensity abnormal state e is inconsistent with the original hidden danger risk consequence label, but the two belong to the same risk consequence category, then the risk consequence similarity is 1. Take similar values ​​for risk consequences of the same type; if the risk consequence label of low intensity abnormal state e is inconsistent with the original risk consequence label of hidden danger, and the two do not belong to the same risk consequence category, then the abnormal state will not be included in the calculation of the weak recurrence rebound index of the same time period as the weak recurrence abnormality of the fire hazard that has been cancelled.

[0128] The similarity value of similar risk consequences is determined based on the hierarchical relationship of risk consequence labels in the urban smart fire protection supervision platform. A similarity value of less than 1 and greater than 0 indicates that although different risk consequence labels are not completely identical, they still have a rebound correlation when belonging to the same risk consequence category. In one implementation, the similarity value of similar risk consequences is determined by the frequency of co-occurrence of different risk consequence labels in historically confirmed rebound samples within the same risk consequence category, or by pre-configuration of the hierarchical relationship of risk consequence labels in the urban smart fire protection supervision platform. Through the above methods, the risk consequence similarity... It can distinguish between completely similar rebounds and weak recurrences of the same type.

[0129] The city's smart fire protection monitoring platform is classified according to the degree of weak anomalies. Spatial isotope Time period overlap Similarity to risk consequences The product of these factors determines the rebound contribution of a single low-intensity anomalous state e, and the product of these factors also determines the rebound contribution of the set of low-intensity anomalous states. The rebound contributions of various low-intensity abnormal states are summed to obtain the weak recurrence rebound index for the same time period. Specifically:

[0130]

[0131] in, This indicates a weak recurrence and rebound index for fire hazards that have been removed from the list during the same time period. This represents the set of low-intensity abnormal states corresponding to the fire hazard that has been removed from the list. This indicates the degree of weak anomalousness of the low-intensity anomalous state e. This indicates the spatial isomorphism between the low-intensity abnormal state e and the fire hazard h that has been cleared. This indicates the degree of overlap in time between the low-intensity abnormal state e and the fire hazard h that has been cleared. This indicates the similarity of risk consequences between a low-intensity abnormal state e and a fire hazard h that has been cleared.

[0132] In the same batch of rebound determinations, the city's smart fire protection supervision platform uses the same length of post-cancellation observation period to calculate the weak recurrence rebound index for the same time period. When different fire hazards that have been rectified have different post-rectification observation periods, the city's smart fire safety monitoring platform will aggregate low-intensity abnormal states. The rebound contribution in the data was normalized according to the observation period length to determine the weak recurrence rebound index in the same time period. Therefore, this avoids the incomparability of weak rebound indices in the same time period due to different observation period lengths.

[0133] The city's smart fire safety monitoring platform determines the criteria for recurrence based on historically closed-off hazard samples. These historically closed-off hazard samples include those that have historically maintained a normal closed-off status and those that have historically recurred upon review. Based on these historically closed-off hazard samples, the city's smart fire safety monitoring platform determines the weak recurrence index for the same time period. Corresponding rebound judgment conditions;

[0134] In one implementation, the urban smart fire protection supervision platform divides historically closed hazard samples into normal closed samples and confirmed rebound samples. The normal closed samples refer to samples that have not rebounded after being closed and subsequently observed or reviewed. The confirmed rebound samples refer to samples that have rebounded after being closed and confirmed by feedback from the responsible unit, online review, on-site verification by grid members, third-party maintenance review, or fire supervision spot checks.

[0135] The city's smart fire protection supervision platform calculates the weak recurrence index of the same time period corresponding to the normal cancellation sample and the confirmed rebound sample respectively; the interval between the highest weak recurrence index of the same time period in the normal cancellation sample and the lowest weak recurrence index of the same time period in the confirmed rebound sample is used as the candidate judgment interval, and the value that can distinguish between the normal cancellation sample and the confirmed rebound sample is selected from the candidate judgment interval as the rebound judgment condition.

[0136] When the weak recurrence index of the same time period corresponding to normal closed-out samples and confirmed rebound samples overlaps, the urban smart fire protection supervision platform will use the value with the fewest misjudgments of normal closed-out samples and confirmed rebound samples as the rebound judgment condition. When the number of confirmed rebound samples is insufficient to form the rebound judgment condition, the urban smart fire protection supervision platform uses manually confirmed suspected rebound samples as the initial sample to determine the initial rebound judgment condition, and updates the rebound judgment condition based on newly added confirmed rebound samples after the accumulation of feedback from subsequent supervision tasks. In this way, the rebound judgment condition is derived from historical closed-out hazard samples, does not rely on fixed empirical thresholds, and can adapt to different supervision objects and different risk consequences.

[0137] When the index shows a weak rebound in the same time period When the rebound judgment criteria are met, the city's smart fire protection supervision platform will mark the previously closed fire hazard as a suspected rebound; when the rebound index shows a weak recurrence in the same time period... If the conditions for determining a rebound are not met, the normally cancelled status of the fire hazard h will be maintained.

[0138] In this embodiment, the rebound state refers to the weak recurrence rebound index based on the same time period. The status of confirmed fire hazards that have been removed from the list for subsequent supervision; the rebound status includes at least the normal removal status and the suspected rebound status; the normal removal status means that no weak signs of recurrence that are sufficient to trigger further supervision have been identified; the suspected rebound status means that the fire hazards that have been removed from the list have experienced low-intensity abnormal accumulation in the same location, with the same risk consequences and the same historical high incidence period after removal from the list, and need to enter the supervision action selection process.

[0139] S5. When the rebound status is a suspected rebound status, the weak recurrence rebound index and rebound status in the same time period are used as the input basis for the reinforcement learning model to select regulatory actions.

[0140] When step S4 determines that the rebound status of the fire hazard h that has been cleared is a suspected rebound status, the city's smart fire protection supervision platform calls the reinforcement learning model to select a supervision action.

[0141] The state of the reinforcement learning model The index of weak rebound from the same period Rebound state, low-intensity abnormal state set The data consists of the ongoing abnormal situation, the duration from the confirmation of account closure to the present moment, the historical rectification feedback from the responsible unit, and the current status of regulatory resource occupancy.

[0142] In this embodiment, the state of the reinforcement learning model This refers to a data combination used to describe the current resurgence of previously identified fire hazards; status. It is not a single risk score, but is composed of rebound identification results, abnormal persistence, duration after cancellation, rectification performance of responsible units, and regulatory resource utilization, enabling the reinforcement learning model to select different regulatory actions in different regulatory scenarios;

[0143] Among them, the abnormal persistence condition is based on the set of low-intensity abnormal states. The occurrence time, duration, and consecutive occurrence of each low-intensity abnormal state are determined; the duration from the confirmation time of closure to the current time is determined based on the closure confirmation time in the hazard anchor point h and the current system time; the historical rectification feedback of the responsible unit is obtained from the closure fire hazard records or fire supervision data, which is used to characterize the responsible unit's past response to fire hazard rectification tasks; the current status of regulatory resource occupancy is determined from the regulatory task records of the city's smart fire supervision platform, which is used to characterize the current regulatory resource occupancy that can be used to execute self-inspection reminders, online reviews, grid member on-site inspections, third-party maintenance reviews, or fire supervision spot checks.

[0144] In one implementation, the urban smart fire protection monitoring platform will use the weak recurrence index of the same time period. Low-intensity abnormal state set The abnormal persistence status, the duration from the confirmation of account closure to the current moment, the historical rectification feedback status of the responsible unit, and the current status of regulatory resource occupancy are each classified into status levels according to the historical sample distribution. Each status level is then combined with the rebound status to form the status of the reinforcement learning model. ;

[0145] Among them, the index of weak rebound in the same period The status level is determined based on the distribution of the weak recurrence and rebound index in the same time period in the historical sample of eliminated hidden dangers; the status level of the abnormal persistence situation is based on the set of low-intensity abnormal states. The duration, frequency, and consecutive occurrence of abnormal states are determined; the status level of the duration from the confirmation time of cancellation to the current time is determined based on the post-cancellation observation period of the cancelled fire hazards; the status level of the responsible unit's historical rectification feedback is determined based on the responsible unit's historical rectification timeliness rate, historical review pass rate, or historical confirmation rebound number; the status level of the current regulatory resource occupancy status is determined based on the number of incomplete tasks, the number of available grid members, the available maintenance task capacity, or the occupancy of supervision and spot check tasks in the regulatory task records of the city's smart fire protection supervision platform.

[0146] Through the aforementioned state level classification, reinforcement learning models can transform regulatory data of different dimensions into states that can be used for action selection. This avoids directly inputting incomparable data into the action selection process;

[0147] Actions of reinforcement learning models This includes sending self-inspection reminders to responsible units, generating online review tasks, assigning grid members to conduct on-site verification tasks, triggering third-party maintenance review tasks, and submitting suggestions for fire safety supervision and spot checks.

[0148] In this embodiment, the regulatory action This refers to the monitoring and handling actions taken by the city's smart fire safety supervision platform for previously closed fire hazards that are suspected of recurring; monitoring actions This is not a regular alarm action, but a type of handling task corresponding to fire safety supervision. It is used to select a handling method that actually restricts the responsible unit or supervisory personnel based on different levels of rebound, different durations of abnormality, and different levels of supervisory resource occupancy.

[0149] In one implementation, the reinforcement learning model uses an action-value table or a neural network to approximate the action-value function. When selecting regulatory actions, the reinforcement learning model considers the current state. For each corresponding regulatory action, select the regulatory action with the highest action value as the output regulatory action. ;

[0150] During the model training or online update phase, the reinforcement learning model selects regulatory actions from the action set according to a preset exploration rule. The preset exploration rule is used to conduct limited exploration of other regulatory actions beyond the existing regulatory actions with the highest action value, so as to obtain feedback on the execution effect of different regulatory actions under different rebound states. In this way, the reinforcement learning model can update the action value of each regulatory action under different states during continuous monitoring.

[0151] In one implementation, the preset exploration rule is an ε-greedy exploration rule, which means that during the training or online update phase, a regulatory action is randomly selected from the action set with a preset exploration probability, and the regulatory action with the highest value is selected with the remaining probability; the preset exploration probability is determined by the reinforcement learning model training process; through the ε-greedy exploration rule, the reinforcement learning model can maintain limited exploration capability when the value of the regulatory action is known, so as to obtain the execution feedback of different regulatory actions in different suspected bounce states;

[0152] Rewards in reinforcement learning models The determination is based on the effects of regulatory actions, including changes in the rebound index, the elimination of low-intensity anomalies, the results of review and confirmation, and the consumption of regulatory resources.

[0153] In this embodiment, the reward This refers to regulatory actions. After execution, the city's smart fire protection monitoring platform generates feedback values ​​based on the effectiveness of the actions; rewards are given. This is used to reflect whether the regulatory action helps reduce the weak recurrence index in the same time period, eliminate low-intensity abnormal conditions, confirm or eliminate the risk of a rebound, and also reflects the regulatory resources occupied by the action; through rewards Reinforcement learning models can adjust regulatory actions in subsequent similar rebound scenarios;

[0154] If the weak recurrence rebound index decreases after the regulatory action is implemented, the low-intensity abnormal state decreases or is eliminated, and the regulatory resource consumption is low, the corresponding reward will be increased; if the low-intensity abnormal state persists after the regulatory action is implemented, the weak recurrence rebound index still meets the rebound judgment conditions, or the regulatory resource consumption is high, the corresponding reward will be decreased.

[0155] In one implementation, if the action execution feedback This indicates that the low-intensity anomaly has been eliminated, and the redefined weak recurrence rebound index for the same time period has been determined. If the rebound judgment condition is not met, the city's smart fire protection supervision platform will give a positive reward to the reinforcement learning model; if the action execution feedback is not met... This indicates that previously identified fire hazards have been confirmed as recurring, or that a weak recurrence of the same time period has been identified as a recurrence of the same type of fire hazard. If the rebound judgment conditions are still met, the city's smart fire protection supervision platform will give a negative reward to the reinforcement learning model; if the supervision action consumes more supervision resources, the corresponding resource consumption item will be deducted from the reward.

[0156] In one specific implementation, the reward This includes rewards for a decrease in the rebound index, rewards for eliminating low-intensity anomalies, rewards for verification and confirmation, and deductions for regulatory resource consumption. The reward for a decrease in the rebound index is determined based on the difference in the weak recurrence rebound index during the same time period before and after the regulatory action; the reward for eliminating low-intensity anomalies is based on the set of low-intensity anomalies after the regulatory action is implemented. The number of low-intensity abnormal states eliminated was determined; the reward was reviewed and confirmed based on the action execution feedback. Whether the low-intensity anomaly has been confirmed to be eliminated is determined; the deduction item for regulatory resource consumption is determined based on the task type corresponding to the regulatory action and the resource consumption of the task execution;

[0157] By using the above-mentioned reward determination method, reinforcement learning models can tend to select regulatory actions that can reduce the rebound index, eliminate low-intensity anomalies, and reduce the consumption of regulatory resources in the subsequent regulatory process.

[0158] In this embodiment, the reinforcement learning model adopts the action value update method in existing reinforcement learning, specifically:

[0159]

[0160] in, Indicates the state Select action value, Indicates the execution of an action The new state that was formed later, Indicates the execution of an action The reward obtained later Indicates the learning rate. a' represents the discount factor, and a' represents the new state. The following actions can be selected; learning rate and discount factor Determined by the reinforcement learning model training process, this embodiment does not set fixed empirical values;

[0161] Reinforcement learning models based on the current state Output regulatory actions The aforementioned regulatory actions Used to generate a regulatory task in step S6 and bound to the cancelled fire hazard h;

[0162] It should be noted that the reinforcement learning model in this embodiment is not used to identify fire images or predict fire occurrences, but rather to select appropriate regulatory actions based on the correlation data of the rebound status when a previously identified fire hazard shows signs of weak recurrence or rebound in the same time period, thereby improving the rationality of regulatory resource allocation.

[0163] S6. Generate regulatory tasks based on regulatory actions, receive action execution feedback, and update the status of closed fire hazards based on action execution feedback;

[0164] The city's smart fire protection monitoring platform outputs monitoring actions based on step S5. Generate a monitoring task and bind it to the fire hazard that has been cleared;

[0165] When regulatory actions When sending self-inspection reminders to responsible entities, the city's smart fire protection supervision platform generates a self-inspection reminder task and sends it to the responsible entity of the supervised object, targeting the original hazardous space unit or the same area under supervision. Self-inspection requirements; when regulatory actions When generating an online review task, the city's smart fire safety supervision platform generates the task and requires the responsible unit to submit review materials corresponding to the low-intensity abnormal state; when the supervision action... When assigning on-site verification tasks to grid workers, the city's smart fire safety supervision platform distributes the tasks to the corresponding grid workers, specifying the verification targets, verification spatial units, and verification risk consequences categories; when supervision actions are taken... When a third-party maintenance review task is triggered, the city's smart fire protection supervision platform will dispatch the review task to the corresponding maintenance unit; when the supervision action is initiated... When submitting fire safety supervision and inspection recommendations, the city's smart fire safety supervision platform generates fire safety supervision and inspection recommendations, and links them to fire hazards that have been resolved (h) and the weak recurrence and rebound index in the same time period. and low-intensity abnormal state set ;

[0166] After the regulatory task is completed, the city's smart fire protection regulatory platform receives feedback on the action execution. ;

[0167] In this embodiment, the action execution feedback This refers to regulatory actions. The corresponding regulatory tasks are completed and the resulting data is returned to the city's smart fire protection regulatory platform; action execution feedback. This includes self-inspection feedback from responsible units, online review results, on-site verification results from grid workers, third-party maintenance review results, or fire safety supervision and spot check results; action execution feedback. It is used both to update the status of fire hazards that have been removed from the list and to form reward feedback for reinforcement learning models.

[0168] The city's smart fire protection monitoring platform provides feedback based on action execution. Update the set of low-intensity abnormal states And redetermine the weak rebound index in the same period. ;

[0169] When the action is executed, feedback is received. This indicates that the low-intensity anomaly has been eliminated, and the re-established weak recurrence rebound index for the same time period has been determined. If the rebound judgment conditions are not met, the city's smart fire protection supervision platform will maintain the normal cancellation status of the cancelled fire hazard h and record the execution result corresponding to this supervision action;

[0170] When the action is executed, feedback is received. This indicates that the original hidden danger has re-emerged as a formal fire hazard, or that a weak recurrence and rebound index has been re-determined within the same time period. If the conditions for determining a rebound are still met, the city's smart fire safety supervision platform will update the status of the previously closed fire hazard h from suspected rebound to confirmed rebound and generate a rectification task.

[0171] In this embodiment, the confirmation of rebound status refers to the feedback of action execution after the fire hazard h that has been cancelled is cancelled. Confirmation of the re-emergence of a formal fire hazard, or a weak recurrence rebound index in the same time period after recalculation. The state formed when the rebound judgment conditions are continuously met; the confirmed rebound state is different from the suspected rebound state. The suspected rebound state is used to trigger the selection of regulatory actions. The confirmed rebound state is used to trigger the generation of re-rectification tasks.

[0172] When the action is executed, feedback is received. When it is indicated that the abnormal status is not related to the fire hazard h that has been cleared, the city's smart fire protection monitoring platform will remove the corresponding abnormal status from the low-intensity abnormal status set. Remove from the middle and provide feedback on the action execution. As a reward feedback for reinforcement learning models;

[0173] In one implementation, the urban smart fire protection monitoring platform provides feedback on action execution. Forming post-execution state The post-execution state The updated weak rebound index for the same time period Updated rebound state, updated set of low-intensity abnormal states The data consists of the ongoing abnormal situation, the duration from the confirmation of account closure to the present moment, the historical rectification feedback from the responsible unit, and the current status of regulatory resource occupancy.

[0174] The city's smart fire protection monitoring platform provides feedback based on action execution. Determine the reward and will execute the pre-state Regulatory actions already taken ,award and post-execution status Input the reinforcement learning model, and the state in the reinforcement learning model Regulatory action The corresponding action value is updated; thus, action execution feedback is generated. It is used not only to update the status of fire hazards that have been deactivated, but also to correct the regulatory action selection results of the reinforcement learning model in subsequent suspected rebound states.

[0175] Example 2: This example provides a safety monitoring system for smart urban fire protection, which is used to execute the method described in Example 1;

[0176] The system includes a module for generating anchor points for closed-loop risks, a module for determining the scope and historical time period of peer supervision, a module for extracting low-intensity abnormal states, a module for calculating the weak recurrence and rebound index, a module for selecting supervisory actions through reinforcement learning, and a module for updating supervision.

[0177] The hazard cancellation anchor point generation module is used to obtain the fire hazard records that have been cancelled and generate the hazard cancellation anchor point h; the hazard cancellation anchor point h includes the regulatory object identifier, the original hazard spatial unit, the original hazard risk consequence label, the cancellation confirmation time, and the historical abnormal time series before cancellation; the hazard cancellation anchor point generation module sends the hazard cancellation anchor point h to the peer regulatory scope and historical time period determination module and the low intensity abnormal state extraction module;

[0178] The module for determining the scope of supervision at the same location and the historical time period is used to determine the scope of supervision at the same location based on the original spatial unit of the hazard in the hazard anchor point h after the hazard is delisted. The historical time period distribution is determined based on the historical abnormal time series before the cancellation of the identified hidden dangers in the anchor point h. The module for determining the scope of corresponding supervision and historical time periods will define the scope of corresponding supervision. Send to the low-intensity anomaly extraction module to distribute historical time periods. Send to the weak recurrence rebound index calculation module;

[0179] The low-intensity abnormal state extraction module is used to obtain fire supervision data after the regulatory object is deregistered, based on the regulatory object identifier, starting from the time of deregistration confirmation, and to filter data occurring within the same regulatory scope. An abnormal state that is consistent with or belongs to the risk consequence category corresponding to the original hazard risk consequence label, and is based on the upper limit of stability. Boundary between formal hazard assessment and Extracting low-intensity anomalous state sets The low-intensity abnormal state extraction module will extract the low-intensity abnormal state set. Stable upper bound Boundary between formal hazard assessment and Send to the weak recurrence rebound index calculation module;

[0180] The weak recurrence rebound index calculation module is used to calculate the index based on a set of low-intensity abnormal states. Stable upper bound Formal hazard assessment boundaries Scope of supervision at the same level Distribution of historical periods Calculate the weak recurrence index in the same time period And based on the weak rebound index in the same period The rebound status is determined; the weak replication rebound index calculation module will calculate the weak replication rebound index for the same time period. The rebound status is sent to the reinforcement learning regulatory action selection module;

[0181] The reinforcement learning-based regulatory action selection module is used to select the weakly recurring rebound index in the same time period when the rebound status is suspected to be a rebound status. Rebound state, low-intensity abnormal state set The state of the reinforcement learning model is composed of the ongoing abnormalities, the duration from the confirmation of account closure to the present moment, the historical rectification feedback from the responsible unit, and the current status of regulatory resource occupancy. The reinforcement learning model then outputs regulatory actions. The reinforcement learning regulatory action selection module will select regulatory actions. Send to the regulatory update module;

[0182] The regulatory update module is used to update information based on regulatory actions. Generate corresponding regulatory tasks and receive action execution feedback. Based on action execution feedback Update the set of low-intensity abnormal states And re-determine the weak rebound index in the same time period. And based on the weak rebound index in the newly determined same time period and action execution feedback Update the status of fire hazards that have been removed from the list; the monitoring update module will also provide action execution feedback. The reward feedback from the reinforcement learning model is sent to the reinforcement learning supervisory action selection module;

[0183] In the above system, each module is deployed on the same server, the same city smart fire protection supervision platform, or multiple servers that are interconnected; data interaction between modules is achieved through internal platform data interfaces, task message queues, or database read / write methods; this embodiment does not limit the specific software and hardware deployment form, as long as the above data processing flow can be achieved.

[0184] Example 3: Taking a commercial complex as an example, the commercial complex was included in the key supervision objects by the city's smart fire protection supervision platform; the platform once found that there was a hidden danger of goods stacking in the evacuation passage on the east side of the first floor of the commercial complex; after the responsible unit completed the cleanup and submitted the rectification materials, the supervisors reviewed and confirmed that the hidden danger had been eliminated, and the platform marked the hidden danger as closed.

[0185] The city's smart fire safety supervision platform generates a closed-out hazard anchor point h based on the closed-out fire hazard record; among them, the supervision object identifier is the unit number of the commercial complex in the platform, the original hazard space unit is the evacuation passage on the east side of the first floor, the original hazard risk consequence label is the evacuation escape obstruction category, the closure confirmation time is the review pass time, and the historical abnormal time series record before the closure of the hazard records the time when goods were piled up multiple times before the hazard was closed.

[0186] The platform determines the scope of supervision based on the original potential hazard spatial unit. In this application example, the scope of peer supervision This includes the evacuation route on the east side of the first floor, adjacent safety exits, adjacent stairwell entrances, and adjacent areas along the same evacuation route; the platform determines the historical time period distribution based on the historical anomaly time series before the anomaly was deactivated. It was found that this potential problem mainly occurred during the evening restocking period before the account was closed.

[0187] After the account is deregistered, the platform will retrieve inspection records, passageway occupancy records, unit self-inspection records, and grid worker verification records based on the regulatory object identifier of the commercial complex from the time of account deregulation confirmation. The platform will then filter out instances occurring within the same regulatory scope. The abnormal state is categorized as a risk consequence of obstructed evacuation and escape; subsequently, the platform determines the upper bound of stability based on stable operation samples after the account is closed. The formal hazard determination boundary shall be determined according to fire safety supervision rules or manually verified samples. ;

[0188] During the subsequent observation period, the platform discovered that short-term container congestion occurred multiple times near the evacuation passage on the east side of the first floor of the commercial complex and near the adjacent safety exit. While this situation did not yet warrant a re-identification as a potential hazard of evacuation passage blockage, its abnormal intensity was higher than the stable state after rectification. Therefore, the platform identified it as a low-intensity abnormal state and included it in the low-intensity abnormal state set. ;

[0189] The platform targets sets of low-intensity abnormal states. Calculate the weak recurrence index in the same time period Because the low-intensity anomaly occurred in the original hazardous spatial unit or in an area adjacent to the same evacuation route, and the time of occurrence coincided with the historical high-incidence period before the closure of the hazard, and the risk consequences were classified as obstructed evacuation, the low-intensity anomaly had a weak recurrence rebound index for the same period. It makes a significant contribution; when the index shows a weak rebound in the same period, it will contribute significantly. When the conditions for determining a potential rebound are met, the platform marks the previously closed fire hazard as a suspected rebound.

[0190] The platform will show a weak rebound index during the same period. Rebound state, low-intensity abnormal state set The status of the reinforcement learning model consists of the ongoing abnormal situation, the duration from the confirmation time of cancellation to the current time, the historical rectification feedback of the responsible unit, and the current status of regulatory resource occupancy; the reinforcement learning model selects regulatory actions from self-inspection reminders, online review tasks, grid member on-site verification tasks, third-party maintenance review tasks, and fire supervision spot check suggestions;

[0191] If the supervisory action output by the reinforcement learning model is to assign a grid worker to conduct an on-site inspection, the platform generates the on-site inspection task and binds the task to the already resolved fire hazard h; after the grid worker conducts the on-site inspection, the inspection result is used as the action execution feedback. Upload to the platform;

[0192] When the action is executed, feedback is received. This indicates that the area has been cleared, and the platform has updated the low-intensity anomaly status set. The newly determined weak rebound index in the same time period If the rebound judgment conditions are not met, the platform will maintain the normal cancellation status of the cancelled fire hazard; when the action execution feedback is received... This indicates that the area has once again become a potential source of blockages in evacuation routes, or that a weak recurrence of the rebound index has been identified at the same time point. If the conditions for a rebound are still met, the platform will update the status of the previously closed fire hazard from suspected rebound to confirmed rebound and regenerate the rectification task.

[0193] As can be seen from the above application examples, this embodiment does not only trigger the action after the fire hazard reaches a serious level again, but can identify the risk rebound trend in advance when a low-intensity anomaly with the same location, the same consequences, and the same high-incidence period appears after the hazard has been cleared, thereby improving the timeliness and closed-loop authenticity of urban smart fire supervision.

[0194] Example 4: Taking the hidden danger of insufficient fire hydrant pressure in a high-rise building as an example, the hidden danger was marked as cancelled after being repaired by the maintenance unit, reported by the responsible unit, and reviewed by the monitoring platform; the platform generated a cancelled hidden danger anchor point h, in which the original hidden danger spatial unit was the fire hydrant point on a certain floor of the high-rise building, the original hidden danger risk consequence label was fire protection facility failure, the cancellation confirmation time was the review approval time, and the historical abnormal time series before cancellation recorded the time when the fire hydrant pressure was insufficient;

[0195] The platform determines the scope of supervision based on the original potential hazard spatial unit. The scope of this peer supervision This includes the fire hydrant location, adjacent fire hydrant locations on the same fire water supply branch, and associated fire water pressure monitoring points; the platform determines the historical time period distribution based on the historical anomaly time series before the issue was closed. ;

[0196] After the account is deactivated, the platform obtains the status records, maintenance records, and inspection records of the fire protection facilities of the high-rise building based on the identifier of the regulated object; if adjacent points on the same fire water supply branch repeatedly experience short-term pressure fluctuations, maintenance review delays, or abnormal inspection records, and the intensity of the abnormal state is greater than the upper limit of stability. And smaller than the formal hazard assessment threshold The platform includes it in the low-intensity abnormal state set as a low-intensity abnormal state. ;

[0197] The platform is based on a set of low-intensity abnormal states. Calculate the weak recurrence index in the same time period If the low-intensity abnormal state is located in the same fire water supply branch as the original hidden danger space unit, the time of occurrence has a high degree of overlap with the historical high-incidence time before the cancellation, and the risk consequences belong to the category of fire protection facility failure, then the platform will identify it as a weak recurrence and rebound of the original cancelled fire hazard.

[0198] When the index shows a weak rebound in the same time period When the rebound judgment conditions are met, the platform marks the fire hazard that has been cleared as a suspected rebound and selects regulatory actions through the reinforcement learning model. If the current regulatory resources are occupied at a high level and the responsible unit has timely historical rectification feedback, the reinforcement learning model will prioritize outputting online review tasks or triggering third-party maintenance review tasks. If low-intensity anomalies continue to occur and the responsible unit has poor historical rectification feedback, the reinforcement learning model will output grid member on-site verification tasks or fire supervision spot check suggestions.

[0199] After the task is executed, the platform will provide feedback on the action execution. Update the set of low-intensity abnormal states Re-determine the weak rebound index in the same time period. And based on the weak rebound index in the newly determined same time period. and action execution feedback Update the status of the now-closed fire hazard;

[0200] Through this embodiment, the platform can identify the low-intensity rebound trend of fire protection facility hazards after they have been cleared, avoiding the need to initiate the handling process only after the fire protection facilities have returned to a formal fault state.

[0201] Example 5: Taking the hidden danger of illegally installed electrical wires in a street-side restaurant as an example, the hidden danger was marked as cancelled by the platform after the responsible unit removed the illegal wires and passed the review. The platform generates a cancelled hidden danger anchor point h, in which the original hidden danger spatial unit is the kitchen electricity area, the original hidden danger risk consequence label is the category of illegal electricity use, the cancellation confirmation time is the review pass time, and the historical abnormal time series before cancellation records the time when the hidden danger appeared.

[0202] The platform determines the scope of supervision based on the original potential hazard spatial unit. This includes the kitchen's electricity usage area, areas using the same electrical circuit, and adjacent electricity usage areas managed by the same responsible entity; the platform determines the historical time period distribution based on the historical anomaly time sequence before the case is closed. ;

[0203] After the account is deactivated, the platform obtains records of abnormal electricity use, self-inspection records of the unit, and verification records of grid members based on the identifier of the regulated object. If the platform finds that the kitchen's electricity use area repeatedly experiences short-term load anomalies, temporary power strip usage records, or minor electricity violations during peak operating hours, and the intensity of the anomaly is higher than the stable state after rectification but has not yet reached the threshold for formal hazard assessment, then it will be included in the low-intensity anomaly state set as a low-intensity anomaly state. ;

[0204] The platform further calculates the weak recurrence index for the same time period. Since the abnormal state occurred in the original hidden danger space unit or the same power circuit area, the time of occurrence overlapped with the historical high-incidence time before the case was closed, and the risk consequences belonged to the category of illegal power use, the platform was able to identify the abnormal state as a weak recurrence and rebound of the closed hidden danger in the same time period.

[0205] When the index shows a weak rebound in the same time period When the rebound judgment conditions are met, the platform marks the fire hazard that has been cleared as a suspected rebound status; the reinforcement learning model selects regulatory actions based on the current status; if the system judges that the responsible unit has a good history of rectification feedback and the current abnormal intensity is close to the stable upper limit, it outputs a self-inspection reminder or an online review task; if the system judges that low-intensity abnormalities continue to occur and the current regulatory resources allow, it outputs a grid member on-site verification task or submits a fire supervision spot check suggestion.

[0206] After the regulatory task is completed, the platform will analyze the feedback from the action execution. Update the set of low-intensity abnormal states And redetermine the weak rebound index in the same period. When the action is executed, feedback is received. When the anomaly has been resolved, the platform maintains the normal closed status of the fire hazard; when the action execution feedback... When it is indicated that illegal electricity use has re-emerged as a formal fire hazard, the platform updates the previously closed fire hazard to a confirmed rebound status and regenerates the rectification task.

[0207] This embodiment demonstrates that the present invention can be applied to the identification of rebounds after the cancellation of illegal electricity use hazards, and can automatically select regulatory actions based on the rebound status, thereby achieving continuous monitoring of the rectification effect of urban fire hazards.

[0208] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A safety supervision method for smart urban fire protection, characterized in that, Includes the following steps: S1. Obtain the records of fire hazards that have been cancelled and generate anchor points for cancelled hazards. The anchor points for cancelled hazards include the regulatory object identifier, the original hazard spatial unit, the original hazard risk consequence label, the cancellation confirmation time, and the historical abnormal time series before cancellation. S2. Determine the scope of supervision based on the original spatial unit of the hidden danger, and determine the distribution of historical time periods based on the historical abnormal time series before the cancellation. S3. From the time of confirmation of cancellation, obtain the fire supervision data of the supervised object after cancellation based on the identifier of the supervised object, screen the abnormal status that occurs within the same supervision scope and is consistent with the original hidden danger risk consequence label or belongs to the risk consequence category corresponding to the original hidden danger risk consequence label, and extract the set of low intensity abnormal status based on the stable upper bound determined by the stable operation sample after cancellation and the formal hidden danger judgment boundary determined by the fire supervision rules, platform hidden danger judgment rules or manual confirmation sample. S4. Calculate the weak recurrence rebound index for the same time period based on the set of low-intensity abnormal states, the scope of supervision at the same time, and the distribution of historical time periods, and determine the rebound status based on the weak recurrence rebound index for the same time period. S5. When the rebound status is a suspected rebound status, the weak recurrence rebound index and rebound status in the same time period are used as the input basis for the reinforcement learning model to select regulatory actions. S6. Generate regulatory tasks based on regulatory actions, receive action execution feedback, and update the status of closed fire hazards based on the action execution feedback.

2. The safety supervision method for smart urban fire protection according to claim 1, characterized in that, The original hazard risk consequence labels are used to characterize the fire safety impact caused by fire hazards that have been eliminated, including: classifying obstructed evacuation routes, locked safety exits, and closed escape windows as evacuation obstruction risk consequences; classifying insufficient fire hydrant pressure, abnormal sprinkler systems, and alarm host malfunctions as fire facility failure risk consequences; and classifying electric vehicle charging in homes, unauthorized wiring, and illegal connection of high-power electrical appliances as illegal electricity use risk consequences.

3. A safety supervision method for smart urban fire protection according to claim 1, characterized in that, The scope of supervision is determined based on the original potential hazard spatial unit, including: The original hazardous space units will be included in the scope of same-location supervision. Identify adjacent spatial units that have a direct fire risk connection with the original hazardous spatial unit; The adjacent spatial units shall be included in the scope of co-location supervision; The direct fire risk associations include adjacent evacuation routes, connected fire protection facility circuits, connected fire compartments on the same floor, connected electrical circuits, or managed by the same responsible entity.

4. A safety supervision method for smart urban fire protection according to claim 1, characterized in that, The historical time period distribution is determined based on the historical anomaly time series before the account was closed, including: Time slots are divided according to the regulatory cycle set by the regulatory platform; Statistical analysis of the number of abnormal occurrences of fire hazards that have been removed from the list within each time period prior to their removal; The historical occurrence probability for each time slice is determined by the proportion of the number of anomalies occurring in each time slice to the total number of anomalies occurring in all time slices. The historical time period distribution is formed by combining each time slice with its corresponding historical occurrence probability.

5. A safety supervision method for smart urban fire protection according to claim 1, characterized in that, Based on the stability upper bound and the formal hazard determination boundary, a set of low-intensity abnormal states is extracted, including: The upper bound of stability is determined based on stable operating samples after the cancellation of the number; The formal hazard determination boundary shall be determined in accordance with fire safety supervision rules, platform hazard determination rules, or manually confirmed samples. Determine the intensity of the abnormality in the abnormal state; When the abnormal intensity is greater than the upper limit of stability but less than the threshold for formal hazard identification, the abnormal state is regarded as a low-intensity abnormal state. Summarize low-intensity abnormal states to form a set of low-intensity abnormal states; The abnormal intensity is calculated based on the patrol deduction points, the degree of deviation of facility status, the degree of passage occupancy, the maintenance abnormality level, or the existing abnormality level of the platform, according to the unified abnormal intensity scale preset by the urban smart fire supervision platform; the fire supervision data includes patrol records, maintenance records, rectification and review records, fire facility status records, passage occupancy records, electricity abnormality records, unit self-inspection records, and grid member verification records.

6. A safety supervision method for smart urban fire protection according to claim 1, characterized in that, The calculation of the weak recurrence rebound index in the same time period includes: The degree of weak anomaly is determined by the position of the anomalous intensity of the low-intensity anomaly state between the upper limit of stability and the boundary for formal hazard assessment. The spatial isomorphism is determined based on the degree of correlation between the spatial unit where the low-intensity abnormal state is located and the original hidden danger spatial unit in the fire protection spatial topology. The overlap of time periods is determined by the historical probability of occurrence of time slices of low-intensity abnormal states in the historical time period distribution. The similarity of risk consequences is determined based on the consistency between the risk consequence label of the low-intensity abnormal state and the original risk consequence label of the hidden danger, or whether the two belong to the same risk consequence category. The rebound contribution of a single low-intensity anomaly is determined by multiplying the degree of weak anomaly, spatial co-location, temporal overlap, and similarity of risk consequences. The rebound contributions of each low-intensity anomaly in the set of low-intensity anomalies are then summed to obtain the weak recurrence rebound index in the same time period.

7. A safety supervision method for smart urban fire protection according to claim 1, characterized in that, Determining the rebound status includes: The criteria for determining a rebound are determined based on historical samples of previously identified hidden risks. When the weak recurrence rebound index in the same time period meets the rebound judgment conditions, the fire hazards that have been cancelled will be marked as suspected rebound status; When the weak rebound index in the same time period does not meet the rebound judgment conditions, the normal cancellation status of the cancelled fire hazards shall be maintained.

8. A safety supervision method for smart urban fire protection according to claim 1, characterized in that, Regulatory actions are selected based on reinforcement learning models, including: The reinforcement learning model consists of the following: the weak recurrence rebound index, rebound status, abnormal persistence in the set of low-intensity abnormal statuses, the duration from the confirmation time of cancellation to the current time, the historical rectification feedback of responsible units obtained from the cancelled fire hazard records or fire supervision data, and the current status of supervision resource occupancy determined from the supervision task records of the city's smart fire supervision platform. The actions of the reinforcement learning model will include sending self-inspection reminders to responsible units, generating online review tasks, assigning grid members to conduct on-site verification tasks, triggering third-party maintenance review tasks, and submitting suggestions for fire supervision spot checks. The reward for the reinforcement learning model is determined based on changes in the rebound index, the elimination of low-intensity anomalies, the results of verification and confirmation, and the consumption of regulatory resources. Regulatory actions are determined based on the actions output by the reinforcement learning model.

9. A safety supervision method for smart urban fire protection according to claim 1, characterized in that, Update the status of closed fire hazards based on action execution feedback, including: The set of low-intensity abnormal states is updated based on the feedback from the action execution, and the weak recurrence rebound index in the same time period is redefined. When the action execution feedback indicates that the low-intensity anomaly has been eliminated, and the re-determined weak recurrence rebound index in the same time period does not meet the rebound judgment condition, the normal cancellation status of the cancelled fire hazard shall be maintained. When the action execution feedback indicates that the original hidden danger has re-formed into a formal fire hazard, or the newly determined weak recurrence rebound index in the same time period still meets the rebound judgment conditions, the fire hazard that has been cancelled will be updated from the suspected rebound status to the confirmed rebound status, and a rectification task will be generated. When the action execution feedback indicates that the abnormal state is not related to the fire hazard that has been cleared, the corresponding abnormal state is removed from the set of low-intensity abnormal states, and the action execution feedback is used as the reward feedback for the reinforcement learning model.

10. A safety monitoring system for smart urban fire protection, used to execute the safety monitoring method for smart urban fire protection as described in any one of claims 1-9, characterized in that, include: The module for generating deregistered hazard anchor points is used to obtain deregistered fire hazard records and generate deregistered hazard anchor points. The deregistered hazard anchor points include the regulatory object identifier, the original hazard spatial unit, the original hazard risk consequence label, the deregistration confirmation time, and the historical abnormal time series before deregistration. The module for determining the scope of supervision and historical time periods is used to determine the scope of supervision based on the original spatial unit of the hidden danger, and to determine the distribution of historical time periods based on the historical abnormal time series before the cancellation of the case. The low-intensity abnormal state extraction module is used to obtain fire supervision data after the supervision object is cancelled based on the supervision object identifier from the time of cancellation confirmation. It filters abnormal states that occur within the same supervision scope and are consistent with or belong to the risk consequence category corresponding to the original hidden danger risk consequence label. Based on the stable upper bound determined by the stable operation sample after cancellation and the formal hidden danger judgment boundary determined by fire supervision rules, platform hidden danger judgment rules or manual confirmation samples, it extracts the set of low-intensity abnormal states. The weak recurrence rebound index calculation module is used to calculate the weak recurrence rebound index for the same time period based on the set of low-intensity abnormal states, the scope of supervision at the same time, and the distribution of historical time periods, and to determine the rebound status based on the weak recurrence rebound index for the same time period. The reinforcement learning-based regulatory action selection module is used to select regulatory actions based on the weak recurrence index and rebound status in the same time period when the rebound status is suspected to be a rebound status. The regulatory update module is used to generate regulatory tasks based on regulatory actions, receive feedback on action execution, and update the status of closed fire hazards based on the action execution feedback.

Images