Double-dimension accurate detection system for subway fire detector false negative and false positive

Abstract

This invention belongs to the field of fire detection technology and discloses a dual-dimensional accurate detection system for missed and false alarms of subway fire detectors. It includes: an observation construction module for forming a zoned observation sequence; a response construction module for forming a zoned reference response sequence and a reference response confidence marker; a deviation generation module for forming a group deviation evidence sequence based on the zoned reference response sequence and the reference response confidence marker; an evidence generation module for forming a drift evidence sequence based on the deviation of the current sampling window from the baseline response interval; an anomaly attribution module for attributing and fusing the group deviation evidence sequence and the drift evidence sequence to form a missed alarm anomaly criterion and a false alarm anomaly criterion; and a result output module for outputting the missed alarm accuracy inaccuracy level, the false alarm accuracy inaccuracy level, and the accuracy compliance conclusion. This invention achieves online accuracy detection of the missed alarm tendency and false alarm tendency of fire detectors.

Description

Technical Field

[0001] This invention relates to the field of fire protection detection technology, and more specifically, to a dual-dimensional accurate detection system for missed and false alarms of subway fire detectors. Background Technology

[0002] Subway stations typically deploy multiple fire detectors according to fire compartments to continuously monitor fire indicators such as smoke and temperature. Current maintenance methods rely heavily on periodic inspections, manual spot checks, equipment self-tests, and single-point threshold exceedance detection. While these methods can detect explicit problems such as offline status, communication interruptions, self-test failures, and serious malfunctions, they lack effective online identification methods for accuracy deviations during online operation, especially hidden inaccuracies that are prone to missed or false alarms.

[0003] In real-world subway scenarios, it's often difficult to continuously obtain accurate smoke concentration or temperature rise data for direct comparison. Therefore, even if a fire detector exhibits low, high, or delayed response, existing methods typically struggle to accurately assess its accuracy as long as it remains online and hasn't triggered a self-test anomaly. Especially within the same fire compartment, the arrival times, self-test status, and missing data of different fire detectors are not entirely consistent. Without unified sampling alignment and validity constraints, it's difficult to establish a reliable and comparable observational basis between different detector outputs, and subsequent analysis results are easily influenced by abnormal data.

[0004] On the other hand, fire detectors of the same type within the same fire compartment typically exhibit a consistent trend of change under the same environmental conditions. However, once an individual fire detector experiences accuracy drift, it often shows a continuous deviation from the collective response of the group or a continuous shift from its own historical stable state. Existing technologies generally lack a mechanism that can simultaneously utilize both the collective response of the group within the compartment and the historical stable response of an individual detector for joint judgment. They also lack methods to limit interference situations such as communication anomalies, unreliable reference responses, and overall synchronous fluctuations within the compartment. Consequently, it is easy to misjudge overall environmental changes as individual accuracy anomalies, or to submerge long-term individual drifts within overall compartment fluctuations, making it difficult to accurately distinguish between missed and false alarm inaccuracies.

[0005] Therefore, there is an urgent need for an online accuracy testing method for subway fire detectors that can reliably determine missed and false alarm inaccuracies in online operation, even in the absence of real-world references, based on the group response relationship between fire detectors within the same fire compartment and the historical response characteristics of individual fire detectors.

[0006] In view of this, the present invention proposes a dual-dimensional accurate detection system for missed and false alarms of subway fire detectors to solve the above problems. Summary of the Invention

[0007] To address the shortcomings of existing technologies, the present invention aims to provide a dual-dimensional accurate detection system for missed and false alarms of subway fire detectors. By constructing a zoned observation sequence, a zoned reference response sequence, a group deviation evidence sequence, and a drift evidence sequence, the system performs attribution fusion on abnormal evidence, thereby achieving the goal of distinguishing between missed and false alarm inaccuracies of fire detectors in online operation under conditions where real-world references are lacking.

[0008] To achieve the above objectives, the present invention adopts the following technical solution:

[0009] The observation construction module is used to align and label the validity of data collected by each fire detector in a fire compartment at a uniform sampling time, forming a compartmental observation sequence.

[0010] Response building module: used to filter and aggregate the probe output values ​​at the same sampling time in the partitioned observation sequence to form a partitioned reference response sequence, and generate a reference response confidence flag for the partitioned reference response at each sampling time;

[0011] Deviation generation module: used to determine the deviation of each fire detector relative to the zone reference response based on the zone reference response sequence and reference response confidence flags, and to form a group deviation evidence sequence through sliding window direction consistency judgment;

[0012] Evidence generation module: used to extract the historical baseline sample set of each fire detector, establish the baseline response interval, and form a drift evidence sequence based on the deviation of the current sampling window from the baseline response interval;

[0013] Anomaly attribution module: used to identify link untrusted windows, reference untrusted windows and environmental common transition windows, and to perform attribution fusion on group deviation evidence sequences and drift evidence sequences to form missed anomaly criteria and false alarm anomaly criteria;

[0014] The results output module is used to determine and output the accuracy inaccuracy level of the missed alarm type, the accuracy inaccuracy level of the false alarm type, and the accuracy qualification conclusion based on the missed alarm type anomaly criteria and the false alarm type anomaly criteria.

[0015] Furthermore, methods for forming the partitioned reference response sequence and reference response confidence markers include:

[0016] Extract valid sampling points within the same fire compartment, of the same type of fire detector, and at the same sampling time to form a candidate reference sample set;

[0017] The candidate reference sample set is subjected to consistency screening to form a set of effective samples after screening.

[0018] The median value of the detection output values ​​in the filtered valid sample set is taken as the partition reference response value at the corresponding sampling time, and a partition reference response sequence is formed according to the sampling time order;

[0019] Based on the number of samples and the dispersion of the effective sample set after screening, a reference response confidence marker is generated for the corresponding sampling time.

[0020] Furthermore, methods for forming an effective sample set include:

[0021] First, calculate the median value of each detection output value in the candidate reference sample set at the corresponding sampling time, and use the larger of the absolute value of the median value and the preset minimum response benchmark value as the denominator of the deviation ratio. Then, divide the absolute value of the difference between each detection output value and the median value by the denominator of the deviation ratio to obtain the deviation ratio of the corresponding detection output value relative to the median value.

[0022] When the deviation ratio of the detected output value is greater than the preset deviation ratio threshold, the corresponding sample will be removed.

[0023] Samples with a deviation ratio not exceeding a preset deviation ratio threshold are retained to form a set of effective samples after screening.

[0024] Furthermore, methods for generating reference response confidence markers include:

[0025] Calculate the deviation ratio of each detection output value in the effective sample set relative to its median value, and use the median value of the deviation ratio as the sample dispersion at the corresponding sampling time;

[0026] When the number of valid samples meets the preset high confidence sample number condition and the sample dispersion does not exceed the preset high confidence dispersion threshold, the reference response confidence flag at the corresponding sampling time is marked as high.

[0027] When the number of valid samples meets the preset confidence sample number condition and the sample dispersion does not exceed the preset confidence dispersion threshold, the reference response confidence flag at the corresponding sampling time is marked as medium.

[0028] In all other cases, the reference response confidence flag is set to low.

[0029] Furthermore, methods for generating group deviation evidence sequences include:

[0030] The reference difference is obtained by subtracting the zone reference response value at the same sampling time from the detection output value of each fire detector at each sampling time.

[0031] The relative deviation ratio is obtained by dividing the reference difference by the larger of the absolute value of the partitioned reference response value and the minimum response benchmark value.

[0032] The relative deviation ratio within consecutive sampling times is used to determine the consistency of the sliding window direction, forming a negative group deviation window or a positive group deviation window.

[0033] The credibility of the negative or positive group deviation window is corrected by combining the reference response confidence flag, thus forming a group deviation evidence sequence.

[0034] Furthermore, methods for forming negative or positive group deviation windows include:

[0035] When the number of sampling points with negative relative deviation ratios within the same sliding window reaches the preset number of points with consistent direction, and the average absolute value of the relative deviation ratios within the sliding window is not less than the preset deviation threshold, the sliding window is recorded as a negative group deviation window.

[0036] When the number of sampling points with a positive relative deviation ratio within the same sliding window reaches the preset number of points with the same direction, and the average value of the absolute value of the relative deviation ratio within the sliding window is not less than the preset deviation threshold, the sliding window is recorded as a positive group deviation window.

[0037] Furthermore, methods for forming a sequence of drift evidence include:

[0038] Extract the effective sampling points of the target fire detectors during historical stable periods, and aggregate them according to the intraday time corresponding to the sampling time to form a historical baseline sample set;

[0039] Statistical analysis was performed on the historical baseline sample set to establish the baseline response interval;

[0040] The deviation of the average value of the sampling window from the baseline response interval and the continuous changes are determined to form a drift evidence sequence.

[0041] Furthermore, methods for establishing the baseline response interval include:

[0042] The detection output values ​​in the historical baseline sample set are statistically analyzed, and the low quantile statistics representing the lower boundary of the interval and the high quantile statistics representing the upper boundary of the interval are extracted. The normal response range formed by the low quantile statistics and the high quantile statistics is taken as the self-baseline response interval.

[0043] When the number of historical baseline samples is less than a preset sample number threshold, a baseline response interval is not established.

[0044] Furthermore, the drift evidence sequence includes both descending drift evidence and ascending drift evidence;

[0045] Calculate the average value of the detection output value corresponding to each sampling point within multiple consecutive sampling windows, and use it as the average value of the sampling window;

[0046] The difference between adjacent sampling windows is obtained by subtracting the average value of the sampling window with the later sampling time from the average value of the sampling window with the earlier sampling time.

[0047] When the average value of multiple consecutive sampling windows is lower than the lower boundary of the baseline response interval, and the number of negative differences in the differences between adjacent windows reaches a preset trend point number or the differences between adjacent windows are all zero, evidence of downward drift is formed.

[0048] Evidence of upward drift is formed when the average value of multiple consecutive sampling windows is higher than the upper boundary of the baseline response interval, and the number of positive differences in the differences between adjacent windows reaches a preset trend point or the differences between adjacent windows are all zero.

[0049] Furthermore, the methods for identifying link untrusted windows, reference untrusted windows, and the common transition window of the environment include:

[0050] When there is a communication abnormal sampling point at the corresponding sampling time, and multiple consecutive adjacent sampling times are marked as link abnormal times, or the number of communication abnormal sampling points in a single current sampling window reaches the abnormal point count threshold, the current sampling window is recorded as a link untrusted window.

[0051] When there is a sampling moment in the current sampling window where the reference response confidence flag is low, the current sampling window is recorded as a reference untrusted window;

[0052] Calculate the output change of each fire detector in the same fire compartment at adjacent sampling times. When the output change direction is consistent and the absolute value of the output change is not less than the preset change amplitude threshold, the current sampling window is recorded as the common transition window of the environment.

[0053] Furthermore, methods for attributing and fusing group deviation evidence sequences and drift evidence sequences to form criteria for missed detection anomalies and false positive anomalies include:

[0054] Within a sampling window that does not belong to the link untrusted window, the reference untrusted window, or the common transition window of the environment, the corresponding window matching is performed on the population deviation evidence sequence and the drift evidence sequence; when negative population deviation evidence and descent drift evidence exist simultaneously in the same sampling window, a missed anomaly criterion is formed.

[0055] When both positive population deviation evidence and rising drift evidence exist within the same sampling window, a false alarm anomaly criterion is formed.

[0056] In other cases, no exception criteria are output.

[0057] The technical effects and advantages of the dual-dimensional accurate detection system for missed and false alarms in subway fire detectors proposed in this invention are as follows:

[0058] This invention addresses the problem of insufficient on-site reference and difficulty in accurately identifying online accuracy inaccuracies in subway fire detectors during online operation. By constructing a zoned basic observation sequence, a zoned reference response sequence, a group deviation evidence sequence, a baseline drift evidence sequence, and an anomaly attribution gating mechanism, it achieves online differentiation and judgment of missed and false alarm inaccuracies in fire detectors, thereby improving the reliability of online accuracy detection.

[0059] This invention aligns and constrains the detection output, self-test status, and message arrival status of each fire detector within the same fire compartment at a unified sampling time, forming a comparable basic observation sequence for the compartment. This reduces judgment bias caused by inconsistent message timing, missing data, and failure to remove abnormal states, providing a consistent data foundation for subsequent accuracy analysis.

[0060] This invention filters and aggregates the detection outputs of similar fire detectors within the same fire compartment to form a compartment reference response sequence and reference response confidence flags. This provides a stable compartment reference for the current detection output of a single fire detector, thereby improving the problem of difficulty in determining online accuracy status when there is a lack of on-site true values.

[0061] This invention further constructs group deviation evidence and baseline drift evidence, which can identify anomalies from two dimensions: the common response relationship of the group within the partition and the historical stable response relationship of a single fire detector. By limiting the link untrusted window, the reference untrusted window and the environmental common transition window, the interference of communication anomalies, reference distortion and overall environmental synchronous fluctuations on the judgment results is reduced, thereby improving the pertinence and reliability of the results of distinguishing between missed and false alarm inaccuracies.

[0062] Based on the above technical means, the present invention can distinguish between missed alarms and false alarms in the online operation of fire detectors in the context of subway fire compartments, providing a basis for subsequent calibration, maintenance and risk management. Attached Figure Description

[0063] Figure 1 This is a module diagram of a dual-dimensional accurate detection system for missed and false alarms in a subway fire detector according to Embodiment 1 of the present invention;

[0064] Figure 2 This is a schematic diagram of the group deviation evidence generation process in Embodiment 1 of the present invention;

[0065] Figure 3 This is a schematic diagram of the anomaly attribution gating process in Embodiment 1 of the present invention. Detailed Implementation

[0066] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0067] Example 1

[0068] See Figure 1 As shown, this embodiment discloses a dual-dimensional accurate detection system for missed and false alarms in subway fire detectors, including:

[0069] The observation construction module aligns and marks the validity of the data collected by each fire detector at a unified sampling time, based on the detection output, self-test status and message arrival status of each fire detector in the fire compartment, to form a compartmentalized observation sequence.

[0070] The fire compartment refers to a local monitoring area in a subway station hall, platform, passageway, equipment room, or ancillary space of the station section, which is designated according to fire separation requirements and is constrained by the same smoke exhaust, ventilation, or fire control conditions.

[0071] The data collected by each fire detector at a uniform sampling time are aligned and their validity is marked. The specific method is as follows:

[0072] The data of each fire detector in the fire compartment is uniformly sampled and managed; specifically, the detection output value, self-test status and arrival time of the reported data message of each fire detector are collected, and a continuous sampling time sequence is generated according to the sampling period and a unified clock, so that the data of different fire detectors in the same fire compartment are all corresponding to the same time reference; the sampling period is set according to the data refresh rhythm and time sequence alignment requirements of the fire detectors in the fire compartment.

[0073] The detection output value refers to the detection value output by the fire detector at the corresponding sampling time, which represents the current smoke concentration response level or temperature response level; the self-test status is used to indicate whether the fire detector is in normal working condition; the arrival time of the reported data message is used to indicate whether the corresponding sampling data is late, missing, or has communication abnormalities.

[0074] A time mapping is performed on the reported data packets corresponding to each sampling time. Specifically, when the time difference between the arrival time of a reported data packet corresponding to a certain fire detector and the current sampling time does not exceed the time delay threshold, the detection output value corresponding to the reported data packet is mapped to the current sampling time. When the time difference exceeds the time delay threshold, the sampling time is marked as a communication abnormal sampling point. The time delay threshold is determined based on the 95th percentile of the arrival time delay of similar detector packets in the past 30 days.

[0075] A determination is made regarding continuous missing data; specifically, when a fire detector fails to receive a valid reported data message at multiple consecutive sampling times, the corresponding sampling time is marked as a communication anomaly sampling point; wherein, the condition of multiple consecutive sampling times is set according to the continuity requirements required for link anomaly identification.

[0076] The device self-test status is analyzed; specifically, when the device self-test status at a certain sampling time is "pass", the sampling time is recorded as the self-test pass sampling point; when the device self-test status at a certain sampling time is "fail", the sampling time is recorded as the self-test failure sampling point.

[0077] After completing message time mapping, communication anomaly determination, and self-test status determination, the fire detector number, sampling time, and detection output value are organized. Among them, sampling points that are not marked as communication anomaly sampling points and self-test anomaly sampling points are recorded as valid sampling points, and the remaining sampling points are recorded as restricted sampling points. Then, a zoned observation sequence is formed according to the fire detector number and sampling time.

[0078] The response construction module performs consistency screening and robust aggregation on the detection output values ​​of the same type of fire detectors at the same sampling time within the same fire compartment based on the valid sampling points in the partition observation sequence. This forms a partition reference response sequence and reference response confidence markers corresponding to each sampling time. The partition reference response sequence is used to provide dynamic reference for subsequent group deviation analysis, and the reference response confidence markers corresponding to each sampling time are used to provide credibility constraints for subsequent accuracy anomaly attribution.

[0079] Consistency screening and robust aggregation are performed on the detection output values ​​of the same type of fire detectors within the same fire compartment at the same sampling time. The specific implementation is as follows:

[0080] Valid sampling points corresponding to the same fire detector type and the same sampling time within the same fire compartment are extracted from the zonal observation sequence, and the detection output values ​​of each valid sampling point are read to form a candidate reference sample set for the corresponding sampling time. Restricted sampling points do not participate in the construction of the candidate reference sample set.

[0081] A consistency screening process is performed on the candidate reference sample set to remove outlier samples that significantly deviate from the common trend of the population. Specifically, the median value of each detection output value in the candidate reference sample set at the corresponding sampling time is first calculated. The larger of the absolute value of the median value and the preset minimum response benchmark value is used as the denominator of the deviation ratio. Then, the absolute value of the difference between each detection output value and the median value is divided by the denominator of the deviation ratio to obtain the deviation ratio of the corresponding detection output value relative to the median value. When the deviation ratio of a certain detection output value is greater than the preset deviation ratio threshold, the corresponding sample is removed. Samples with deviation ratios not greater than the preset deviation ratio threshold are retained to form the screening. The effective sample set is then used. When the number of effective samples in the filtered effective sample set is lower than the preset effective sample lower limit, the sampling time is recorded as the low confidence sampling time. The deviation ratio threshold is set according to the statistical distribution of the deviation ratio of the same type of fire detectors in the fire compartment during the historical stable operation period, preferably determined according to the 95th percentile of the historical deviation ratio sample set. The preset effective sample lower limit is set according to the minimum number of participating samples required to construct the partition reference response. The minimum response benchmark value is determined according to the lower limit of the effective range of the fire detector, which is used to avoid the deviation ratio calculation distortion caused by the median value being zero or close to zero.

[0082] The median value of the detection output values ​​in the filtered valid sample set is taken as the partition reference response value at the corresponding sampling time, and a partition reference response sequence is formed according to the sampling time order.

[0083] A reference response confidence marker is generated for the partition reference response at each sampling time. Specifically, the deviation ratio of each probe output value in the filtered valid sample set relative to its median value is calculated, and the median value of the deviation ratio is used as the sample dispersion at that sampling time. When the number of filtered valid samples meets the preset high-confidence sample quantity condition and the sample dispersion does not exceed the preset high-confidence dispersion threshold, the reference response confidence marker at that sampling time is recorded as high. When the number of filtered valid samples meets the preset medium-confidence sample quantity condition and the sample dispersion does not exceed the preset medium-confidence dispersion threshold, the reference response confidence marker at that sampling time is recorded as medium. All other cases are recorded as low. The preset high-confidence sample quantity condition and the preset medium-confidence sample quantity condition are set according to the partitioned reference response stability requirements and the participation sample coverage requirements. Preferably, the preset high-confidence sample quantity condition is that the number of effective samples after screening is not less than 4, and the preset medium-confidence sample quantity condition is that the number of effective samples after screening is 3. The high-confidence dispersion threshold and the medium-confidence dispersion threshold are set according to the statistical distribution of the historical dispersion sample set. The high-confidence dispersion threshold is preferably taken as the 75th percentile of the historical dispersion sample set, and the medium-confidence dispersion threshold is preferably taken as the 90th percentile of the historical dispersion sample set. The high-confidence dispersion threshold is less than the medium-confidence dispersion threshold.

[0084] After completing candidate reference sample extraction, consistency screening, partition reference response value generation, and confidence tag generation, the partition reference response value and reference response confidence tag corresponding to each sampling time are output together to form the partition reference response sequence and the reference response confidence tag corresponding to each sampling time.

[0085] The deviation generation module calculates the reference difference and relative deviation ratio of each fire detector relative to the partition reference response value at each sampling time based on the partition reference response sequence, reference response confidence flags, and the effective sampling points corresponding to each fire detector in the partition observation sequence. It then performs sliding window direction consistency judgment and credibility correction on the relative deviation ratio to form a group deviation evidence sequence. The group deviation evidence sequence is used to provide external reference evidence for subsequent accuracy anomaly attribution.

[0086] See Figure 2 As shown, the deviation of each fire detector from the zone reference response is calculated, and the specific implementation is as follows:

[0087] For each fire detector, the effective sampling points at each sampling time are extracted, and the detection output value corresponding to the effective sampling point is read; the zone reference response value corresponding to the same sampling time is read; for each sampling time, the detection output value of the fire detector at that sampling time is subtracted from the zone reference response value at the same sampling time to obtain the reference difference value at that sampling time, and the reference difference value is divided by the larger value between the absolute value of the zone reference response value and the minimum response benchmark value to obtain the relative deviation ratio.

[0088] The relative deviation ratio within consecutive sampling times is used to determine directional consistency. Specifically, multiple consecutive sampling points are used as a sliding window, and the directional consistency of the relative deviation ratio within the sliding window is statistically analyzed. When the number of sampling points with negative relative deviation ratios within the same sliding window reaches a preset number of directional consistency points, and the average absolute value of the relative deviation ratios within the sliding window is greater than or equal to a preset deviation threshold, the sliding window is recorded as a negative group deviation window. When the number of sampling points with positive relative deviation ratios within the same sliding window reaches a preset number of directional consistency points, and the average absolute value of the relative deviation ratios within the sliding window is greater than or equal to a preset deviation threshold, the sliding window is recorded as a positive group deviation window. The length of the sliding window is set according to the time smoothing requirements for continuous deviation identification; the number of directional consistency points is set according to the anti-fluctuation requirements for stable deviation identification; and the deviation threshold is set according to the relative deviation ratio distribution of normal fire detectors during stable operation, preferably determined based on the 95th percentile of the relative deviation ratio distribution of normal fire detectors during stable operation over the past 90 days.

[0089] The credibility of negative and positive group deviation windows is corrected by combining the confidence markers of the reference response to form evidence of group deviation. Specifically, the confidence markers of the reference response at corresponding sampling times within each negative and positive group deviation window are statistically analyzed. When the proportion of high-confidence sampling times is not lower than a preset high-confidence proportion threshold, the group deviation window is recorded as high-confidence group deviation evidence. When the proportions of medium-confidence and high-confidence sampling times reach a preset qualified proportion requirement and the proportion of high-confidence sampling times is lower than the preset high-confidence proportion threshold, the group deviation window is recorded as medium-confidence group deviation evidence. All other cases are recorded as weak group deviation evidence. The preset high-confidence proportion threshold and the preset qualified proportion requirement are set according to the reference response stability requirement and the evidence credibility differentiation requirement.

[0090] After completing the relative deviation ratio calculation, direction consistency determination, and credibility correction, the group deviation evidence corresponding to each sliding window is output to form a group deviation evidence sequence. The group deviation evidence includes negative group deviation evidence and positive group deviation evidence, and each group deviation evidence has a corresponding credibility level. Among them, negative group deviation evidence indicates that the detection output of the fire detector is consistently lower than the common response level of the same zone, and positive group deviation evidence indicates that the detection output of the fire detector is consistently higher than the common response level of the same zone.

[0091] The evidence generation module extracts the historical baseline sample set of each fire detector within a historical stable period based on the effective sampling points corresponding to each fire detector in the partitioned observation sequence, establishes a baseline response interval, and forms a drift evidence sequence based on the positional relationship and continuous change of the average sampling window value relative to the baseline response interval. The drift evidence sequence serves as internal historical reference evidence.

[0092] The historical baseline sample set of each fire detector during a historical stable period is extracted, and the baseline response interval is established. The specific implementation is as follows:

[0093] Effective sampling points of target fire detectors within historical stable periods are extracted from the zonal observation sequence and aggregated according to the intraday time period corresponding to the sampling time to form a historical baseline sample set. Specifically, the historical stable period is the historical sampling period during which the corresponding fire detector is not marked as a communication anomaly sampling point or a self-test anomaly sampling point. For each fire detector, historical effective sampling points that are consistent with the intraday time period corresponding to the sampling window are selected as historical baseline samples so that the established baseline can reflect the normal response level of the fire detector in the same time period.

[0094] The self-baseline response interval of the fire detector is established based on the historical baseline sample set. Specifically, when the number of historical baseline samples reaches a preset sample number threshold, the detection output values ​​in the historical baseline sample set are statistically analyzed, and the low quantile statistics representing the lower boundary of the interval and the high quantile statistics representing the upper boundary of the interval are extracted. The normal response range formed by the low quantile statistics and the high quantile statistics is taken as the self-baseline response interval of the fire detector in the corresponding time period. When the number of historical baseline samples is lower than the preset sample number threshold, no self-baseline response interval is formed for the corresponding time period. The self-baseline response interval represents the self-response boundary of the fire detector under normal operating conditions. Preferably, the self-baseline response interval is determined based on the 10th percentile and 90th percentile values ​​of the historical baseline samples. The preset sample number threshold is preferably 20.

[0095] The deviation of the sampling window from the baseline response interval is calculated. Specifically, a sampling window is formed by multiple consecutive sampling points. The average value of the detection output value corresponding to each sampling point within the sampling window is calculated as the average value of the sampling window. The average value of the sampling window is then compared with the baseline response interval. When the average value of the sampling window is lower than the lower boundary of the baseline response interval, the sampling window is recorded as a falling deviation window. When the average value of the sampling window is higher than the upper boundary of the baseline response interval, the sampling window is recorded as a rising deviation window. When the average value of the sampling window is within the baseline response interval, no deviation window is formed. The sampling window length is preferably set to 8 sampling points.

[0096] The changing trend of consecutive sampling windows is determined to form evidence of baseline drift. Specifically, the average values ​​of multiple consecutive sampling windows are arranged in chronological order of sampling time, and the average value of the next sampling window is subtracted from the average value of the previous sampling window to obtain the difference between adjacent windows. When the average values ​​of multiple consecutive sampling windows are all below the lower boundary of the baseline response interval, and the number of negative differences among the differences between adjacent windows reaches a preset trend point number, or when the differences between adjacent windows are all zero and the number of multiple consecutive sampling windows reaches a preset duration window number, evidence of downward drift is formed. When the average values ​​of multiple consecutive sampling windows are all above the upper boundary of the baseline response interval, and the number of positive differences among the differences between adjacent windows reaches a preset trend point number, or when the differences between adjacent windows are all zero and the number of multiple consecutive sampling windows reaches a preset duration window number, evidence of upward drift is formed. In other cases, no evidence of baseline drift is formed.

[0097] The preset number of trend points is determined based on the number of differences between adjacent windows formed by multiple consecutive sampling windows, preferably half of the number of differences between adjacent windows rounded up; the preset number of continuous windows is set according to the anti-accidental fluctuation requirements required for continuous deviation identification, preferably 3 consecutive sampling windows.

[0098] After completing the extraction of historical baseline samples, the establishment of baseline response intervals, the determination of sampling window deviation, and the determination of continuous change trends, the downward drift evidence or upward drift evidence corresponding to each sampling window is output to form a drift evidence sequence. Among them, downward drift evidence indicates that the current detection output of the fire detector is continuously lower than its historical normal level, while upward drift evidence indicates that the current detection output of the fire detector is continuously higher than its historical normal level.

[0099] The anomaly attribution module identifies link untrusted windows, reference untrusted windows, and environmental common transition windows based on the communication anomaly sampling points corresponding to each sampling time in the partitioned observation sequence, the reference response confidence flags corresponding to each sampling time, the group deviation evidence sequence, and the drift evidence sequence. Within sampling windows that do not belong to the link untrusted window, reference untrusted window, and environmental common transition window, it performs attribution fusion on the group deviation evidence and the baseline drift evidence to form a missed anomaly criterion and a false alarm anomaly criterion. The missed anomaly criterion and the false alarm anomaly criterion serve as the basis for determining the accuracy detection result output.

[0100] See Figure 3 As shown, restrictions are imposed on the link untrusted window, the reference untrusted window, and the environment co-transition window. Within sampling windows that do not belong to the link untrusted window, the reference untrusted window, and the environment co-transition window, attribution fusion is performed on group deviation evidence and baseline drift evidence. The specific implementation is as follows:

[0101] The link untrusted window is identified based on the communication anomaly sampling points corresponding to each sampling time in the partitioned observation sequence. Specifically, when there is a communication anomaly sampling point corresponding to a certain fire detector, the corresponding sampling time is marked as a link anomaly time. When multiple consecutive adjacent sampling times are marked as link anomaly times, or when the number of communication anomaly sampling points contained in a single sampling window reaches a preset anomaly point threshold, the corresponding sampling period is recorded as a link untrusted window.

[0102] The preset threshold for the number of abnormal points is set according to the length of the sampling window and the requirements for data integrity in subsequent abnormal judgments. Preferably, the number of communication abnormal sampling points in the sampling window is not less than 2. Preferably, the multiple consecutive adjacent sampling times are marked as link abnormal times.

[0103] The reference untrusted window is identified based on the reference response confidence flag; specifically, when there is a sampling moment in the sampling window where the reference response confidence flag is low, the sampling window is recorded as the reference untrusted window.

[0104] The environmental common transition window is identified based on the synchronous changes of multiple fire detectors within the same fire compartment. Specifically, the output change of each fire detector within the same fire compartment at adjacent sampling times is calculated. When the output change of fire detectors reaching a preset proportion is in the same direction and the absolute value of the output change is not less than a preset change amplitude threshold, the corresponding sampling window is recorded as the environmental common transition window. The environmental common transition window is used to characterize the overall response change caused by ventilation switching, changes in passenger flow concentration, or other common disturbances in the compartment.

[0105] The fire detectors that reach the preset proportion are preferably no less than 80% of the fire detectors in the same fire compartment; the preset change range threshold is preferably determined based on the statistical distribution of the absolute value of the output change at adjacent sampling times during historical stable operation periods, and preferably the 95th percentile value of the statistical distribution.

[0106] Restrictions are imposed on the link untrusted window, the reference untrusted window, and the environmental co-transition window; specifically, when a sampling window belongs to any of the link untrusted window, the reference untrusted window, or the environmental co-transition window, attribution fusion is not performed on the group deviation evidence and the self-baseline drift evidence within that sampling window.

[0107] After limiting the untrusted sampling window, attribution fusion is performed on the population deviation evidence and baseline drift evidence within the trustworthy sampling window. Specifically, when negative population deviation evidence and falling drift evidence exist simultaneously within the same sampling window, a missed anomaly criterion is formed; when positive population deviation evidence and rising drift evidence exist simultaneously within the same sampling window, a false alarm anomaly criterion is formed; in other cases, no anomaly criterion is output.

[0108] After completing the identification of untrusted sampling windows and the fusion of evidence attribution, the system outputs the criteria for missed alarms, the criteria for false alarms, and the information of untrusted sampling windows. The criteria for missed alarms indicate that the fire detector is consistently lower than the zone's common response level and is continuously decreasing compared to its own historical stable state. The criteria for false alarms indicate that the fire detector is consistently higher than the zone's common response level and is continuously increasing compared to its own historical stable state.

[0109] The results output module, based on the missed detection anomaly criteria, false alarm anomaly criteria, and unreliable sampling window information, statistically analyzes the anomalies of each fire detector within the reliable sampling window during the target detection period. Based on the statistical results, it determines the missed detection accuracy misalignment level, false alarm accuracy misalignment level, and accuracy compliance conclusion. The missed detection accuracy misalignment level, false alarm accuracy misalignment level, and accuracy compliance conclusion serve as the basis for cleaning, calibration, repair, or replacement.

[0110] The system statistically analyzes the anomalies of each fire detector within the reliable sampling window during the target detection period and outputs the accuracy misalignment level (missed detection), the accuracy misalignment level (false alarm), and the accuracy pass / fail conclusion. The specific implementation is as follows:

[0111] Based on the criteria for missed anomalies, the criteria for false alarms, and the information on unreliable sampling windows, the reliable sampling windows corresponding to each fire detector within the target detection period are determined. Specifically, sampling windows that do not belong to unreliable sampling windows are determined as reliable sampling windows, and the number of missed anomaly windows and the number of false alarm anomaly windows corresponding to each fire detector in the reliable sampling windows are counted respectively.

[0112] The accuracy inaccuracy levels for missed detection and false alarms are determined based on the proportions of the number of missed detection anomaly windows and the number of false alarms within the trusted sampling windows. Specifically, when the proportion of missed detection anomaly windows to the total number of trusted sampling windows is not less than a first preset proportion, the accuracy inaccuracy level for missed detection is determined to be high; when the proportion is lower than the first preset proportion but not lower than a second preset proportion, the accuracy inaccuracy level for missed detection is determined to be medium; when the proportion is lower than the second preset proportion but there are missed detection anomaly windows, the accuracy inaccuracy level for missed detection is determined to be low. The accuracy inaccuracy level for false alarms is determined in the same way. The first preset proportion and the second preset proportion are comprehensively set based on the distribution of anomaly window proportions during the target detection period, the sensitivity requirements of on-site maintenance, and the need to distinguish between missed detection risks and false alarm risks. Preferably, the first preset proportion is 40% and the second preset proportion is 20%.

[0113] The accuracy compliance conclusion is determined based on the levels of missed detection, false alarms, and the proportion of reliable sampling windows. Specifically, when both the levels of missed detection and false alarms are lower than the preset inaccuracy levels, and the proportion of reliable sampling windows to all sampling windows within the target detection period is not lower than the preset reliable proportion, the accuracy compliance conclusion is determined to be qualified. When the proportion of reliable sampling windows is lower than the preset reliable proportion, the accuracy compliance conclusion is determined to be pending review. When either the level of missed detection or the level of false alarms reaches the preset inaccuracy level, the accuracy compliance conclusion is determined to be inaccurate. The preset reliable proportion is set according to the coverage requirement of reliable sampling windows to all sampling windows within the target detection period. The preset inaccuracy level is set according to the on-site maintenance triggering conditions and risk handling requirements, preferably medium.

[0114] While outputting the accuracy inaccuracy level of missed detection, the accuracy inaccuracy level of false alarm, and the accuracy compliance conclusion, the corresponding trigger evidence fragment is also output. Specifically, the sampling window that forms the missing detection anomaly criterion or the false alarm anomaly criterion is output as the trigger evidence fragment, along with the corresponding sampling time, reference difference, relative deviation ratio, average value of the sampling window, and evidence of falling drift or rising drift.

[0115] After completing the statistics of the reliable sampling window, the classification of anomaly levels, and the accuracy qualification judgment, the following are output for each fire detector: the accuracy failure level of the missed alarm type, the accuracy failure level of the false alarm type, the accuracy qualification conclusion, and the trigger evidence fragment. The accuracy failure level of the missed alarm type is used to characterize the fire detector as having a risk of continuously low response, the accuracy failure level of the false alarm type is used to characterize the fire detector as having a risk of continuously high response, and the accuracy qualification conclusion is used to characterize whether the fire detector meets the online accuracy requirements during the current target detection period.

[0116] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

[0117] In conclusion, the above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A double-dimension accurate detection system for subway fire detector false negative and false positive, characterized in that, include: The observation construction module is used to align and label the validity of data collected by each fire detector in a fire compartment at a uniform sampling time, forming a compartmental observation sequence. Response building module: used to filter and aggregate the probe output values ​​at the same sampling time in the partitioned observation sequence to form a partitioned reference response sequence, and generate a reference response confidence flag for the partitioned reference response at each sampling time; Deviation generation module: used to determine the deviation of each fire detector relative to the zone reference response based on the zone reference response sequence and reference response confidence flag, and to form a group deviation evidence sequence through sliding window direction consistency determination, wherein the group deviation evidence includes negative group deviation evidence and positive group deviation evidence; Evidence generation module: used to extract the historical baseline sample set of each fire detector, establish the baseline response interval, and form a drift evidence sequence based on the deviation of the current sampling window from the baseline response interval. The drift evidence sequence includes falling drift evidence and rising drift evidence. Anomaly attribution module: used to identify link untrusted windows, reference untrusted windows and environmental common transition windows, and to perform attribution fusion on group deviation evidence sequences and drift evidence sequences to form missed anomaly criteria and false alarm anomaly criteria; The method for attributing and fusing group deviation evidence sequences and drift evidence sequences to form false negative and false positive anomaly criteria includes: Within the sampling window that does not belong to the link untrusted window, the reference untrusted window, and the common transition window of the environment, the corresponding window matching is performed on the group deviation evidence sequence and the drift evidence sequence; When both negative population deviation evidence and descent drift evidence exist within the same sampling window, a missed anomaly criterion is formed. When both positive population deviation evidence and rising drift evidence exist within the same sampling window, a false alarm anomaly criterion is formed. In other cases, no exception criteria are output; The results output module is used to determine and output the accuracy inaccuracy level of the missed alarm type, the accuracy inaccuracy level of the false alarm type, and the accuracy qualification conclusion based on the missed alarm type anomaly criteria and the false alarm type anomaly criteria.

2. The subway fire detector false alarm and false negative double dimension accurate detection system according to claim 1, characterized in that, Methods for forming partitioned reference response sequences and reference response confidence tags include: Extract valid sampling points within the same fire compartment, of the same type of fire detector, and at the same sampling time to form a candidate reference sample set; The candidate reference sample set is subjected to consistency screening to form a set of effective samples after screening. The median value of the detection output values ​​in the filtered valid sample set is taken as the partition reference response value at the corresponding sampling time, and a partition reference response sequence is formed according to the sampling time order; Based on the number of samples and the dispersion of the effective sample set after screening, a reference response confidence marker is generated for the corresponding sampling time.

3. The dual-dimensional accurate detection system for missed and false alarms in subway fire detectors according to claim 2, characterized in that, The method for obtaining the effective sample set includes: First, calculate the median value of each detection output value in the candidate reference sample set at the corresponding sampling time, and use the larger of the absolute value of the median value and the preset minimum response benchmark value as the denominator of the deviation ratio. Then, divide the absolute value of the difference between each detection output value and the median value by the denominator of the deviation ratio to obtain the deviation ratio of the corresponding detection output value relative to the median value. When the deviation ratio of the detected output value is greater than the preset deviation ratio threshold, the corresponding sample will be removed. Samples with a deviation ratio not exceeding a preset deviation ratio threshold are retained to form a set of effective samples after screening.

4. The dual-dimensional accurate detection system for missed and false alarms in subway fire detectors according to claim 2, characterized in that, Methods for generating reference response confidence markers include: Calculate the deviation ratio of each detection output value in the effective sample set relative to its median value, and use the median value of the deviation ratio as the sample dispersion at the corresponding sampling time; When the number of valid samples meets the preset high confidence sample number condition and the sample dispersion does not exceed the preset high confidence dispersion threshold, the reference response confidence flag at the corresponding sampling time is marked as high. When the number of valid samples meets the preset confidence sample number condition and the sample dispersion does not exceed the preset confidence dispersion threshold, the reference response confidence flag at the corresponding sampling time is marked as medium. In all other cases, the reference response confidence flag is set to low.

5. The dual-dimensional accurate detection system for missed and false alarms in subway fire detectors according to claim 1, characterized in that, The method for generating the group deviation evidence sequence includes: The reference difference is obtained by subtracting the zone reference response value at the same sampling time from the detection output value of each fire detector at each sampling time. The relative deviation ratio is obtained by dividing the reference difference by the larger of the absolute value of the partitioned reference response value and the minimum response benchmark value. The relative deviation ratio within consecutive sampling times is used to determine the consistency of the sliding window direction, forming a negative group deviation window or a positive group deviation window. The credibility of the negative or positive group deviation window is corrected by combining the reference response confidence flag, thus forming a group deviation evidence sequence.

6. The dual-dimensional accurate detection system for missed and false alarms in subway fire detectors according to claim 5, characterized in that, The method for forming a negative or positive population deviation window includes: When the number of sampling points with negative relative deviation ratios within the same sliding window reaches the preset number of points with consistent direction, and the average absolute value of the relative deviation ratios within the sliding window is not less than the preset deviation threshold, the sliding window is recorded as a negative group deviation window. When the number of sampling points with a positive relative deviation ratio within the same sliding window reaches the preset number of points with the same direction, and the average value of the absolute value of the relative deviation ratio within the sliding window is not less than the preset deviation threshold, the sliding window is recorded as a positive group deviation window.

7. The dual-dimensional accurate detection system for missed and false alarms in subway fire detectors according to claim 1, characterized in that, The method for forming the drift evidence sequence includes: Extract the effective sampling points of the target fire detectors during historical stable periods, and aggregate them according to the intraday time corresponding to the sampling time to form a historical baseline sample set; Statistical analysis was performed on the historical baseline sample set to establish the baseline response interval; The deviation of the average value of the sampling window from the baseline response interval and the continuous changes are determined to form a drift evidence sequence.

8. The dual-dimensional accurate detection system for missed and false alarms in subway fire detectors according to claim 7, characterized in that, The method for establishing the baseline response interval includes: The detection output values ​​in the historical baseline sample set are statistically analyzed, and the low quantile statistics representing the lower boundary of the interval and the high quantile statistics representing the upper boundary of the interval are extracted. The normal response range formed by the low quantile statistics and the high quantile statistics is taken as the self-baseline response interval. When the number of historical baseline samples is less than a preset sample number threshold, a baseline response interval is not established.

9. The dual-dimensional accurate detection system for missed and false alarms in subway fire detectors according to claim 7, characterized in that, Calculate the average value of the detection output value corresponding to each sampling point within multiple consecutive sampling windows, and use it as the average value of the sampling window; The difference between adjacent sampling windows is obtained by subtracting the average value of the sampling window with the later sampling time from the average value of the sampling window with the earlier sampling time. When the average value of multiple consecutive sampling windows is lower than the lower boundary of the baseline response interval, and the number of negative differences in the differences between adjacent windows reaches a preset trend point number or the differences between adjacent windows are all zero, evidence of downward drift is formed. Evidence of upward drift is formed when the average value of multiple consecutive sampling windows is higher than the upper boundary of the baseline response interval, and the number of positive differences in the differences between adjacent windows reaches a preset trend point or the differences between adjacent windows are all zero.

10. The dual-dimensional accurate detection system for missed and false alarms in subway fire detectors according to claim 1, characterized in that, Methods for identifying link untrusted windows, reference untrusted windows, and joint transition windows of the environment include: When there is a communication abnormal sampling point at the corresponding sampling time, and multiple consecutive adjacent sampling times are marked as link abnormal times, or the number of communication abnormal sampling points in a single current sampling window reaches the abnormal point count threshold, the current sampling window is recorded as a link untrusted window. When there is a sampling moment in the current sampling window where the reference response confidence flag is low, the current sampling window is recorded as a reference untrusted window; Calculate the output change of each fire detector in the same fire compartment at adjacent sampling times. When the output change direction is consistent and the absolute value of the output change is not less than the preset change amplitude threshold, the current sampling window is recorded as the common transition window of the environment.

Images