An aviation obstacle light remote state intelligent monitoring method based on big data analysis

Abstract

The application discloses an aviation obstacle light remote state intelligent monitoring method based on big data analysis, which comprises the following steps: collecting multi-source monitoring data, running topology data and historical operation and maintenance data; standardizing preprocessing according to a monitoring time window to generate aviation obstacle light multi-source time sequence data; generating comprehensive state representation and environmental disturbance representation according to the multi-source time sequence data; combining a static running topology graph to generate dynamic topology weight through a common cause inhibition Hawkes process; generating a health degree score through an improved MTAD-GAT model; and generating an alarm level, a maintenance priority and an operation and maintenance suggestion according to the health degree score. Through the introduction of the common cause inhibition Hawkes process and the improved MTAD-GAT model, timely sensing of the running state, early warning of the fault trend, accurate identification of the fault type and intelligent generation of the operation and maintenance decision in the aviation obstacle light remote monitoring scene are realized.

Description

Technical Field

[0001] This invention relates to the field of aviation obstruction light operation monitoring technology, and in particular to a remote intelligent monitoring method for aviation obstruction lights based on big data analysis. Background Technology

[0002] With the increasing number of tall structures such as high-rise buildings, communication towers, wind turbines, power towers, chimneys, and bridges in cities, aviation obstruction lights, as aviation safety warning facilities, play a crucial role in ensuring low-altitude aircraft can identify obstacles and prevent flight accidents. Aviation obstruction lights are typically installed at high locations, widely distributed, and operate in open environments for extended periods, making them susceptible to factors such as power supply fluctuations, light source attenuation, communication anomalies, rain and fog, temperature and humidity changes, and equipment aging. Current operation and maintenance methods largely rely on manual inspections, periodic checks, or simple remote alarms, making it difficult to promptly grasp the actual operating status of distributed aviation obstruction lights. This results in delayed fault detection, high inspection costs, and low maintenance efficiency.

[0003] Existing aviation obstruction light monitoring systems typically determine equipment malfunctions by detecting voltage, current, on / off status, or communication online status. While this method provides basic status monitoring, its judgment criteria are relatively singular and fail to comprehensively reflect the correlation between light intensity attenuation, flickering anomalies, power supply fluctuations, communication instability, and environmental disturbances. In actual operation, a decrease in the intensity of a single light may be caused by light source aging, lamp cover contamination, rain or fog obstruction, or changes in ambient brightness; simultaneous malfunctions of multiple lights on the same power supply line may be related to line faults; and multiple lights going offline under the same communication gateway may be related to communication link failures. Existing technologies struggle to distinguish between synchronous anomalies originating from the same source and actual fault triggering relationships, easily leading to false alarms, missed alarms, and inaccurate fault type identification.

[0004] Furthermore, existing monitoring methods primarily rely on alarms triggered after a fault occurs, lacking in-depth utilization of historical operation and maintenance data such as fault time, fault type, number of fault recurrences, equipment service life, and maintenance intervals. They also lack the ability to uniformly model multi-source monitoring data, operational topology data, environmental disturbance information, and historical operation and maintenance data, making it difficult to identify true fault propagation trends and equipment degradation trends. Even when some systems can output alarm information, they typically only indicate equipment abnormalities, failing to combine factors such as health status, fault trends, installation height, building location, power supply lines, and communication gateways to generate maintenance priorities and recommendations, resulting in a lack of data support for operation and maintenance decisions.

[0005] Therefore, how to provide a remote intelligent monitoring method for the status of aviation obstruction lights based on big data analysis is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0006] One objective of this invention is to propose a remote intelligent monitoring method for the status of aviation obstruction lights based on big data analysis. This invention fully utilizes multi-source time-series data fusion, common-cause suppression Hawkes process, and improved MTAD-GAT model. It details the implementation process of remote acquisition of aviation obstruction light operating status, suppression of synchronous anomalies from the same source, fault type identification, health assessment, and maintenance priority generation. It has the advantages of timely status perception, accurate fault identification, low false alarm and missed alarm rates, early fault warning, intelligent operation and maintenance decision-making, and low maintenance costs.

[0007] A remote intelligent monitoring method for the status of aviation obstruction lights based on big data analysis according to an embodiment of the present invention includes the following steps: Collect multi-source monitoring data, operational topology data, and historical maintenance data of aviation obstruction lights; The multi-source monitoring data, operational topology data, and historical operation and maintenance data are standardized and preprocessed to generate multi-source time-series data of aviation obstruction lights; The operating characteristics, communication status characteristics, and environmental impact characteristics of aviation obstruction lights are extracted from multi-source time-series data. The operating characteristics and communication status characteristics of the lights are fused to generate a comprehensive status characterization, and the environmental impact characteristics are encoded to generate an environmental disturbance characterization. A static operational topology map is constructed based on the operational topology data. Anomaly occurrence time, device identifier, fault type marker, and topology attribution marker are extracted from historical operation and maintenance data and environmental disturbance characterization to form an anomaly event sequence. The contribution of synchronous anomaly triggering from the same source is weakened through the common cause suppression Hawkes process. The fault triggering intensity is calculated and converted into dynamic topology weights. The comprehensive state representation, environmental disturbance representation, static operating topology graph and dynamic topology weight are input into the improved MTAD-GAT model. The attention of nodes is allocated through the topology constraint graph attention layer, the contribution of abnormal residuals is corrected through the environmental disturbance correction layer, and the health score is generated by fusing abnormal scores, fault types and fault trends through the health decision layer. Based on health scores, historical operation and maintenance data, and operational topology data, alarm levels, repair priorities, and operation and maintenance recommendations are generated, and the repair results are written into historical operation and maintenance data.

[0008] Optionally, the multi-source monitoring data includes electrical operation data, optical status data, flicker control data, communication status data, and environmental impact data. The electrical operation data includes voltage, current, and power; the optical status data includes light intensity and on / off status; the flicker control data includes flicker frequency and flicker period; the communication status data includes online status, signal strength, data packet loss rate, and communication delay; and the environmental impact data includes ambient brightness, temperature, humidity, rain / fog conditions, and visibility. Optionally, the operational topology data includes the identification of the aviation obstruction light equipment, its installation location, installation height, the building to which it belongs, the controller to which it belongs, the power supply line to which it belongs, the communication gateway to which it belongs, and the identification of adjacent aviation obstruction light equipment. Optionally, the historical operation and maintenance data includes historical failure time, failure type, repair time, repair measures, replaced parts, number of failure recurrences, equipment commissioning years and maintenance intervals; Optionally, the generation of multi-source time-series data for aviation obstruction lights includes: Multi-source monitoring data are time-mapped according to equipment identification and collection timestamp, mapping electrical operation data, optical status data, flashing control data, communication status data and environmental impact data to the same monitoring time window, and linking operation topology data and historical operation and maintenance data to the corresponding aviation obstruction lights; The time-mapped multi-source monitoring data is processed for integrity and outlier handling. The missing monitoring items are supplemented by the valid sampled values ​​within adjacent monitoring time windows of the same aviation obstruction light, the valid sampled values ​​of the same power supply line and the valid sampled values ​​of the same communication gateway. Abnormal monitoring items that exceed the working range of the equipment and the range of the acquisition terminal are removed. The multi-source monitoring data, which has undergone integrity processing and outlier processing, the associated operational topology data, and the associated historical operation and maintenance data are quantified and normalized. Sliding window samples are generated according to the aviation obstruction light equipment identification, monitoring time window, operational topology data, and historical operation and maintenance data to obtain multi-source time-series data of aviation obstruction lights.

[0009] Optionally, the fusion of luminaire operating characteristics and communication status characteristics generates a comprehensive state characterization, and the encoding of environmental impact characteristics generates an environmental disturbance characterization, including: Based on the identification of aviation obstruction lights and the monitoring time window, extract electrical operation data, optical status data, flashing control data, communication status data and environmental impact data from the multi-source time-series data of aviation obstruction lights; Statistical features, rate of change features, and periodic consistency features are extracted from electrical operation data, optical status data, and flicker control data within a window to generate luminaire operation features; Extract online continuous features, signal strength change features, packet loss change features, and communication delay change features from communication status data to generate communication status features; Environmental impact features are generated by extracting ambient light variation characteristics, temperature and humidity variation characteristics, rain and fog characteristics, and visibility variation characteristics from environmental impact data. The operating characteristics and communication status characteristics of the lighting fixtures are fused and encoded to generate a comprehensive status representation, and the environmental impact characteristics are independently encoded to generate an environmental disturbance representation.

[0010] Optionally, the step of weakening the contribution of homogeneous synchronization anomaly triggering through the common-cause suppression Hawkes process, calculating the fault triggering strength, and converting it into dynamic topology weights includes: Based on the operational topology data, establish an aviation obstruction light node table and a topology edge table, configure a node number for each aviation obstruction light device identifier, and write nodes that match each other in the topology edge table, such as those belonging to the same building, the same controller, the same power supply line, the same communication gateway, or adjacent aviation obstruction light device identifiers, to generate a static operational topology diagram. Read historical fault time, fault type, aviation obstruction light equipment identification and fault recurrence number, match the building, controller, power supply line and communication gateway to which the aviation obstruction light equipment belongs according to the aviation obstruction light equipment identification, match environmental disturbance characteristics according to historical fault time, configure fault type mark, topology attribution mark and environmental disturbance mark for each historical fault record, and form an abnormal event sequence in ascending order of historical fault time; Event triggering channels are divided according to fault type marking and topology attribution marking. Within each event triggering channel, homogeneous synchronous anomalies are identified, and the anomaly event number, target anomaly event number, and homogeneous synchronous anomaly identification results are written into the homogeneous synchronous anomaly marking table. The homogeneous synchronous anomaly is an anomaly event in which the topology attribution marking is consistent within the same monitoring time window or adjacent monitoring time windows and the environmental disturbance marking reaches the similarity condition formed by the statistical representation of historical environmental disturbances. Read the same-source synchronization anomaly labeling table, and configure the basic triggering parameters, time decay parameters, fault type consistency coefficient, topology affiliation consistency coefficient, node connection coefficient and common cause suppression coefficient during the common cause suppression Hawkes process. Configure the common cause suppression coefficient lower than the benchmark value for the anomaly event number and the target anomaly event number identified as same-source synchronization anomaly in the same-source synchronization anomaly labeling table, and configure the common cause suppression coefficient of the benchmark value for the anomaly event number and the target anomaly event number identified as non-same-source synchronization anomaly. For the target abnormal event, select the preceding abnormal event whose historical failure time is earlier than the target abnormal event and is located in the same event triggering channel, and obtain the time interval decay value by multiplying the negative time decay parameter by the time interval with the natural constant as the base. The triggering contribution of preceding abnormal events to the target abnormal event is calculated based on the time interval decay value, fault type consistency coefficient, topology affiliation consistency coefficient, node connection coefficient, and common cause suppression coefficient. The triggering contributions corresponding to the same target abnormal event are summed to obtain the fault triggering intensity. The fault trigger intensity is normalized to obtain a dynamic weight value, and the dynamic weight value is written into the corresponding node connection relationship in the static running topology graph to generate dynamic topology weight.

[0011] Optionally, the steps of allocating node attention through the topology constraint graph attention layer, correcting abnormal residual contributions through the environmental disturbance correction layer, and generating a health score by fusing abnormal scores, fault types, and fault trends through the health decision layer include: An improved MTAD-GAT model is constructed, which includes a topological constraint graph attention layer, a temporal prediction reconstruction layer, an environmental disturbance correction layer, an anomaly scoring layer, and a health decision layer. The topology constraint graph attention layer reads the comprehensive state representation according to the aviation obstruction light equipment identification, determines the set of adjacent nodes according to the static operating topology graph, calculates the state similarity between nodes, uses dynamic topology weights to weight and normalize the state similarity to obtain the node attention coefficient, and aggregates the comprehensive state representation of adjacent nodes according to the node attention coefficient to generate topology-enhanced state features. The temporal prediction and reconstruction layer arranges the topology-enhanced state features according to the monitoring time window order and performs temporal dependency encoding to generate state prediction representation and state reconstruction representation. The difference between each of the two representations is calculated with the comprehensive state representation to obtain the prediction residual and reconstruction residual. The environmental disturbance correction layer reads the environmental disturbance characterization, extracts the environmental light influence item, temperature and humidity influence item, rain and fog influence item and visibility influence item, divides the prediction residual and reconstruction residual into electrical residual, optical residual, scintillation residual and communication residual, generates residual correction coefficients and corrects various residuals, and generates environmental correction residuals; The anomaly scoring layer performs weighted fusion of environmental correction residuals, prediction residuals, and reconstruction residuals to generate anomaly scores. It generates fault trends based on the direction, magnitude, and duration of change of anomaly scores within a continuous monitoring time window, and generates fault types based on the matching results of anomaly scores and fault types in historical operation and maintenance data. The health assessment decision layer reads abnormal scores, fault types, fault trends, fault recurrence times, equipment service life, and maintenance intervals, and generates a health assessment score through normalized coding and weighted fusion. The training samples are formed by selecting the monitoring time window corresponding to the historical failure time and the previous monitoring time window. Failure type labels are generated by reading the failure type in the historical operation and maintenance data. Health score supervision items are generated according to maintenance time, failure recurrence number and maintenance interval. The weighted sum of state prediction bias, state reconstruction bias, failure type classification bias and health score bias is calculated to obtain the training target. Backpropagation is used to update and improve the MTAD-GAT model parameters.

[0012] Optionally, generating alarm levels, maintenance priorities, and operational recommendations, and writing the maintenance results into historical operational data includes: Read the health score, fault type and fault trend output by the improved MTAD-GAT model, perform interval normalization on the health score, and generate corresponding alarm levels according to the alarm level to which the health score belongs. The alarm levels include emergency, severe, warning, attention and normal. The alarm levels are maintained at the same level or adjusted towards emergency according to the fault trend. The system reads the number of fault recurrences, equipment service life, and maintenance intervals from historical operation and maintenance data. It also reads the installation height, building, power supply line, communication gateway, and adjacent aviation obstruction light equipment identifiers from the operation topology data. The system then sums the scores corresponding to the alarm level, the number of fault recurrences, the equipment service life, the maintenance interval, and the operation topology data to generate maintenance priorities. Maintenance recommendations are generated based on fault type, alarm level, and maintenance priority. These recommendations include suggestions for lighting fixture inspection, power supply inspection, communication inspection, environmental verification, and component replacement. The system receives maintenance time, maintenance measures, replaced components, and fault recurrence confirmation results and writes them into historical maintenance data.

[0013] The beneficial effects of this invention are: This invention proposes a remote intelligent monitoring method for the status of aviation obstruction lights based on big data analysis. By constructing a multi-source monitoring data acquisition system, standardized preprocessing, status feature fusion, static operational topology modeling, dynamic topology weight generation using the Hawkes process for common-cause suppression, and an improved MTAD-GAT model intelligent diagnostic process, it effectively enhances the accuracy of remote status perception, fault identification, and trend early warning for aviation obstruction lights. Compared to traditional monitoring methods relying on manual inspections or single-parameter alarms, this invention comprehensively utilizes electrical operation data, optical status data, flashing control data, communication status data, environmental impact data, operational topology data, and historical maintenance data to perform continuous, remote, and multi-dimensional analysis of the operational status of aviation obstruction lights. This avoids false alarms and missed alarms caused by simple judgments based solely on voltage, current, on / off status, or communication online status, thus improving the real-time performance and completeness of distributed aviation obstruction light operation monitoring.

[0014] This invention extracts operational characteristics, communication status characteristics, and environmental impact characteristics from multi-source time-series data of aviation obstruction lights to generate comprehensive status and environmental disturbance characteristics. Combined with operational topology data, a static operational topology map is constructed, enabling the monitoring system to distinguish between single-light faults, anomalies on the same power supply line, anomalies on the same communication gateway, and environmental disturbance anomalies. In particular, by using a common-cause suppression Hawkes process to analyze the triggering relationships of anomalous event sequences, it weakens the triggering contribution of synchronous anomalies caused by the same environmental disturbance, the same power supply line, or the same communication gateway, reducing the interference of common-cause anomalies on fault triggering intensity calculation. This allows the dynamic topology weights to more accurately reflect the true fault correlations between aviation obstruction lights, improving the reliability of fault propagation trend identification and fault type judgment.

[0015] This invention improves the MTAD-GAT model by jointly analyzing comprehensive state representation, environmental disturbance representation, static operating topology diagram, and dynamic topology weights. It extracts operational correlation features between aviation obstruction lights using the topology constraint graph attention layer, corrects abnormal residual contributions using the environmental disturbance correction layer, and generates a health score by fusing abnormal scores, fault types, and fault trends using the health decision layer. This achieves a shift from post-fault alarm to pre-fault prediction and intelligent maintenance. This method can identify degradation trends such as light intensity attenuation, current fluctuations, flickering anomalies, frequent communication anomalies, and equipment aging in advance. It also generates alarm levels, maintenance priorities, and maintenance recommendations by combining historical faults, maintenance records, fault recurrence counts, equipment service life, and maintenance intervals. This invention offers advantages such as timely remote state perception, accurate fault identification, strong suppression of synchronous anomalies, early fault trend warning, sufficient basis for maintenance scheduling, and low maintenance costs. It is applicable to intelligent operation and maintenance scenarios for aviation obstruction lights in tall structures such as high-rise buildings, communication towers, wind turbine towers, power towers, chimneys, and bridges. Attached Figure Description

[0016] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings: Figure 1 This is a flowchart of a remote intelligent monitoring method for the status of aviation obstruction lights based on big data analysis proposed in this invention; Figure 2 This is a flowchart illustrating the generation of dynamic topological weights based on the Hawkes process of common cause suppression, a remote intelligent monitoring method for aviation obstruction lights based on big data analysis proposed in this invention. Figure 3 This is a schematic diagram of the improved MTAD-GAT model of the remote intelligent monitoring method for aviation obstruction lights based on big data analysis proposed in this invention. Detailed Implementation

[0017] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0018] refer to Figure 1 , Figure 2 and Figure 3 A remote intelligent monitoring method for the status of aviation obstruction lights based on big data analysis includes the following steps: Collect multi-source monitoring data, operational topology data, and historical maintenance data of aviation obstruction lights; The multi-source monitoring data, operational topology data, and historical operation and maintenance data are standardized and preprocessed to generate multi-source time-series data of aviation obstruction lights; The operating characteristics, communication status characteristics, and environmental impact characteristics of aviation obstruction lights are extracted from multi-source time-series data. The operating characteristics and communication status characteristics of the lights are fused to generate a comprehensive status characterization, and the environmental impact characteristics are encoded to generate an environmental disturbance characterization. A static operational topology map is constructed based on the operational topology data. Anomaly occurrence time, device identifier, fault type marker, and topology attribution marker are extracted from historical operation and maintenance data and environmental disturbance characterization to form an anomaly event sequence. The contribution of synchronous anomaly triggering from the same source is weakened through the common cause suppression Hawkes process. The fault triggering intensity is calculated and converted into dynamic topology weights. The comprehensive state representation, environmental disturbance representation, static operating topology graph and dynamic topology weight are input into the improved MTAD-GAT model. The attention of nodes is allocated through the topology constraint graph attention layer, the contribution of abnormal residuals is corrected through the environmental disturbance correction layer, and the health score is generated by fusing abnormal scores, fault types and fault trends through the health decision layer. Based on health scores, historical operation and maintenance data, and operational topology data, alarm levels, repair priorities, and operation and maintenance recommendations are generated, and the repair results are written into historical operation and maintenance data.

[0019] In this embodiment, the multi-source monitoring data includes electrical operation data, optical status data, flicker control data, communication status data, and environmental impact data. The electrical operation data includes voltage, current, and power. The optical status data includes light intensity and on / off status. The flicker control data includes flicker frequency and flicker period. The communication status data includes online status, signal strength, data packet loss rate, and communication delay. The environmental impact data includes ambient brightness, temperature, humidity, rain / fog conditions, and visibility.

[0020] In this embodiment, the operational topology data includes the identification of the aviation obstruction light equipment, its installation location, installation height, the building to which it belongs, the controller to which it belongs, the power supply line to which it belongs, the communication gateway to which it belongs, and the identification of adjacent aviation obstruction light equipment.

[0021] In this embodiment, the historical operation and maintenance data includes historical fault time, fault type, repair time, repair measures, replaced parts, number of fault recurrences, equipment commissioning years, and maintenance interval.

[0022] In this embodiment, the generation of multi-source time-series data for aviation obstruction lights includes: Multi-source monitoring data is time-mapped according to equipment identification and acquisition timestamps, mapping electrical operation data, optical status data, flashing control data, communication status data, and environmental impact data to the same monitoring time window. Furthermore, operational topology data and historical maintenance data are associated with corresponding aviation obstruction lights. Specifically, the association of operational topology data and historical maintenance data with corresponding aviation obstruction lights is as follows: A data index table is established based on the identification of aviation obstruction light equipment. The installation location, installation height, building, controller, power supply line, communication gateway, and adjacent aviation obstruction light equipment identification are read from the operational topology data. The operational topology data under the same equipment identification is written into the monitoring time window corresponding to the same equipment identification. The historical fault time, fault type, maintenance time, maintenance measures, replaced parts, number of fault recurrences, equipment commissioning years, and maintenance interval are read from the historical operation and maintenance data. Equipment matching is completed according to the aviation obstruction light equipment identification, and time matching is completed according to the time inclusion relationship between the historical fault time and the monitoring time window. An aviation obstruction light window sample with operational topology data and historical operation and maintenance data is generated. The time-mapped multi-source monitoring data is processed for integrity and outlier handling. The missing monitoring items are supplemented by the valid sampled values ​​within adjacent monitoring time windows of the same aviation obstruction light, the valid sampled values ​​of the same power supply line and the valid sampled values ​​of the same communication gateway. Abnormal monitoring items that exceed the working range of the equipment and the range of the acquisition terminal are removed. Multi-source monitoring data (after integrity and outlier processing), correlated operational topology data, and correlated historical maintenance data are quantized and normalized. Sliding window samples are generated according to the aviation obstruction light equipment identifier, monitoring time window, operational topology data, and historical maintenance data to obtain multi-source time-series data for aviation obstruction lights. Specifically, the generation of sliding window samples according to the aviation obstruction light equipment identifier, monitoring time window, operational topology data, and historical maintenance data yields the following multi-source time-series data for aviation obstruction lights: Data sequences are divided by the identification of aviation obstruction lights. Multi-source monitoring data with normalized encoding are arranged in chronological order according to the monitoring time window. The operational topology data and historical maintenance data corresponding to the same equipment identification are written into the corresponding monitoring time window. Multi-source monitoring data, operational topology data and historical maintenance data within a continuous monitoring time window are extracted with a fixed window length to form a sliding window sample of a single aviation obstruction light. Data sequences corresponding to different aviation obstruction light equipment identifications are traversed according to the same window length and sliding step size to form multi-source time-series data of aviation obstruction lights.

[0023] In this embodiment, the generation of a comprehensive state characterization by integrating the luminaire's operating characteristics and communication status characteristics, and the generation of an environmental disturbance characterization by encoding environmental impact characteristics, include: Based on the identification of aviation obstruction lights and the monitoring time window, extract electrical operation data, optical status data, flashing control data, communication status data and environmental impact data from the multi-source time-series data of aviation obstruction lights; Statistical features, rate of change features, and periodic consistency features within a window are extracted from electrical operation data, optical status data, and flicker control data to generate luminaire operation features. Specifically, the extraction of statistical features, rate of change features, and periodic consistency features within a window to generate luminaire operation features involves: According to the identification of aviation obstruction lights and the monitoring time window, read electrical operation data, optical status data and flashing control data, calculate the window mean, maximum value, minimum value, range, standard deviation and effective sampling percentage for voltage, current, power, light intensity, on / off state, flashing frequency and flashing period respectively, and form statistical characteristics within the window; The voltage change amplitude, current change amplitude, power change amplitude, light intensity change amplitude, and scintillation frequency change amplitude are calculated based on the numerical differences between adjacent sampling times within the monitoring time window. The initial sampling value and the final sampling value within the monitoring time window are read. The sign of the difference obtained by subtracting the initial sampling value from the final sampling value is determined. When the difference is positive, it is recorded as an upward direction; when the difference is negative, it is recorded as a downward direction; and when the difference is zero, it is recorded as a stable direction. The ratio of the absolute value of the difference to the duration of the monitoring time window is recorded as the overall change amplitude, forming a change rate feature. The actual number of flashes and the adjacent switching interval are counted according to the switching time of the on / off state. The difference between the actual number of flashes and the corresponding number of flashes at the flash frequency is recorded as the flash number deviation, and the difference between the adjacent switching interval and the flash period is recorded as the period interval deviation. The period stability is generated according to the dispersion of the adjacent switching interval. The flash number deviation and the period interval deviation are normalized with the corresponding number of flashes at the flash frequency and the flash period, respectively. The normalization result is subtracted from 1 to obtain the consistency component whose value decreases as the deviation increases. The consistency component and the period stability are encoded and combined in the order of the fields to form the period consistency feature. The statistical characteristics, rate of change characteristics, and periodic consistency characteristics within the window are coded and combined according to the aviation obstruction light equipment identification and the monitoring time window to generate the light fixture operation characteristics; Online persistence features, signal strength variation features, packet loss variation features, and communication delay variation features are extracted from communication status data to generate communication status features. Specifically, the generation of communication status features includes: According to the identification of aviation obstruction light equipment and the monitoring time window, the communication status data is read, the online status is continuously statistically analyzed, and the number of online samplings, the number of offline samplings, the continuous online duration and the continuous offline duration within the monitoring time window are recorded to generate online continuous characteristics; The signal strength is calculated according to the sampling time sequence, including the window mean, maximum value, minimum value, standard deviation, and the intensity difference between adjacent sampling times, to generate signal strength variation characteristics; The packet loss rate is calculated according to the sampling time sequence, including the window mean, peak value, fluctuation range, and the difference in packet loss rate between adjacent sampling times, to generate packet loss change characteristics. The communication delay is calculated according to the sampling time order, including the window mean, maximum delay, minimum delay, delay standard deviation, and delay difference between adjacent sampling times, to generate communication delay variation characteristics. The online continuous characteristics, signal strength change characteristics, packet loss change characteristics, and communication delay change characteristics are written into the same communication characteristic field group according to the aviation obstruction light equipment identification and monitoring time window to generate communication status characteristics; Environmental impact data are extracted to extract features of ambient light variation, temperature and humidity variation, rain and fog conditions, and visibility variation, generating environmental impact features. Specifically, the generation of environmental impact features includes: According to the identification of aviation obstruction light equipment and monitoring time window, environmental impact data is read. The ambient brightness is calculated according to the sampling time sequence, including window mean, maximum value, minimum value, standard deviation, brightness difference between adjacent sampling times, and brightness difference at the beginning and end of the monitoring time window, to generate ambient light change characteristics. The mean, maximum, minimum, standard deviation, variation range between adjacent sampling times, and variation range at the start and end of the monitoring time window are calculated for temperature and humidity respectively to generate temperature and humidity variation characteristics; The rain and fog status characteristics are generated by statistically monitoring the number of rain and fog samples, the number of no-rain and fog samples, the duration of rain and fog, the number of rain and fog status switching, and the continuous duration of rain and fog status within the rain and fog status monitoring time window. The visibility is calculated according to the sampling time sequence, including the window mean, minimum value, maximum value, standard deviation, visibility difference between adjacent sampling times, and visibility difference at the beginning and end of the monitoring time window, to generate visibility change characteristics; The characteristics of changes in ambient light, temperature and humidity, rain and fog, and visibility are written into the same environmental characteristic field group according to the aviation obstruction light equipment identification and monitoring time window to generate environmental impact characteristics. The luminaire operating characteristics and communication status characteristics are fused and encoded to generate a comprehensive status representation, while environmental impact characteristics are independently encoded to generate an environmental disturbance representation. Specifically: According to the aviation obstruction light equipment identification and monitoring time window, read the light operation characteristics, communication status characteristics and environmental impact characteristics, configure the feature field order for the light operation characteristics and communication status characteristics respectively, and concatenate the fields of the light operation characteristics and communication status characteristics under the same aviation obstruction light equipment identification and the same monitoring time window to form a status fusion characteristic; Normalize and linearly map the state fusion features to map state fusion features of different field lengths to a unified representation dimension to generate a comprehensive state representation. The environmental impact characteristics, including ambient light variation, temperature and humidity variation, rain and fog conditions, and visibility variation, are arranged and combined in the order of the fields to form combined environmental impact characteristics. These combined environmental impact characteristics are then normalized, calibrated, and linearly mapped to generate an environmental disturbance characterization. The comprehensive state characterization and environmental disturbance characterization are established in correspondence according to the aviation obstruction light equipment identification and monitoring time window.

[0024] In this embodiment, the step of weakening the contribution of homogeneous synchronization anomaly triggering through the common-cause suppression Hawkes process, calculating the fault triggering strength, and converting it into dynamic topology weights includes: Based on the operational topology data, an aviation obstruction light node table and a topology edge table are established. A node number is assigned to each aviation obstruction light device identifier. Nodes belonging to the same building, controller, power supply line, communication gateway, or adjacent aviation obstruction light device identifiers that match are written into the topology edge table. A static operational topology diagram is generated, specifically as follows: The operational topology data is read according to the identification of the aviation obstruction light equipment. A unique node number is assigned to each aviation obstruction light equipment identification, and the node number, installation location, installation height, building to which it belongs, controller to which it belongs, power supply line to which it belongs, communication gateway to which it belongs, and adjacent aviation obstruction light equipment identifications are written into the aviation obstruction light node table. The aviation obstruction light node table is matched according to the building to which it belongs, controller to which it belongs, power supply line to which it belongs, communication gateway to which it belongs, and adjacent aviation obstruction light equipment identifications correspond to each other. When the matching results meet the conditions of mutual correspondence of the same building to which it belongs, controller to which it belongs, power supply line to which it belongs, communication gateway to which it belongs, or adjacent aviation obstruction light equipment identifications, the topology edge table is written between the corresponding node numbers. A node connection matrix is ​​generated according to the node number connection relationship in the topology edge table, and the node connection matrix is ​​mapped to the aviation obstruction light node table to obtain the static operational topology diagram. The system reads historical fault times, fault types, aviation obstruction light equipment identifiers, and fault recurrence counts. It matches the associated building, controller, power supply line, and communication gateway based on the aviation obstruction light equipment identifier. It also matches environmental disturbance characteristics based on historical fault times. For each historical fault record, it assigns fault type, topology, and environmental disturbance tags, and then forms an abnormal event sequence in ascending order of historical fault time. Specifically, the process of matching environmental disturbance characteristics based on historical fault times and assigning fault type, topology, and environmental disturbance tags to each historical fault record, followed by forming an abnormal event sequence in ascending order of historical fault time, is as follows: Read the monitoring time window corresponding to the historical fault record according to the aviation obstruction light equipment identification, and locate the corresponding monitoring time window by using the time inclusion relationship between the historical fault time and the monitoring time window. Within the corresponding monitoring time window, read the environmental disturbance characteristics under the same aviation obstruction light equipment identifier, extract the characteristics of ambient light change, temperature and humidity change, rain and fog condition characteristics and visibility change characteristics, and encode them to form an environmental disturbance marker; A fault type tag is generated based on the field value of the fault type, and a topology attribution tag is generated based on the matching results of the building, controller, power supply line and communication gateway to which it belongs. The historical fault time, aviation obstruction light equipment identification, fault type tag, topology attribution tag, environmental disturbance tag and fault recurrence number are written into the same abnormal event record, and the abnormal event records are arranged in ascending order according to the historical fault time to form an abnormal event sequence. Event triggering channels are divided according to fault type and topology attribution. Within each event triggering channel, synchronous anomalies from the same source are identified, and the anomaly event number, target anomaly event number, and synchronous anomaly identification result are written into a synchronous anomaly labeling table. A synchronous anomaly from the same source is defined as an anomaly where the topology attribution is consistent within the same or adjacent monitoring time windows, and the environmental disturbance label meets the similarity criteria formed by statistical analysis of historical environmental disturbance characteristics. Specifically, the event triggering channels are divided according to fault type and topology attribution as follows: Read each abnormal event record in the abnormal event sequence and extract the abnormal event number, historical failure time, aviation obstruction light equipment identifier, failure type marker, and topology attribution marker; combine the failure type marker and topology attribution marker to generate an event trigger channel number; write abnormal event records with consistent combined codes into the same event trigger channel and sort them in ascending order according to historical failure time; establish an event index table for abnormal event numbers within the same event trigger channel to form an event trigger channel jointly defined by the failure type marker and topology attribution marker; Identify synchronization anomalies originating from the same source within each event-triggered channel, specifically as follows: Read abnormal event records within the same event triggering channel and compare them in groups according to the monitoring time window; for abnormal event records located within the same or adjacent monitoring time windows, read the topology attribution marker and environmental disturbance marker; when the topology attribution marker is consistent, and the changes in ambient light, temperature and humidity, rain and fog conditions, and visibility between the environmental disturbance markers meet the similarity conditions formed by the statistical analysis of historical environmental disturbance characteristics, the corresponding abnormal event record is identified as a synchronous anomaly from the same source; write the preceding abnormal event number, the target abnormal event number, and the synchronous anomaly identification result into the synchronous anomaly marker table from the same source. The similarity conditions formed by the statistical representation of historical environmental disturbances are as follows: Historical environmental disturbances were grouped according to their respective buildings, power supply lines, and communication gateways. The differences in ambient light, temperature and humidity, visibility, and rain / fog conditions within each group were statistically analyzed. Similarity boundaries were determined based on the historical concentrated distribution intervals of these differences. Conditions for consistency in rain / fog conditions were determined based on the consistency in rain / fog conditions. The similarity boundaries and conditions for consistency in rain / fog conditions were combined to form similarity conditions. Read the homogeneous synchronization anomaly labeling table. During the Hawkes common cause suppression process, configure the basic trigger parameters, time decay parameters, fault type consistency coefficient, topology affiliation consistency coefficient, node connection coefficient, and common cause suppression coefficient. For anomaly event numbers identified as homogeneous synchronization anomalies in the homogeneous synchronization anomaly labeling table, configure a common cause suppression coefficient lower than the baseline value for the target anomaly event number. For anomaly event numbers identified as non-homogeneous synchronization anomalies, configure a common cause suppression coefficient at the baseline value for the target anomaly event number. Specifically, the configuration of the basic trigger parameters, time decay parameters, fault type consistency coefficient, topology affiliation consistency coefficient, node connection coefficient, and common cause suppression coefficient during the Hawkes common cause suppression process is as follows: Read the sequence of abnormal events according to the event triggering channel, count the number of abnormal events in the same event triggering channel, read the earliest and latest historical fault times, calculate the time span between the latest and earliest historical fault times, divide the number of abnormal events by the time span, and obtain the basic triggering parameters. Arrange the abnormal events in the same event triggering channel in ascending order of historical fault time, calculate the time interval between adjacent abnormal events one by one, extract the median of the time interval, and divide 1 by the median to obtain the time decay parameter. Read the fault type flags of the preceding abnormal event and the target abnormal event. When the two fault type flags are the same, the fault type consistency coefficient is 1; when the two fault type flags are different, the fault type consistency coefficient is 0. Read the topology attribution tags of the preceding abnormal event and the target abnormal event, compare the building, controller, power supply line and communication gateway to which they belong, count the number of consistent items, divide the number of consistent items by the total number of topology attribution tags, and obtain the topology attribution consistency coefficient. Read the node connection matrix in the static running topology graph. When there is a node connection relationship between the node corresponding to the preceding abnormal event and the node corresponding to the target abnormal event, the node connection coefficient is 1. When there is no node connection relationship between the node corresponding to the preceding abnormal event and the node corresponding to the target abnormal event, the node connection coefficient is 0. Read the same-source synchronous anomaly marker table. When the identification result corresponding to the preceding anomaly number and the target anomaly number is a non-same-source synchronous anomaly, the common cause suppression coefficient is set to 1. When the identification result corresponding to the preceding anomaly number and the target anomaly number is a same-source synchronous anomaly, read the changes in ambient light, temperature and humidity, rain and fog conditions, and visibility from the corresponding environmental disturbance markers. Calculate the differences in ambient light, temperature and humidity, and visibility between the preceding anomaly and the target anomaly, and average each difference by the corresponding similarity boundary to obtain the environmental difference value. When the rain and fog conditions are consistent, the environmental difference value is reduced by the correction amount obtained from the statistical analysis of historical samples with consistent rain and fog conditions. When the rain and fog conditions are inconsistent, the environmental difference value remains unchanged. Count the number of anomalies belonging to the same event triggering channel within the same monitoring time window or adjacent monitoring time windows. Divide the counted number by the total number of anomalies within the same event triggering channel to obtain the time setpoint. Average the environmental difference value with the result of 1 minus the time setpoint to obtain a common cause suppression coefficient less than 1. For the target anomaly, select preceding anomalies whose historical fault times are earlier than the target anomaly and are located within the same event triggering channel. Calculate the time interval decay value by multiplying the time interval by the negative time decay parameter with the natural constant as the base. Specifically, the time interval decay value is as follows: Read the anomaly number, historical fault time, fault type marker, and topology attribution marker of the target anomaly event. Locate the event triggering channel to which the target anomaly event belongs based on the fault type marker and topology attribution marker. Within the same event triggering channel, filter anomaly event records whose historical fault time is earlier than the target anomaly event to obtain the preceding anomaly event. Read the historical fault time of the preceding anomaly event and the historical fault time of the target anomaly event. Subtract the historical fault time of the preceding anomaly event from the historical fault time of the target anomaly event to obtain the time interval. Read the time decay parameter, multiply the time decay parameter by the time interval and take the negative value to obtain the decay exponent. Perform exponential calculation with the natural constant as the base and the decay exponent as the exponent to obtain the time interval decay value of the preceding anomaly event relative to the target anomaly event. The triggering contribution of preceding abnormal events to the target abnormal event is calculated based on the time interval decay value, fault type consistency coefficient, topology attribution consistency coefficient, node connectivity coefficient, and common cause suppression coefficient. The triggering contributions corresponding to the same target abnormal event are then summed to obtain the fault triggering intensity. Specifically, the calculation of the triggering contribution of preceding abnormal events to the target abnormal event based on the time interval decay value, fault type consistency coefficient, topology attribution consistency coefficient, node connectivity coefficient, and common cause suppression coefficient is as follows: Read the time interval decay value, fault type consistency coefficient, topology attribution consistency coefficient, node connection coefficient, and common cause suppression coefficient corresponding to the preceding abnormal event and the target abnormal event within the same event triggering channel; multiply the fault type consistency coefficient, topology attribution consistency coefficient, and node connection coefficient to obtain the event association strength; multiply the event association strength by the time interval decay value to obtain the time-decayed trigger value; read the common cause suppression coefficient. When the common cause suppression coefficient is 1, keep the time-decayed trigger value unchanged. When the common cause suppression coefficient is less than 1, multiply the time-decayed trigger value by the common cause suppression coefficient to obtain the triggering contribution of the preceding abnormal event to the target abnormal event; sum the triggering contributions of all preceding abnormal events corresponding to the same target abnormal event to obtain the fault triggering strength corresponding to the target abnormal event. The fault trigger intensity is normalized to obtain a dynamic weight value, and the dynamic weight value is written into the corresponding node connection relationship in the static running topology graph to generate dynamic topology weight.

[0025] In this embodiment, the steps of allocating node attention through the topology constraint graph attention layer, correcting abnormal residual contributions through the environmental disturbance correction layer, and generating a health score by fusing abnormal scores, fault types, and fault trends through the health decision layer include: An improved MTAD-GAT model is constructed, comprising a topological constraint graph attention layer, a temporal prediction reconstruction layer, an environmental perturbation correction layer, an anomaly scoring layer, and a health decision layer. Specifically, the construction of the improved MTAD-GAT model involves: Set up input interfaces for comprehensive state representation, environmental disturbance representation, static running topology graph and dynamic topology weight, and establish hierarchical connection relationships in the order of topology constraint graph attention layer, time series prediction reconstruction layer, environmental disturbance correction layer, anomaly scoring layer and health decision layer; Configure a node feature reading unit, an adjacent node filtering unit, a state similarity calculation unit, and a dynamic topology weight adjustment unit for the attention layer of the topology constraint graph, so that the comprehensive state representation, static running topology graph, and dynamic topology weight can complete the topology association encoding in the same layer; Configure the temporal prediction reconstruction layer with a time window sorting unit, a temporal dependency coding unit, a state prediction unit, and a state reconstruction unit so that the topology-enhanced state features form prediction residuals and reconstruction residuals. An environmental disturbance reading unit, a residual category classification unit, and a residual correction unit are configured for the environmental disturbance correction layer, so that the environmental disturbance characterization participates in the correction processing of the predicted residual and the reconstructed residual. Configure residual fusion unit, fault type matching unit and fault trend generation unit for the anomaly scoring layer so that environmental correction residual, prediction residual and reconstruction residual can form anomaly score, fault type and fault trend; Configure historical operation and maintenance data reading unit, normalization coding unit and weighted fusion unit for the health status decision layer, so that the anomaly score, fault type, fault trend, fault recurrence number, equipment commissioning years and maintenance interval form a health status score; The parameters of each layer are initialized, and a parameter update channel from health score to training target is established to form an improved MTAD-GAT model; The topology constraint graph attention layer reads the comprehensive state representation according to the aviation obstruction light equipment identification, determines the set of adjacent nodes according to the static operating topology graph, calculates the state similarity between nodes, and uses dynamic topology weights to weight and normalize the state similarity to obtain node attention coefficients. The comprehensive state representations of adjacent nodes are aggregated according to the node attention coefficients to generate topology-enhanced state features. The topology constraint graph attention layer includes: Node feature reading unit: Reads the comprehensive status representation corresponding to the aviation obstruction light equipment identifier, and maintains the time sequence of node features according to the monitoring time window; Adjacent node filtering unit: Reads the node connection relationship in the static running topology graph and determines the set of adjacent nodes corresponding to each aviation obstruction light node; State similarity calculation unit: Performs field alignment, difference statistics and change direction comparison on the comprehensive state representation of aviation obstruction light nodes and adjacent nodes to obtain the state similarity between nodes; Dynamic topology weight adjustment unit: Reads dynamic topology weights, performs weighted adjustment and normalization on the state similarity between nodes to obtain node attention coefficients, and aggregates the comprehensive state representation of adjacent nodes according to the node attention coefficients. In the topology constraint graph attention layer, the node feature reading unit reads the comprehensive state representation within the corresponding monitoring time window according to the aviation obstruction light equipment identifier and writes the comprehensive state representation into the corresponding aviation obstruction light node; the adjacent node filtering unit retains adjacent nodes with node connection relationships with the current aviation obstruction light node according to the node connection matrix in the static running topology graph, and filters out aviation obstruction light nodes without node connection relationships; the state similarity calculation unit aligns the fields of the comprehensive state representation of the current aviation obstruction light node and adjacent nodes, calculates the absolute value of the difference between corresponding fields, the consistency of change direction, and the closeness of state fluctuation, and synthesizes the calculation results into the state similarity between nodes; the dynamic topology weight adjustment unit reads the dynamic topology weight between the current aviation obstruction light node and adjacent nodes, writes the dynamic topology weight into the adjustment process of the state similarity between nodes, normalizes the adjusted state similarity to obtain the node attention coefficient, and performs weighted aggregation on the comprehensive state representation of adjacent nodes according to the node attention coefficient to generate topology-enhanced state features; The temporal prediction and reconstruction layer arranges the topology-enhanced state features according to the monitoring time window order and performs temporal dependency encoding to generate state prediction representations and state reconstruction representations. The differences between these representations and the comprehensive state representation are calculated item by item to obtain prediction residuals and reconstruction residuals. The temporal prediction and reconstruction layer includes: Time window sorting unit: Arrange the topology enhancement state features according to the time order of the monitoring time window to form a continuous topology enhancement state sequence; Temporal Dependency Encoding Unit: Performs temporal location writing, adjacent window change extraction, and long-term change trend encoding on continuous topology augmented state sequences to generate temporal dependency features; State prediction unit: Generates a state prediction representation of the target monitoring time window based on the temporal dependency features prior to the current monitoring time window; State Reconstruction Unit: Based on the temporal dependency characteristics within the current monitoring time window, the topology-enhanced state features are reconstructed to generate a state reconstruction representation; The residual calculation unit calculates the difference between the state prediction representation and the state reconstruction representation and the comprehensive state representation within the same monitoring time window to obtain the prediction residual and the reconstruction residual. In the temporal prediction and reconstruction layer, the time window sorting unit reads the topology enhancement state features according to the aviation obstruction light equipment identifier and arranges them into a continuous topology enhancement state sequence based on the start and end times of the monitoring time window; the temporal dependency encoding unit writes the time window position information into the continuous topology enhancement state sequence, calculates the feature change amplitude and direction between adjacent monitoring time windows, and encodes the change trend within multiple consecutive monitoring time windows to form temporal dependency features; the state prediction unit reads the temporal dependency features before the target monitoring time window and outputs the state prediction representation corresponding to the target monitoring time window; the state reconstruction unit reads the temporal dependency features within the target monitoring time window and outputs the state reconstruction representation corresponding to the target monitoring time window; the residual calculation unit calculates the difference between the state prediction representation and the comprehensive state representation item by item according to the field order in the comprehensive state representation to obtain the prediction residual, and calculates the difference between the state reconstruction representation and the comprehensive state representation item by item to obtain the reconstruction residual; The environmental disturbance correction layer reads the environmental disturbance characterization, extracts the influence terms of ambient light, temperature and humidity, rain and fog, and visibility, and classifies the prediction residuals and reconstruction residuals into electrical residuals, optical residuals, scintillation residuals, and communication residuals. It generates residual correction coefficients and corrects various residuals to generate environmental correction residuals. The environmental disturbance correction layer includes: Environmental disturbance reading unit: Reads environmental disturbance characteristics and extracts the influence items of ambient light, temperature and humidity, rain and fog, and visibility. Residual category classification unit: According to the source of the corresponding fields of the prediction residual and the reconstruction residual, the prediction residual and the reconstruction residual are classified into electrical residual, optical residual, scintillation residual and communication residual; Residual Correction Unit: Generates residual correction coefficients according to the correspondence between various residuals and environmental impact items, corrects various residuals, and generates environmentally corrected residuals; In the environmental disturbance correction layer, the environmental disturbance reading unit reads the environmental disturbance characterization according to the aviation obstruction light equipment identification and monitoring time window. It extracts fields corresponding to changes in ambient light, temperature and humidity, rain and fog conditions, and visibility from the environmental disturbance characterization, forming ambient light impact items, temperature and humidity impact items, rain and fog impact items, and visibility impact items, respectively. The residual classification unit reads the predicted residual and reconstructed residual according to the field sources of the comprehensive state characterization. Residuals from voltage, current, and power fields are classified as electrical residuals; residuals from light intensity and on / off state fields are classified as optical residuals; and residuals from flicker frequency and flicker period fields are classified as flicker residuals. In the future... The residuals derived from online status, signal strength, data packet loss rate, and communication delay fields are classified as communication residuals. The residual correction unit configures the ambient light, rain / fog, and visibility impact items into the optical residual correction process, the temperature and humidity impact items into the electrical and flicker residual correction processes, and the rain / fog and visibility impact items into the communication residual correction process. Environmental impact items participating in the same type of residual correction are normalized, calibrated, and weighted to obtain the corresponding residual correction coefficients. The electrical, optical, flicker, and communication residuals are then multiplied and corrected according to these residual correction coefficients. The corrected residuals are then encoded and combined in field order to generate the environmental correction residuals. The anomaly scoring layer performs weighted fusion of environmental correction residuals, prediction residuals, and reconstruction residuals to generate anomaly scores. It then generates fault trends based on the direction, magnitude, and duration of change in the anomaly scores within a continuous monitoring time window. Finally, it generates fault types based on the matching results between the anomaly scores and fault types in historical operation and maintenance data. The anomaly scoring layer includes: Residual fusion unit: Reads environmentally corrected residuals, prediction residuals, and reconstructed residuals, and performs field alignment and weight configuration on the three types of residuals; Anomaly score generation unit: Performs weighted fusion and normalization mapping on the field-aligned residuals to generate anomaly scores; Fault Trend Generation Unit: Reads abnormal scores according to the continuous monitoring time window, calculates the direction, magnitude and duration of change of abnormal scores, and generates fault trends; Fault type matching unit: Reads fault types from historical operation and maintenance data and generates fault types according to the source of residual fields and the distribution of abnormal scores; In the anomaly scoring layer, the residual fusion unit aligns the environmental correction residual, prediction residual, and reconstruction residual according to the field order of the comprehensive state representation, and writes environmental correction residuals, prediction residuals, and reconstruction residuals from the same field source into the same residual calculation group; it configures residual weights according to the field source of electrical residuals, optical residuals, scintillation residuals, and communication residuals, and performs weighted summation on each residual calculation group to obtain field anomaly values; the anomaly scoring generation unit accumulates the anomaly values ​​of each field and maps the accumulation result to a unified scoring interval to generate an anomaly score; the fault trend generation unit reads continuous monitoring data according to the aviation obstruction light equipment identification. The anomaly score within the time window is calculated by comparing the anomaly score of the current monitoring time window with that of the previous monitoring time window. A positive difference is recorded as an upward trend, a negative difference as a downward trend, and a zero difference as a stable trend. The number of monitoring time windows in which the anomaly score consistently changes in the same direction is also counted to generate a fault trend. The fault type matching unit reads the fault types from historical operation and maintenance data, counts the sources of residual fields and the distribution of anomaly scores corresponding to different fault types, and matches the source of the residual field with the highest proportion in the current monitoring time window with the correspondence of historical fault types to generate the fault type. The health decision-making layer reads anomaly scores, fault types, fault trends, fault recurrence counts, equipment service life, and maintenance intervals. It then generates a health score through normalized coding and weighted fusion. The health decision-making layer includes: Historical maintenance data reading unit: Reads the number of fault recurrences, equipment service life, and maintenance intervals according to the aviation obstruction light equipment identification; Normalized coding unit: unifies the numerical range and fields of anomaly scores, fault types, fault trends, fault recurrence times, equipment service life, and maintenance intervals; Weighted fusion unit: Configures health weights for the normalized encoded fields and performs weighted fusion; Health score generation unit: maps the weighted fusion result into a health score; In the health decision-making layer, the historical operation and maintenance data reading unit reads the anomaly score, fault type, and fault trend according to the aviation obstruction light equipment identifier, and reads the fault recurrence count, equipment service life, and maintenance interval corresponding to the same aviation obstruction light equipment identifier from the historical operation and maintenance data; the normalization coding unit maps the anomaly score to a unified numerical range, converts the fault type into a fault type code according to the fault type field in the historical operation and maintenance data, converts the fault trend into a trend code according to the upward, downward, and stable directions, and normalizes the fault recurrence count, equipment service life, and maintenance interval according to the value range of the corresponding fields; the weighted fusion unit configures the health weights according to the field order of anomaly score, fault type code, trend code, fault recurrence count, equipment service life, and maintenance interval, and performs a weighted sum of the normalization results of each field to obtain the health fusion value; the health score generation unit maps the health fusion value to the range of 0 to 100 to generate the health score; The training samples are constructed by selecting the monitoring time window corresponding to the historical failure time and previous monitoring time windows. Failure type labels are generated from the failure types in the historical operation and maintenance data. Health score supervision items are generated based on maintenance time, failure recurrence frequency, and maintenance interval. The weighted sum of state prediction bias, state reconstruction bias, failure type classification bias, and health score bias is calculated to obtain the training objective. Backpropagation is then used to update and improve the MTAD-GAT model parameters. Specifically, the backpropagation method for updating and improving the MTAD-GAT model parameters is as follows: The training samples are input into the improved MTAD-GAT model, and then pass through the topological constraint graph attention layer, the temporal prediction reconstruction layer, the environmental disturbance correction layer, the anomaly scoring layer, and the health decision layer in sequence to obtain the state prediction representation, the state reconstruction representation, the fault type output result, and the health score output result. The difference between the state prediction representation and the comprehensive state representation within the corresponding monitoring time window is calculated to obtain the state prediction deviation. The difference between the state reconstruction representation and the comprehensive state representation within the corresponding monitoring time window is calculated to obtain the state reconstruction deviation. The difference between the fault type output and the fault type label is calculated to obtain the fault type classification deviation. The difference between the health score output and the health score supervision item is calculated to obtain the health score deviation. The training objective is obtained by weighting and summing the results of the weight configuration of state prediction bias, state reconstruction bias, fault type classification bias and health score bias. Based on the training objective, the gradients of the trainable parameters in the topological constraint graph attention layer, temporal prediction reconstruction layer, environmental perturbation correction layer, anomaly scoring layer, and health decision layer are calculated respectively. The node attention parameters, temporal coding parameters, residual correction parameters, anomaly scoring parameters, and health decision parameters are updated according to the gradient direction until the training objective reaches the convergence condition, and the parameters of the improved MTAD-GAT model after training are obtained.

[0026] In this embodiment, generating alarm levels, maintenance priorities, and operation and maintenance suggestions, and writing the maintenance results into historical operation and maintenance data includes: The health score, fault type, and fault trend output by the improved MTAD-GAT model are read. The health score is then normalized across intervals, and corresponding alarm levels are generated based on the alarm level range to which the health score belongs. These alarm levels include Emergency, Severe, Warning, Attention, and Normal. The alarm levels are adjusted according to the fault trend, maintaining the same level or shifting towards Emergency. Specifically, the adjustment is as follows: Read the fault trend output by the improved MTAD-GAT model, extract the direction, magnitude and duration of change in the fault trend, and establish an alarm level sequence according to the level order of normal, attention, warning, severe and emergency. When the direction of change is downward or stable, the alarm level of the health score remains unchanged; when the direction of change is upward, the magnitude and duration of the change are read, the magnitude of the change is compared with the magnitude boundary obtained from historical operation and maintenance data, and the duration is compared with the duration boundary obtained from historical operation and maintenance data. When the change reaches the magnitude boundary and the duration reaches the duration boundary, the alarm level of the health score will be raised one level towards the emergency direction along the alarm level sequence. When both the magnitude and duration of the change reach the higher level boundary, the alarm level of the health score will be raised two levels towards the emergency level along the alarm level sequence. When the adjusted alarm level exceeds the emergency level, the emergency level will be recorded as the final alarm level. The system retrieves fault recurrence counts, equipment service life, and maintenance intervals from historical maintenance data. It also retrieves installation height, associated building, associated power supply line, associated communication gateway, and adjacent aviation obstruction light equipment identifiers from operational topology data. A maintenance priority is generated by summing the scores corresponding to alarm level, fault recurrence count, equipment service life, maintenance interval, and operational topology data. Specifically, the maintenance priority generation process is as follows: Read the alarm level and configure the corresponding score for the alarm level in the order of emergency, critical, warning, attention and normal; Read the number of fault recurrences, the equipment's years of operation, and the maintenance interval, and normalize and encode them according to the value range in the historical operation and maintenance data to obtain the scores corresponding to the number of fault recurrences, the equipment's years of operation, and the maintenance interval. Read the installation height, the building to which it belongs, the power supply line to which it belongs, the communication gateway to which it belongs, and the identification of adjacent aviation obstruction light equipment. Normalize the installation height and encode it. Count the number of alarm events within the same building, power supply line, communication gateway, and adjacent aviation obstruction light equipment identification range. Normalize the statistical results and encode them to obtain the corresponding score of the running topology data. The maintenance priority score is obtained by summing the scores corresponding to the alarm level, the number of fault recurrences, the equipment's service life, the maintenance interval, and the operation topology data. The aviation obstruction light equipment icons are arranged from highest to lowest according to their maintenance priority scores to form a maintenance priority. Maintenance recommendations are generated based on fault type, alarm level, and maintenance priority. These recommendations include suggestions for lighting fixture inspection, power supply inspection, communication inspection, environmental verification, and component replacement. The system receives maintenance time, maintenance measures, replaced components, and fault recurrence confirmation results and writes them into historical maintenance data.

[0027] Example 1: To verify the feasibility of this invention in practice, it was applied to a remote monitoring scenario for aviation obstruction lights in an airport industrial park. The park includes two high-rise buildings, two communication towers, one chimney, and two wind turbine anemometer towers, with a total of 96 aviation obstruction lights installed, involving 14 controllers, 16 power supply lines, and 9 communication gateways. The system ran continuously for 30 days, with on-site data acquisition terminals collecting data at 30-second intervals, resulting in 8,294,400 raw monitoring records. Multi-source monitoring data included voltage, current, power, light intensity, on / off status, flicker frequency, communication status, ambient brightness, temperature, humidity, rain / fog conditions, and visibility. Operational topology data included equipment identification, installation height, affiliated building, affiliated controller, affiliated power supply line, affiliated communication gateway, and identification of adjacent aviation obstruction light equipment. Historical maintenance data included 183 fault records from the past 12 months, including 52 records of light intensity attenuation, 47 records of communication downtime, 31 records of power supply anomalies, 23 records of flicker anomalies, and 30 records of light fixture aging. The raw data showed voltage ranging from 214.6 to 229.8 V, current from 0.18 to 1.42 A, luminous intensity from 28.5 to 2054.3 cd, communication delay from 38 to 860 ms, and visibility from 0.18 to 11.6 km. Four rain and fog events occurred during the monitoring period. The system mapped 30-second sampling data to 5-minute monitoring time windows, creating 288 windows per light per day, totaling 276,480 monitoring windows over 30 days. 21,864 records (0.26%) were missing from the raw data. The system supplemented the missing monitoring items by using valid sampling values ​​from adjacent windows of the same aviation obstruction light, valid sampling values ​​from the same power supply line during the same period, and valid sampling values ​​from the same communication gateway during the same period. After supplementation, the percentage of valid windows increased from 96.8% to 99.4%. Meanwhile, the system removed 1,732 abnormal monitoring items that exceeded the operating range of the equipment and the measurement range of the acquisition terminal, such as abnormal samples with voltage below 180V, current of 0 but light intensity greater than 100cd, and communication delay exceeding 2,000ms. After completing dimensional unification and normalized encoding, multi-source time-series data of aviation obstruction lights with a size of 276,480×32 was obtained.

[0028] During feature processing, the system extracts lamp operation features, communication status features, and environmental impact features from multi-source time-series data of aviation obstruction lights. For electrical operation data, optical status data, and flicker control data, the system extracts the mean, standard deviation, rate of change, and periodic consistency within the window to form 18-dimensional lamp operation features. The mean rate of change of light intensity for normal lamps is 0.031, while the mean rate of change of light intensity for lamps with diminishing intensity is 0.184. For communication status data, the system extracts online duration, signal strength decrease, packet loss rate changes, and communication delay fluctuations to form 8-dimensional communication status features. During periods of communication gateway anomaly, the variance of communication delay for lamps under the same gateway increases from 64.2 to 531.7. For environmental impact data, the system extracts changes in ambient light, temperature and humidity, rain and fog conditions, and visibility changes to form 6-dimensional environmental impact features. The system fuses and encodes the lamp operation features and communication status features to generate a 64-dimensional comprehensive status representation, and independently encodes the environmental impact features to generate a 16-dimensional environmental disturbance representation. The system then established a node table based on the identifiers of 96 aviation obstruction lights and a topology edge table based on the associated building, controller, power supply line, communication gateway, and adjacent equipment relationships, resulting in 426 static topology edges. The system retrieved historical fault times, fault types, aviation obstruction light identifiers, and fault recurrence counts from historical maintenance data and combined this with environmental disturbance characteristics to form an abnormal event sequence. During the heavy fog in the early morning of day 12, 34 lights experienced a decrease in light intensity within the same monitoring time window. Of these, 31 lights showed a decrease in visibility from 8.1 km to 0.42 km, with the light intensity decrease concentrated between 18% and 26%. Voltage, current, and flicker frequency remained stable, leading the system to identify this as a synchronous anomaly. The Hawkes process, which suppresses common causes, generates a common cause suppression coefficient of 0.73 when calculating trigger contributions, reducing the average trigger contribution among the batch of abnormal events from 0.61 to 0.16. When three lamps on the same power supply line experience consecutive voltage fluctuations, current drops, and abnormal light intensity, their trigger contributions remain between 0.68 and 0.74 due to the low similarity of environmental disturbances and consistent topological affiliation. After normalization, the average dynamic topological weight of normal adjacent lamps is 0.21, while the dynamic topological weight of abnormal lamps on the same faulty power supply line rises to 0.82. The dynamic topological weight of synchronous light intensity anomalies under the same rainy / foggy environment is suppressed to 0.19.

[0029] The system integrates state representation, environmental disturbance representation, static operating topology graph, and dynamic topology weights into the improved MTAD-GAT model. The topology constraint graph attention layer determines the set of adjacent nodes according to the static operating topology graph and uses dynamic topology weights to adjust the state similarity between nodes. In power supply line anomaly samples, the node attention coefficient between abnormal lamps increases from 0.34 to 0.78, while the attention coefficient between lamps on different lines decreases to 0.12. The temporal prediction and reconstruction layer performs temporal dependency encoding according to the 5-minute monitoring window sequence, generating state prediction representations and state reconstruction representations. The mean prediction residual for normal samples is 0.047, and the mean prediction residual for real fault samples is 0.286. The environmental disturbance correction layer reads the environmental disturbance representation and corrects the electrical, optical, flicker, and communication residuals. In rain and fog samples, the optical residual is corrected from 0.241 to 0.083, and in lamp aging samples, the optical residual is corrected from 0.263 to 0.247. The anomaly scoring layer generates anomaly scores in the 0–1 range, with a mean score of 0.13 for normal states, 0.31 for common environmental anomalies, and 0.76 for actual equipment failures. The health decision layer integrates anomaly scores, failure types, failure trends, failure recurrence counts, equipment service life, and maintenance intervals. The mean score for healthy lighting fixtures is 87.6, for fixtures requiring attention is 68.4, for warning fixtures is 49.7, and for severely faulty fixtures is 28.3. During the training phase, the monitoring window corresponding to the historical failure time and the 24 monitoring windows preceding it were selected as training samples, resulting in 4392 positive samples and 18000 normal samples. After 50 rounds of training, the state prediction bias decreased from 0.214 to 0.061, the state reconstruction bias decreased from 0.187 to 0.054, the failure type classification accuracy increased from 72.6% to 93.8%, and the mean absolute error of the health score decreased from 14.7 to 4.9. Compared with traditional single-threshold remote alarm methods, the fault identification accuracy of the traditional method is 79.3%, while that of this invention is 94.8%; the false alarm rate of the traditional method is 18.6%, while that of this invention is 5.7%; the false negative rate of the traditional method is 11.4%, while that of this invention is 3.2%; the average fault detection time of the traditional method is 26.5 hours, while that of this invention is 2.8 hours; the misjudgment rate of common environmental anomalies of the traditional method is 42.1%, while that of this invention is 8.9%; the consistency rate between maintenance priority and manual review results of the traditional method is 68.5%, while that of this invention is 91.2%. These results demonstrate that this invention can solve the problems of untimely remote status perception and low fault identification accuracy caused by reliance on manual inspections or single-parameter alarms in aviation obstruction light monitoring, as well as the problems of difficulty in predicting fault trends and lack of data support for operation and maintenance decisions due to the lack of comprehensive analysis of operational data, environmental data, and historical maintenance data.

[0030] 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 remote intelligent monitoring method for the status of aviation obstruction lights based on big data analysis, characterized in that, Includes the following steps: Collect multi-source monitoring data, operational topology data, and historical maintenance data of aviation obstruction lights; The multi-source monitoring data, operational topology data, and historical operation and maintenance data are standardized and preprocessed to generate multi-source time-series data of aviation obstruction lights; The operating characteristics, communication status characteristics, and environmental impact characteristics of aviation obstruction lights are extracted from multi-source time-series data. The operating characteristics and communication status characteristics of the lights are fused to generate a comprehensive status characterization, and the environmental impact characteristics are encoded to generate an environmental disturbance characterization. A static operational topology map is constructed based on the operational topology data. Anomaly occurrence time, device identifier, fault type marker, and topology attribution marker are extracted from historical operation and maintenance data and environmental disturbance characterization to form an anomaly event sequence. The contribution of synchronous anomaly triggering from the same source is weakened through the common cause suppression Hawkes process. The fault triggering intensity is calculated and converted into dynamic topology weights. The comprehensive state representation, environmental disturbance representation, static operating topology graph and dynamic topology weight are input into the improved MTAD-GAT model. The attention of nodes is allocated through the topology constraint graph attention layer, the contribution of abnormal residuals is corrected through the environmental disturbance correction layer, and the health score is generated by fusing abnormal scores, fault types and fault trends through the health decision layer. Based on health scores, historical operation and maintenance data, and operational topology data, alarm levels, repair priorities, and operation and maintenance recommendations are generated, and the repair results are written into historical operation and maintenance data. The process of weakening the contribution of homogeneous synchronization anomaly triggering through the common-cause suppression Hawkes process, calculating the fault triggering strength and converting it into dynamic topology weights includes: Based on the operational topology data, establish an aviation obstruction light node table and a topology edge table, configure a node number for each aviation obstruction light device identifier, and write nodes that match each other in the topology edge table, such as those belonging to the same building, the same controller, the same power supply line, the same communication gateway, or adjacent aviation obstruction light device identifiers, to generate a static operational topology diagram. Read historical fault time, fault type, aviation obstruction light equipment identification and fault recurrence number, match the building, controller, power supply line and communication gateway to which the aviation obstruction light equipment belongs according to the aviation obstruction light equipment identification, match environmental disturbance characteristics according to historical fault time, configure fault type mark, topology attribution mark and environmental disturbance mark for each historical fault record, and form an abnormal event sequence in ascending order of historical fault time; Event triggering channels are divided according to fault type marking and topology attribution marking. Within each event triggering channel, homogeneous synchronous anomalies are identified, and the anomaly event number, target anomaly event number, and homogeneous synchronous anomaly identification results are written into the homogeneous synchronous anomaly marking table. The homogeneous synchronous anomaly is an anomaly event in which the topology attribution marking is consistent within the same monitoring time window or adjacent monitoring time windows and the environmental disturbance marking reaches the similarity condition formed by the statistical representation of historical environmental disturbances. Read the same-source synchronization anomaly labeling table. During the Hawkes common cause suppression process, configure the basic trigger parameters, time decay parameters, fault type consistency coefficient, topology affiliation consistency coefficient, node connection coefficient, and common cause suppression coefficient. For the anomaly event numbers identified as same-source synchronization anomalies in the same-source synchronization anomaly labeling table, configure a common cause suppression coefficient lower than the baseline value for the target anomaly event number. For the anomaly event numbers identified as non-same-source synchronization anomalies, configure a common cause suppression coefficient of the baseline value for the target anomaly event number. Where: Read the sequence of abnormal events according to the event triggering channel, count the number of abnormal events in the same event triggering channel, read the earliest and latest historical fault times, calculate the time span between the latest and earliest historical fault times, divide the number of abnormal events by the time span, and obtain the basic triggering parameters. Arrange the abnormal events in the same event triggering channel in ascending order of historical fault time, calculate the time interval between adjacent abnormal events one by one, extract the median of the time interval, and divide 1 by the median to obtain the time decay parameter. Read the fault type flags of the preceding abnormal event and the target abnormal event. When the two fault type flags are the same, the fault type consistency coefficient is 1; when the two fault type flags are different, the fault type consistency coefficient is 0. Read the topology attribution tags of the preceding abnormal event and the target abnormal event, compare the building, controller, power supply line and communication gateway to which they belong, count the number of consistent items, divide the number of consistent items by the total number of topology attribution tags, and obtain the topology attribution consistency coefficient. Read the node connection matrix in the static running topology graph. When there is a node connection relationship between the node corresponding to the preceding abnormal event and the node corresponding to the target abnormal event, the node connection coefficient is 1. When there is no node connection relationship between the node corresponding to the preceding abnormal event and the node corresponding to the target abnormal event, the node connection coefficient is 0. For the target abnormal event, select the preceding abnormal event whose historical failure time is earlier than the target abnormal event and is located in the same event triggering channel, and obtain the time interval decay value by multiplying the negative time decay parameter by the time interval with the natural constant as the base. The triggering contribution of preceding abnormal events to the target abnormal event is calculated based on the time interval decay value, fault type consistency coefficient, topology affiliation consistency coefficient, node connection coefficient, and common cause suppression coefficient. The triggering contributions corresponding to the same target abnormal event are summed to obtain the fault triggering intensity. The fault trigger intensity is normalized to obtain a dynamic weight value, and the dynamic weight value is written into the corresponding node connection relationship in the static running topology graph to generate dynamic topology weight.

2. The method for remote intelligent monitoring of the status of aviation obstruction lights based on big data analysis according to claim 1, characterized in that, The multi-source monitoring data includes electrical operation data, optical status data, flicker control data, communication status data, and environmental impact data. The electrical operation data includes voltage, current, and power. The optical status data includes light intensity and on / off status. The flicker control data includes flicker frequency and flicker period. The communication status data includes online status, signal strength, data packet loss rate, and communication delay. The environmental impact data includes ambient brightness, temperature, humidity, rain / fog conditions, and visibility.

3. The method for remote intelligent monitoring of the status of aviation obstruction lights based on big data analysis according to claim 1, characterized in that, The operational topology data includes the identification of the aviation obstruction light equipment, its installation location, installation height, the building to which it belongs, the controller to which it belongs, the power supply line to which it belongs, the communication gateway to which it belongs, and the identification of adjacent aviation obstruction light equipment.

4. The method for remote intelligent monitoring of the status of aviation obstruction lights based on big data analysis according to claim 1, characterized in that, The historical operation and maintenance data includes historical fault time, fault type, repair time, repair measures, replaced parts, number of fault recurrences, equipment commissioning years, and maintenance intervals.

5. The method for remote intelligent monitoring of the status of aviation obstruction lights based on big data analysis according to claim 1, characterized in that, The generated multi-source time-series data for aviation obstruction lights includes: Multi-source monitoring data are time-mapped according to equipment identification and collection timestamp, mapping electrical operation data, optical status data, flashing control data, communication status data and environmental impact data to the same monitoring time window, and linking operation topology data and historical operation and maintenance data to the corresponding aviation obstruction lights; The time-mapped multi-source monitoring data is processed for integrity and outlier handling. The missing monitoring items are supplemented by the valid sampled values ​​within adjacent monitoring time windows of the same aviation obstruction light, the valid sampled values ​​of the same power supply line and the valid sampled values ​​of the same communication gateway. Abnormal monitoring items that exceed the working range of the equipment and the range of the acquisition terminal are removed. The multi-source monitoring data, which has undergone integrity processing and outlier processing, the associated operational topology data, and the associated historical operation and maintenance data are quantified and normalized. Sliding window samples are generated according to the aviation obstruction light equipment identification, monitoring time window, operational topology data, and historical operation and maintenance data to obtain multi-source time-series data of aviation obstruction lights.

6. The method for remote intelligent monitoring of the status of aviation obstruction lights based on big data analysis according to claim 1, characterized in that, The integrated lighting fixture operation characteristics and communication status characteristics generate a comprehensive state characterization, and the encoded environmental impact characteristics generate an environmental disturbance characterization, including: Based on the identification of aviation obstruction lights and the monitoring time window, extract electrical operation data, optical status data, flashing control data, communication status data and environmental impact data from the multi-source time-series data of aviation obstruction lights; Statistical features, rate of change features, and periodic consistency features are extracted from electrical operation data, optical status data, and flicker control data within a window to generate luminaire operation features; Extract online continuous features, signal strength change features, packet loss change features, and communication delay change features from communication status data to generate communication status features; Environmental impact features are generated by extracting ambient light variation characteristics, temperature and humidity variation characteristics, rain and fog characteristics, and visibility variation characteristics from environmental impact data. The operating characteristics and communication status characteristics of the lighting fixtures are fused and encoded to generate a comprehensive status representation, and the environmental impact characteristics are independently encoded to generate an environmental disturbance representation.

7. The method for remote intelligent monitoring of the status of aviation obstruction lights based on big data analysis according to claim 1, characterized in that, The process of allocating node attention through a topology constraint graph attention layer, correcting abnormal residual contributions through an environmental disturbance correction layer, and generating a health score by fusing abnormal scores, fault types, and fault trends through a health decision layer includes: An improved MTAD-GAT model is constructed, which includes a topological constraint graph attention layer, a temporal prediction reconstruction layer, an environmental disturbance correction layer, an anomaly scoring layer, and a health decision layer. The topology constraint graph attention layer reads the comprehensive state representation according to the aviation obstruction light equipment identification, determines the set of adjacent nodes according to the static operating topology graph, calculates the state similarity between nodes, uses dynamic topology weights to weight and normalize the state similarity to obtain the node attention coefficient, and aggregates the comprehensive state representation of adjacent nodes according to the node attention coefficient to generate topology-enhanced state features. The temporal prediction and reconstruction layer arranges the topology-enhanced state features according to the monitoring time window order and performs temporal dependency encoding to generate state prediction representation and state reconstruction representation. The difference between each of the two representations is calculated with the comprehensive state representation to obtain the prediction residual and reconstruction residual. The environmental disturbance correction layer reads the environmental disturbance characterization, extracts the environmental light influence item, temperature and humidity influence item, rain and fog influence item and visibility influence item, divides the prediction residual and reconstruction residual into electrical residual, optical residual, scintillation residual and communication residual, generates residual correction coefficients and corrects various residuals, and generates environmental correction residuals; The anomaly scoring layer performs weighted fusion of environmental correction residuals, prediction residuals, and reconstruction residuals to generate anomaly scores. It generates fault trends based on the direction, magnitude, and duration of change of anomaly scores within a continuous monitoring time window, and generates fault types based on the matching results of anomaly scores and fault types in historical operation and maintenance data. The health assessment decision layer reads abnormal scores, fault types, fault trends, fault recurrence times, equipment service life, and maintenance intervals, and generates a health assessment score through normalized coding and weighted fusion. The training samples are formed by selecting the monitoring time window corresponding to the historical failure time and the previous monitoring time window. Failure type labels are generated by reading the failure type in the historical operation and maintenance data. Health score supervision items are generated according to maintenance time, failure recurrence number and maintenance interval. The weighted sum of state prediction bias, state reconstruction bias, failure type classification bias and health score bias is calculated to obtain the training target. Backpropagation is used to update and improve the MTAD-GAT model parameters.

8. The method for remote intelligent monitoring of the status of aviation obstruction lights based on big data analysis according to claim 1, characterized in that, The process of generating alarm levels, repair priorities, and maintenance recommendations, and writing the repair results into historical maintenance data, includes: Read the health score, fault type and fault trend output by the improved MTAD-GAT model, perform interval normalization on the health score, and generate corresponding alarm levels according to the alarm level to which the health score belongs. The alarm levels include emergency, severe, warning, attention and normal. The alarm levels are maintained at the same level or adjusted towards emergency according to the fault trend. The system reads the number of fault recurrences, equipment service life, and maintenance intervals from historical operation and maintenance data. It also reads the installation height, building, power supply line, communication gateway, and adjacent aviation obstruction light equipment identifiers from the operation topology data. The system then sums the scores corresponding to the alarm level, the number of fault recurrences, the equipment service life, the maintenance interval, and the operation topology data to generate maintenance priorities. Maintenance recommendations are generated based on fault type, alarm level, and maintenance priority. These recommendations include suggestions for lighting fixture inspection, power supply inspection, communication inspection, environmental verification, and component replacement. The system receives maintenance time, maintenance measures, replaced components, and fault recurrence confirmation results and writes them into historical maintenance data.

Images