Project operation intelligent evaluation and prediction method based on global device health situation

Abstract

The application belongs to the technical field of intelligent city lighting operation and maintenance management, and relates to a project operation and maintenance intelligent evaluation and prediction method based on global device health situation, which is applied to a lamp networking intelligent control system. The method acquires electrical, optical, environmental, communication and operation parameters, and constructs a power supply-communication-space three-edge heterogeneous graph. After multi-time scale feature extraction and cross-scale modulation, a multi-head graph attention network, a GRU and a Bayesian neural network-dynamic linear model cascade architecture are used to realize full-link uncertainty propagation and residual life probability estimation. Based on the differentiated Copula tail correlation correction health situation index of fault type perception, an operation and maintenance strategy is generated, and model parameter and cross-device knowledge migration are optimized in a closed loop. The application solves the problems of physical constraint loss, uncertainty fracture and group fault early warning deficiency in the prior art, and significantly improves the evaluation accuracy and operation and maintenance efficiency.

Description

Technical Field

[0001] This invention belongs to the field of smart city lighting operation and maintenance management technology, and relates to a project operation and maintenance intelligent assessment and prediction method based on the health status of all devices. Background Technology

[0002] With the deepening of IoT technology and smart city construction, urban lighting systems have evolved from traditional independent lamp control to "Internet of Lights" intelligent control systems. An Internet of Lights system typically consists of tens of thousands of LED street light nodes, individual lamp controllers integrated into each lamp, circuit-level centralized controllers, edge gateways, power distribution circuits, and a backend cloud platform, forming a complex heterogeneous network with deep coupling of power supply topology, communication topology, and spatial topology.

[0003] In actual operation, lighting network equipment faces the combined effects of multiple stresses: voltage fluctuations, harmonic distortion, and three-phase imbalance within the power supply circuit affect the electrical lifespan of the equipment; environmental stresses such as high temperature, high humidity, dust, and vibration accelerate the degradation of LED light sources and driver power supplies; changes in dimming depth and load rate alter the levels of internal thermal and electrical stresses; and deterioration in communication quality leads to blind spots and data loss in equipment condition monitoring. Under the combined effect of these multiple stresses, the health status of the equipment exhibits complex characteristics such as spatial clustering, temporal lag, and fault propagation.

[0004] In existing technologies, the health assessment and fault prediction of lighting network devices have the following main shortcomings:

[0005] First, the graph construction methods lack physical topological constraints. Some existing solutions construct dynamic topology graphs based on electrical waveform correlation and channel feature similarity, which is essentially a purely data-driven adjacency matrix learning. However, data correlation in a network of lights is not always consistent with physical power supply and communication logic. For example, two lights far apart in the same circuit may exhibit low waveform correlation due to differences in load conditions, but when an electrical fault occurs, the fault will propagate along the power supply line, indicating a strong physical coupling between the two. Simply relying on data correlation to construct the graph structure will introduce systematic biases into fault propagation modeling.

[0006] Second, spatiotemporal feature extraction lacks multi-timescale fusion. The light decay of LED light sources is a slow degradation process spanning several months, the aging of capacitors in the driver power supply spans thousands of hours, while voltage drops and harmonic shocks in the power supply circuit are transient events on the order of seconds. Existing methods typically use a single time window for feature extraction, which cannot simultaneously capture slow degradation trends and transient anomalous shocks. Furthermore, when slow degradation has entered an accelerated phase, the meaning of transient anomalous features changes fundamentally—the same current spike indicates completely different failure risks in healthy equipment and equipment on the verge of failure.

[0007] Third, the uncertainty propagation link is broken. The latent states output by deep temporal networks are often directly used as observations in degradation models or directly mapped to health index point estimates. Uncertainties in each stage of feature extraction, temporal modeling, and state estimation are not explicitly modeled and propagated, resulting in a lack of credibility measurement for the final output health index and lifetime prediction. Operations personnel cannot distinguish between "high-confidence low-health warnings" and "low-confidence low-health warnings," and it is difficult to formulate differentiated operation and maintenance strategies accordingly.

[0008] Fourth, the tail-dependent nature of fault propagation among multiple devices is not effectively utilized. Devices within the same power supply circuit share electrical stress, and devices within the same communication subnet share communication quality degradation. When a device in a circuit experiences a voltage drop or abnormal leakage current, the electrical stress borne by other devices in the same circuit increases accordingly, exhibiting a lower tail-dependent characteristic in fault risk. When the performance of the gateway within the communication subnet deteriorates, the communication quality of all nodes in the subnet deteriorates synchronously, exhibiting an upper tail-dependent characteristic. Existing solutions typically evaluate each device as an independent entity, ignoring this tail-dependent structure of fault propagation, resulting in insufficient early warning capabilities for group faults. Summary of the Invention

[0009] The purpose of this invention is to provide a method for intelligent assessment and prediction of project operation and maintenance based on the health status of all devices, thereby solving the above-mentioned technical problems.

[0010] The technical solution adopted in this invention is as follows:

[0011] 1. A project operation and maintenance intelligent assessment and prediction method based on the health status of all devices, comprising: acquiring multi-source monitoring data of all-domain lighting network devices, and constructing a device heterogeneity graph based on the power supply topology, communication topology, and spatial location of the lighting network devices, wherein the device heterogeneity graph includes nodes and edges; extracting multi-time-scale time-frequency features from the multi-source monitoring data of each node to obtain a health feature vector for each node, and then inputting the health feature vectors of each node into a graph attention network for spatial aggregation to obtain the spatial coupling features of each node; inputting the spatial coupling features of each node into a gated recurrent unit network in a time sequence to obtain a spatiotemporal degradation hidden state sequence for each node, and then mapping the spatiotemporal degradation hidden state sequence to an observation distribution through a Bayesian neural network layer, wherein the observation... The observed distribution includes mean and variance. The observed distribution is input into a Bayesian dynamic linear model, using both mean and variance as the observation uncertainty. The health status index and remaining lifetime probability distribution of each node are estimated online. Based on a fault type-aware differentiated Copula (connection function) selection strategy, tail correlation analysis is performed on the health status index of nodes within the same power supply circuit or communication subnet. According to the fault type, a corresponding Copula is selected from a preset Copula candidate set to calculate the tail correlation coefficient. The health status index is then corrected based on the tail correlation coefficient to obtain a comprehensive device health status assessment result. An operation and maintenance strategy is generated based on the comprehensive device health status assessment result and the remaining lifetime probability distribution.

[0012] Optionally, the multi-source monitoring data includes electrical parameters, optical parameters, environmental parameters, communication parameters, and operating parameters;

[0013] The heterogeneous graph of the devices includes nodes and edges. The nodes include lighting fixture nodes, single-lamp controller nodes, centralized controller nodes, gateway nodes, and power distribution loop nodes. The weight of the power supply edge is determined by the loop impedance, power supply distance, and load rate. The weight of the communication edge is dynamically updated by the received signal strength indicator, signal-to-noise ratio, and packet loss rate. The weight of the spatial edge is determined by the geographical distance at exponential decay. The edge weights of the heterogeneous graph are iteratively updated based on the feedback from the subsequent graph attention network aggregation output, forming a dynamic evolution closed loop between graph construction and aggregation.

[0014] The multiple time scales include short-term windows, medium-term windows, and long-term windows; wherein, the trend features extracted by the long-term window are used as degradation priors to modulate the attention weights of the short-term window features;

[0015] The graph attention network adopts a multi-head attention mechanism, in which multiple attention heads are divided according to edge type and aggregate messages from neighboring nodes of the corresponding type. During the aggregation process, the attention temperature coefficient is dynamically adjusted according to the observation variance output by the Bayesian neural network layer. When the observation variance of a node exceeds a preset threshold, the weight of that node as a message transmission source is reduced.

[0016] The state equation of the Bayesian dynamic linear model includes an environmental stress-accelerated degradation term, and its process noise variance is dynamically determined by the Copula tail correlation coefficient of the loop in which the node is located.

[0017] Optionally, the construction of the device heterogeneity graph also includes the following physical constraints:

[0018] The weight of the power supply side decreases with increasing loop impedance and power supply distance, and increases with increasing load rate;

[0019] The weight of the communication side increases with the increase of the received signal strength indication and the signal-to-noise ratio, and decreases with the increase of the packet loss rate;

[0020] Spatial edges are established only when the geographical distance is less than a set threshold, and their weight decreases as the geographical distance increases.

[0021] When the weight of any power supply edge is lower than the preset power supply threshold, the spatial edge weight between the downstream load node of that power supply edge and its spatially adjacent node is forcibly reset to zero.

[0022] When the weight of any communication edge is lower than the preset communication threshold, the node does not participate in neighbor message passing in the graph attention network aggregation, but it is injected into the Bayesian neural network layer as a prior uncertainty of the observation blind zone.

[0023] Optionally, in the multi-timescale time-frequency feature extraction:

[0024] The short-time window is used to capture abrupt changes and transient anomalies, and the extracted features include kurtosis and peak factor.

[0025] The mid-time window is used to capture daily fluctuation patterns, and the extracted features include mean, variance, and sample entropy.

[0026] The long-term window is used to capture trend degradation, and the extracted features include the slope of the trend term and the energy decay rate;

[0027] Features from three time scales are weighted and fused using a time-scale attention mechanism;

[0028] When the slope of the trend term in the long-term window indicates that the device is in an accelerated degradation phase, the attention weight of the short-term window features is increased; when the slope of the trend term in the long-term window indicates that the device is in a healthy and stable phase, the attention weight of the short-term window features is suppressed.

[0029] Optionally, in the multi-head attention mechanism of the graph attention network:

[0030] The attention heads are divided into three groups: power supply side aggregation heads, communication side aggregation heads, and spatial side aggregation heads, which aggregate messages from neighboring nodes of the corresponding types.

[0031] During splicing, weighted fusion is performed based on the average edge weight of each edge type;

[0032] The attention temperature coefficient is dynamically adjusted by the observation variance output by the Bayesian neural network layer. When the observation variance of a node exceeds a preset threshold, the attention weight of that node as a message transmission source is multiplied by a decay factor less than 1, thereby reducing its influence on neighboring nodes.

[0033] Optionally, the Bayesian neural network layer maps the hidden states output by the gated recurrent unit network to the mean and variance of the observation distribution, including:

[0034] The state equation of the Bayesian dynamic linear model includes an environmental stress accelerated degradation term, which is determined by both the ambient temperature and the load rate.

[0035] During the particle filter update process, the weight of each particle simultaneously considers the variance of the observation distribution and the observation noise variance of the Bayesian dynamic linear model.

[0036] The process noise variance is dynamically determined by the Copula tail correlation coefficient of the loop in which the node is located. The larger the tail correlation coefficient, the larger the process noise variance.

[0037] The environmental stress accelerated degradation term is obtained by multiplying the ratio of ambient temperature to rated temperature by a power factor and then by a load acceleration factor determined by the ratio of load rate to rated load.

[0038] Optionally, the differentiated Copula selection strategy specifically includes:

[0039] For electrical faults, the Clayton Copula function is used to calculate the lower-tail correlation coefficient;

[0040] For communication failures, the Gumbel Copula function is used to calculate the upper tail correlation coefficient;

[0041] The Gaussian Copula function was used to calculate the rank correlation coefficient for optical degradation.

[0042] For complex faults, a hybrid Copula is constructed by a convex combination of multiple Copulas, and its weights are estimated online by the EM algorithm.

[0043] When the difference in AIC values ​​between two candidate Copulas is less than a preset threshold, a Bayesian model averaging is used instead of hard selection, and the weighted average tail correlation coefficient is output.

[0044] The weights of the hybrid Copula are only used to correct the health status index after they reach a steady state; otherwise, the weights of the previous period are used and marked as an uncertain state of Copula.

[0045] The correction is performed according to the following rules: the health status index of a node is reduced based on the maximum tail correlation coefficient of its neighboring nodes with the same fault type in its loop or communication subnet, and the reduction range is controlled by a preset correction strength coefficient.

[0046] Optionally, after the operation and maintenance strategy is executed, it also includes:

[0047] Calculate multidimensional maintenance effectiveness indicators, adjust model parameters in a closed loop based on the maintenance effectiveness indicators, and aggregate maintenance effectiveness data into the prior degradation coefficient of the same type of equipment to achieve cross-equipment knowledge transfer;

[0048] The multidimensional maintenance effect indicators include health recovery indicators and uncertainty calibration improvement indicators;

[0049] The health recovery indicators reflect the degree of improvement in the health status index before and after maintenance;

[0050] The uncertainty calibration improvement index reflects the degree of improvement in the observed variance before and after maintenance relative to the nominal observed variance;

[0051] When the uncertainty calibration improvement index is lower than expected, priority should be given to adjusting the dropout rate of the Bayesian neural network layer or the process noise of the Bayesian dynamic linear model.

[0052] The maintenance performance data of the same batch and model of lamps are aggregated and used to update the prior degradation coefficient of all equipment of that model.

[0053] In summary, due to the adoption of the above technical solution, the beneficial effects of the present invention are:

[0054] 1. Deep Coupling of Physical Topology Constraints and Data-Driven Approach. In constructing the heterogeneous graph of devices, this invention explicitly embeds the physical constraints of power supply topology, communication topology, and spatial topology into the calculation of edge weights. Power supply edge weights are determined by loop impedance, power supply distance, and load rate, reflecting the physical attenuation law of electrical fault propagation along power lines. Communication edge weights are dynamically updated by RSSI, SNR, and packet loss rate, reflecting the impact of communication quality degradation on the reliability of information aggregation. Spatial edge weights are determined by exponential decay based on geographical distance, reflecting the spatial clustering effect of device faults. More importantly, the edge weights are not statically calculated and fixed, but iteratively updated based on the aggregation output of the graph attention network, forming a dynamic evolutionary closed loop of "graph construction—aggregation—edge weight recalibration." Furthermore, when the power supply edge weight is below a threshold, the spatial edges of its downstream load nodes are forcibly set to zero, preventing false spatial correlations from occurring in power-destroyed devices. When the communication edge weight is below a threshold, the observation information of this node does not participate in neighbor message passing, but is injected into the Bayesian layer as a priori uncertainty of the "observation blind zone." The aforementioned physical constraint-driven graph construction mechanism solves the problem of erroneous aggregation when "data correlation is inconsistent with physical logic" in pure data-driven schemes.

[0055] 2. End-to-End Uncertainty Calibrable Architecture. This invention constructs an end-to-end uncertainty propagation mechanism from GRU to Bayesian dynamic linear model. The Bayesian neural network layer maps the hidden state of GRU to an observation distribution containing mean and variance, rather than point estimation; the weight of each particle in the particle filter update simultaneously considers the observation variance and the observation noise variance of the Bayesian dynamic linear model; the process noise variance of the Bayesian dynamic linear model is dynamically determined by the Copula tail correlation coefficient of the loop to which the node belongs—the stronger the correlation with the degradation of other devices, the greater the process noise, reflecting the influence of common stress disturbances. More uniquely, the observation variance output by the BNN is fed back to the attention temperature coefficient of the graph attention network. When a node's own sensor is unreliable (observation variance is too large), the model automatically reduces the weight of that node as a message passing source and instead borrows health status information from neighboring nodes. This "listen to neighbors when uncertain" mechanism achieves an organic linkage between spatial aggregation and uncertainty quantification.

[0056] 3. Differentiated Copula Tail Correlation Correction for Fault Type Perception. This invention does not uniformly select a single Copula function for correlation analysis, but rather differentiates the selection based on the physical matching relationship between fault type and Copula tail structure: electrical faults exhibit lower tail correlation within the power supply circuit (abnormal events such as voltage drops and harmonic impulses propagate downwards within the circuit), so Clayton Copula is selected; communication faults exhibit upper tail correlation within the subnet (synchronous degradation of the node group), so Gumbel Copula is selected; optical degradation exhibits symmetrical asymptotic correlation, so Gaussian Copula is selected. When AIC values ​​are similar, a Bayesian model is used to average and avoid the vulnerability of hard selection. The mixed Copula weights are only used for correction after reaching a steady state. The correction rule is clearly defined: the node health status index is lowered according to the maximum tail correlation coefficient of neighboring nodes with the same fault type within its circuit or subnet. This mechanism does not use Copula for prediction, but rather uses Copula to quantify the tail dependency strength between nodes in the same circuit or subnet and corrects the health status index accordingly, providing an effective mathematical tool for early warning of group faults.

[0057] 4. Multi-scale Feature Extraction and Cross-scale Modulation. This invention captures transient shocks, daily fluctuations, and trend degradation through three time windows: short-term, medium-term, and long-term. A cross-scale interactive modulation mechanism is introduced: the slope of the trend term in the long-term window serves as a degradation prior, modulating the attention weights of the features in the short-term window. When the device is in an accelerated degradation phase, the same transient anomaly is assigned a higher attention weight; when the device is in a healthy and stable phase, transient anomalies are suppressed. This cross-scale modulation solves the key problem in the Internet of Lights where "the same observed feature has different meanings at different degradation stages," avoiding misjudging occasional noise from healthy devices as a precursor to failure.

[0058] 5. Closed-loop self-evolution and cross-device knowledge transfer. This invention introduces a multi-dimensional maintenance effect evaluation after the execution of the operation and maintenance strategy. It not only evaluates the degree of recovery of the health index but also the degree of improvement of uncertainty calibration. When the calibration improvement is lower than expected, the dropout rate or process noise of the BNN is adjusted first, rather than simply increasing the noise variance. After aggregating the maintenance effect data of the same batch and model of lamps, the prior degradation coefficient of all devices of that model is updated, realizing a swarm intelligence mechanism of "repairing one lamp and optimizing a group of lamps". This enables the evaluation and prediction model to continuously learn from actual maintenance experience, fundamentally improving the adaptability and accuracy of the model in long-term operation. Attached Figure Description

[0059] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the embodiments will be briefly described below. It should be understood that the following drawings only show some embodiments of the present invention and should not be regarded as a limitation of the scope. For those skilled in the art, other related drawings can be obtained based on these drawings without creative effort, wherein:

[0060] Figure 1 This is a schematic diagram illustrating the principle and process of a project operation and maintenance intelligent assessment and prediction method based on the health status of all devices. Detailed Implementation

[0061] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only for explaining the invention and are not intended to limit the invention; that is, the described embodiments are merely some embodiments of the invention, and not all embodiments. The components of the embodiments of the invention described and shown in the accompanying drawings can generally be arranged and designed in various different configurations.

[0062] Therefore, the following detailed description of the embodiments of the invention provided in the accompanying drawings is not intended to limit the scope of the claimed invention, but merely to illustrate selected embodiments of the invention. All other embodiments obtained by those skilled in the art based on the embodiments of the invention without inventive effort are within the scope of protection of the invention.

[0063] It should be noted that relational terms such as "first" and "second" are used merely to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0064] The features and performance of the present invention will be further described in detail below with reference to embodiments.

[0065] Example 1

[0066] like Figure 1As shown, a preferred embodiment of the present invention provides a project operation and maintenance intelligent assessment and prediction method based on the overall device health status, comprising:

[0067] I. System Architecture and Data Foundation.

[0068] 1.1 Architecture of the Internet of Lights Intelligent Control System:

[0069] The intelligent lighting network control system applicable to this embodiment includes the following device levels:

[0070] Lighting fixture node layer: Composed of LED streetlights, each fixture integrates an LED light source module, driver power supply, heat dissipation structure, and optical lens. Lighting fixture nodes are the core objects of health status assessment.

[0071] Single-lamp controller node layer: Each lamp integrates a single-lamp controller, which is responsible for collecting the lamp's electrical parameters (input voltage, input current, active power, power factor, leakage current, total harmonic distortion), optical parameters (luminous flux attenuation rate, color temperature drift, illuminance uniformity), environmental parameters (ambient temperature, relative humidity, PM2.5 / PM10, vibration amplitude), and operating parameters (cumulative running time, number of switching, dimming frequency, dimming depth, average load rate), and uploading the data through narrowband communication methods such as NB-IoT.

[0072] Centralized controller node layer: Each power supply circuit is equipped with a centralized controller, which is responsible for the acquisition of circuit-level electrical parameters (three-phase voltage imbalance, neutral line current, line impedance, circuit load rate) and the issuance of circuit-level control commands.

[0073] Gateway node layer: Multiple single-lamp controllers converge to the edge gateway through communication links. The gateway collects communication parameters (received signal strength indication, signal-to-noise ratio, packet loss rate, retransmission rate, communication latency) and uploads them to the cloud platform via broadband communication methods such as 4G.

[0074] Distribution circuit node layer: includes circuit breakers, contactors, power supply lines and three-phase power supply busbars, which provide power supply for lighting fixtures in the same circuit.

[0075] The data sampling period is 5 minutes, the online evaluation period is 15 minutes, and the remaining lifetime prediction time domain is the next 30 days.

[0076] 1.2 Acquisition and Preprocessing of Multi-Source Monitoring Data:

[0077] Before the raw monitoring data enters the evaluation process, data cleaning and missing data filling are performed:

[0078] Electrical Parameter Missing Data Filling: Within the same power supply circuit, electrical parameters are constrained by Kirchhoff's laws. When voltage or current data for a node is missing, linear interpolation is performed using measured values ​​from adjacent normal equipment within the circuit and circuit topology constraints. For example, in a series power supply segment, if the input voltage of node i is missing, constrained interpolation can be performed based on the voltage measurements of its upstream and downstream nodes, combined with the line impedance of the power supply segment. The physical basis of this filling method is that electrical quantities within the same circuit follow circuit laws under steady-state operation; missing values ​​are not arbitrary interpolations, but rather deterministic interpolations subject to physical constraints.

[0079] Missing Communication Parameters Filling: Communication link quality exhibits strong autocorrelation over time. When a node's RSSI or SNR is missing, a weighted moving average is used to fill the gap by utilizing the temporal correlation of the node's historical communication quality sequence. The weights decrease as the time distance increases.

[0080] Abnormal Pulse Removal: Abnormal pulses in the monitoring data are smoothed using median filtering, and unreliable data points exceeding the equipment's rated operating range are removed using physical limiting rules. The basis for physical limiting is the equipment's rated parameters (e.g., a rated voltage range of 220V±15%; data points exceeding this range are considered sensor malfunctions or communication errors, rather than actual physical quantities).

[0081] II. Construction and dynamic evolution of tri-sided heterogeneous graphs with physical topological constraints.

[0082] 2.1 Node Types and Edge Types:

[0083] The heterogeneous equipment graph G consists of a set of nodes V and a set of edges E. Node types include lighting fixture nodes, individual lamp controller nodes, centralized controller nodes, gateway nodes, and power distribution loop nodes. Edge types include three categories:

[0084] Power supply edge: A node connecting nodes within the same power supply circuit, representing the physical path of an electrical fault propagating along the power supply line. Power supply edges are established based on the circuit topology, meaning that power supply edges exist between devices belonging to the same distribution circuit.

[0085] Communication edge: Connects individual light controllers to the gateway, and centralized controllers to the gateway, representing the transmission path of status monitoring data. Communication edges are established based on communication affiliation, meaning that devices uploading data through the same gateway have indirect communication relationships.

[0086] Spatial edge: Connects geographically adjacent luminaire nodes, representing the spatial correlation of environmental stresses (temperature, humidity, dust) and the spatial clustering effect of faults. Spatial edges are only established when the geographical distance is less than a set threshold.

[0087] 2.2 Calculation of physical constraints for edge weights:

[0088] The physical meaning and calculation principle of power supply edge weight:

[0089] The power supply edge weight characterizes the intensity of electrical fault propagation between two nodes within the same circuit. In this embodiment, the power supply edge weight... Calculate using the following formula:

[0090] ;

[0091] in The attenuation coefficient is adjustable. For loop impedance, For power supply distance, For load rate, To prevent small amounts with a denominator of zero, this formula reflects three physical principles: the greater the loop impedance and the longer the power supply distance, the more significant the attenuation of the fault current along the transmission path, and the smaller the weight; the higher the load rate, the more sensitive the node is to electrical disturbances, and the greater the weight.

[0092] The principle of dynamic update of communication edge weights:

[0093] Communication edge weights Update dynamically as follows:

[0094] ;

[0095] in For the Sigmoid function, For receiving signal strength indication, For signal-to-noise ratio, For packet loss rate, , , This is an adjustable coefficient. The weight increases as the received signal strength indicator and signal-to-noise ratio increase, and decreases as the packet loss rate increases, which aligns with the physical intuition of communication link reliability.

[0096] Distance decay principle for spatial edge weights:

[0097] Spatial edge weights Calculate using the following formula:

[0098] ;

[0099] in The spatial attenuation coefficient, This represents the geographical distance between the two nodes. And it only applies to geographical distance. A spatial edge is established when the value is less than the set threshold Dthr; when... When ≥Dthr, the spatial edge either does not exist or has a weight of 0. This formula reflects the exponential decay law of the spatial correlation of environmental stress.

[0100] The threshold Dthr can be determined by statistically analyzing the geographical distance distribution between the faulty device and its nearest faulty neighbor in historical fault records, and taking the 85% to 95% quantile of this distance distribution as Dthr, thereby ensuring that the spatial edge only covers the distance range with a significant fault clustering effect.

[0101] 2.3 Coupling Mechanism between Physical Constraints and Data-Driven Operations

[0102] One of the core features of this embodiment is that the edge weights are not statically determined and fixed after the initial construction stage, but are iteratively updated under the feedback of the aggregated output of the subsequent graph attention network.

[0103] Specifically, in each evaluation cycle:

[0104] The first step is to perform spatial aggregation of the graph attention network based on the current edge weights to obtain the spatial coupling features of each node;

[0105] The second step is to analyze the spatial coupling characteristics and identify nodes with poor aggregation effects (e.g., nodes whose aggregated features deviate too much from those of their neighboring nodes, or nodes whose aggregated features have significantly increased uncertainty).

[0106] The third step is to recalibrate the edge weights of the identified abnormal nodes: if the aggregated features of a node deviate too much from those of its neighbors, it means that the edge weights between the node and its neighbors are set too high, resulting in excessive mixing of information. In this case, the weights of the corresponding edges are reduced. Conversely, if a node has insufficient aggregated information due to its edge weights being too low (manifested as excessive variance of aggregated features), the weights of the corresponding edges are appropriately increased.

[0107] This iterative mechanism of "graph construction-aggregation-edge weight recalibration" enables the structure and weights of heterogeneous graphs to be adaptively fine-tuned based on data-driven signals within the framework of physical constraints, taking into account both physical rationality and data adaptability.

[0108] 2.4 Coupling constraints between power supply edge and spatial edge:

[0109] When the weight of any power supply edge falls below a preset power supply threshold, it indicates that the power supply link has severely deteriorated or even lost power. At this point, the downstream load node of that power supply edge cannot function properly, and the optical and environmental data it generates is no longer of reference value. The physical premise of "spatial correlation of environmental stress" between the node and its spatial neighbors no longer holds. Therefore, the spatial edge weight between the downstream load node and its spatial neighbors is forcibly reset to zero to prevent the power-destroyed device from generating false spatial correlation information and to avoid this unreliable information being transmitted to healthy neighboring nodes through spatial edges.

[0110] 2.5 Handling of observation blind spots when communication edges deteriorate:

[0111] When the weight of any communication edge is lower than a preset communication threshold, it indicates that the monitoring data of that node cannot be reliably uploaded to the cloud platform, and that node becomes an "observation blind spot." At this time, the data of that node does not participate in neighbor message passing in the spatial aggregation of the graph attention network, thus avoiding untrusted data from contaminating the health feature aggregation results of neighboring nodes.

[0112] However, the observation blind spot does not mean ignoring the node. In this embodiment, the fact that "data is unreliable" is itself treated as a priori uncertainty and injected into the observation variance of subsequent Bayesian neural network layers: the observation variance of the node is increased, representing a decrease in the confidence level of its health status estimate. The physical meaning of this approach is that when sensor data is unreliable, the system should "acknowledge its ignorance" rather than "force the use of unreliable data," thereby ensuring the authenticity of subsequent uncertainty propagation.

[0113] III. Multi-timescale time-frequency feature extraction and cross-scale modulation

[0114] 3.1 Physical Basis of the Three Time Windows

[0115] The degradation patterns of Internet of Light devices exhibit significant multi-timescale characteristics:

[0116] Short-time window (1 hour): Captures transient events ranging from seconds to minutes, such as voltage drops in power supply circuits, harmonic impulses, and sudden communication interference. These transient events are characterized by waveform spikes, abrupt changes, and short-term oscillations, with kurtosis and crazing rate as the main time-domain features. Kurtosis reflects the heavy-tailed nature of the signal distribution; the kurtosis of a normal steady-state signal approximates a Gaussian distribution (kurtosis value approximately 3), while transient impulses significantly increase kurtosis. Crazing rate is the ratio of the signal's peak value to its RMS value, used to capture short-term spike amplitudes.

[0117] Mid-time window (24 hours): Captures intraday periodic fluctuation patterns, such as load changes caused by day-night dimming strategy switching, thermal stress fluctuations caused by daily temperature cycles, and vibration amplitude fluctuations caused by traffic flow changes. The characteristics of the mid-time window are mainly statistical: the mean reflects the average daily operating level, the variance reflects the intraday fluctuation amplitude, and the sample entropy reflects the regularity and predictability of the signal sequence. When the equipment is in a healthy state, the daily fluctuation pattern has stable regularity (low sample entropy); after equipment degradation, random degradation noise is superimposed on the signal, the regularity decreases, and the sample entropy increases.

[0118] Long-term window (7 days): Captures weekly or longer-term degradation trends, such as the slow decay of LED luminous flux, the gradual decrease in driver power supply capacitance, and the gradual increase in thermal resistance of the heat dissipation structure. The core features of the long-term window are the trend slope and the energy decay rate. The trend slope is obtained by linear or nonlinear trend fitting of the 7-day series, reflecting the speed and direction of degradation; the energy decay rate reflects the proportion of the decrease in the device's output capacity (luminous flux, active power) relative to the baseline value.

[0119] 3.2 Temporal Scale Attention Fusion

[0120] Features from three time scales cannot be simply concatenated or averaged with equal weights. This embodiment employs a time-scale attention mechanism, which automatically learns the contribution of each time-scale feature to health assessment through learnable attention weights. The learning of attention weights relies on historical operational annotation data: which combinations of features from different time scales best distinguish between healthy devices and devices nearing failure; the attention mechanism automatically learns this discriminative ability. Let the short-term window feature be... The characteristics of the time window are Long-term window features Then the fused health feature vector is:

[0121] ;

[0122] in This represents the feature transformation after attention weighting. It is a short-time window feature. For mid-time window features, This is a long-term window feature.

[0123] 3.3 Cross-scale modulation mechanism:

[0124] The key innovation of this embodiment lies in the modulation effect of long-term window features on the attention weight of short-term windows.

[0125] The physical basis lies in the fact that the same transient anomaly (such as instantaneous current spikes or instantaneous voltage drops) has completely different fault prediction meanings in healthy and deteriorating equipment. For healthy equipment, instantaneous spikes may be sensor noise caused by occasional disturbances on the power grid or traffic vibrations, and do not indicate equipment failure itself; for equipment that is already in the accelerated degradation stage, instantaneous spikes may be precursor signals of drive power supply capacitor failure or solder joint microcrack propagation.

[0126] Therefore, when the slope of the trend term in the long-term window indicates that the device is in a healthy and stable phase, the attention weight of the short-term window features is suppressed—transient anomalies are interpreted as external noise and do not have a significant impact on the health status; when the slope of the trend term in the long-term window indicates that the device has entered a phase of accelerated degradation, the attention weight of the short-term window features is increased—transient anomalies are interpreted as key event signals in the degradation process and have a significant impact on the health status assessment.

[0127] An exemplary implementation of this modulation mechanism is as follows: defining the slope of the long-term window trend term. Multiply the short-term attention weights by the modulation factor ,in For example, a monotonically increasing function:

[0128] ;

[0129] in For steepness parameter, This serves as a reference threshold for the trend slope. The slope of the long-term trend term is denoted as . The steeper the trend slope (the more pronounced the degradation), the larger the modulation factor, and the higher the attention weight of the short-term features; when the trend slope is gentle, the modulation factor is close to 0.5, and the short-term features are moderately suppressed.

[0130] IV. Multi-headed graph attention space aggregation based on edge type perception.

[0131] 4.1 Side-type division of labor in multi-head attention:

[0132] In the multi-head attention mechanism of traditional graph attention networks, each attention head independently learns the attention distribution for all neighbor nodes, and finally the outputs of all heads are concatenated or averaged. This approach implicitly assumes that all neighbor relationships are homogeneous.

[0133] However, the three types of edges in the heterogeneous graph of the light network have completely different physical meanings: the power supply edge propagates electrical faults, the communication edge transmits reliable information, and the spatial edge propagates environmental stress and fault spatial accumulation. Indiscriminately mixing these three types of neighbor messages will lead to semantic confusion.

[0134] In this embodiment, the multiple attention heads of the graph attention network are explicitly divided into three groups: the power supply edge aggregation head group, the communication edge aggregation head group, and the spatial edge aggregation head group. The power supply edge aggregation head group only receives and processes feature messages of neighboring nodes connected through power supply edges, the communication edge aggregation head group only receives and processes feature messages of neighboring nodes associated through communication edges, and the spatial edge aggregation head group only receives and processes feature messages of neighboring nodes connected through spatial edges.

[0135] The three attention heads independently calculate attention weights and perform message aggregation. For node i, the attention coefficients of its neighbor node j are... Calculated by the following formula:

[0136] ;

[0137] Where W is the learnable weight matrix and a is the attention coefficient vector. This represents vector concatenation. Let be the set of neighboring nodes of node i. For the modified linear unit activation function with leakage, Let be the health feature vector of node i. Let be the health feature vector of node j, a neighbor of node i. The set of neighboring nodes of node i The health feature vector of any neighbor node k in the middle; The health feature vector of node ii is obtained by linear transformation of the learnable weight matrix W. The feature vector is obtained by linearly transforming the health feature vector of neighbor node j through the same learnable weight matrix W. The feature vector is obtained by linearly transforming the health feature vector of neighbor node k through the same learnable weight moment W; This represents concatenating the linear transformation feature vectors of node i and its neighbor node j. This involves concatenating the linear transformation feature vectors of node i and its neighbor node k. Simultaneously, the calculated attention coefficient is multiplied by the weight of the corresponding edge to reflect the impact of physical constraints on information propagation.

[0138] The outputs of the three attention heads are weighted and fused according to the average edge weight of each edge type to obtain the spatial coupling feature. :

[0139] ;

[0140] in , , These are the average edge weights for the power supply edge, communication edge, and spatial edge, respectively. It is a non-linear activation function. The output characteristics of the power supply side aggregation head group. The output characteristics of the communication edge aggregation head group. This describes the output characteristics of the spatial edge aggregation head group. The physical meaning of this is that when the power supply edge weight is high (tight electrical coupling), the spatial coupling characteristics reflect more information from the electrical domain; when the power supply edge weight is low but the spatial edge weight is high, the spatial coupling characteristics reflect more information from the environmental stress domain.

[0141] 4.2 Observational variance feedback modulates the attention temperature coefficient:

[0142] Another key mechanism in this embodiment is the feedback adjustment of the output variance of the Bayesian neural network layer on the attention temperature coefficient of the graph attention network.

[0143] The health feature vector of each node is not entirely reliable. The reliability of its feature vector decreases when a node's sensor drifts, communication links deteriorate, or data quality declines. This embodiment quantifies this reliability using the observation variance of a Bayesian neural network layer.

[0144] When the observation variance of a node exceeds a preset threshold, it indicates that the observation data of that node is unreliable. In this case, in the next round of graph attention network aggregation, the attention weight of that node as a message source is multiplied by a decay factor less than 1. ,For example:

[0145] ;

[0146] in To preset the variance threshold, This represents the observation variance of the node. When the observation variance is less than a threshold, the attenuation factor is 1 (no attenuation); when the observation variance exceeds the threshold, the attenuation factor decreases as the variance increases. The physical meaning is: a device whose own sensors are unreliable and whose own state estimation is uncertain should not propagate potentially noisy health status information to its neighbors. When aggregating information, neighboring nodes should listen less to the information provided by this unreliable neighbor and rely more on information from other reliable neighboring nodes.

[0147] This mechanism forms a closed loop of "uncertainty assessment - information aggregation strategy adjustment": the Bayesian layer tells the graph attention network "whose data is untrustworthy", and the graph attention network adjusts the trust weight of message passing accordingly.

[0148] V. GRU time series modeling.

[0149] Spatially coupled feature sequences are sequentially input into a gated recurrent unit (GRU) network. The GRU extracts long-term dependencies from the sequence through update and reset gate mechanisms. The internal computation of the GRU is as follows:

[0150] ;

[0151] ;

[0152] ;

[0153] ;

[0154] in To reset the door, To update the door, This is a candidate hidden state. This is the current hidden state. This represents the hidden state from the previous time step, where ⊙ indicates element-wise multiplication. and All are Sigmoid activation functions, and Wr, Wu, Wh, Ur, Uu, and Uh are the learnable weight matrices of the GRU network. , , For learnable bias terms in GRU networks; Let t be the spatial coupling feature vector input to the gated recurrent unit network at time t.

[0155] The input to GRU is the spatial coupling feature vector generated in each evaluation cycle. After accumulating through multiple evaluation cycles, GRU learns the spatiotemporal degenerate hidden state sequence of each node. The core property of this hidden state sequence is that it is not only a compressed representation of the node's own historical health information, but also incorporates the spatiotemporal influence of the health status of neighboring nodes on the node through the input of spatial coupling features.

[0156] VI. Uncertainty propagation in the cascade of Bayesian neural network layers and Bayesian dynamic linear models.

[0157] 6.1 BNN observation distribution mapping:

[0158] The hidden state output by the GRU is a deterministic vector, containing no uncertainty information. This embodiment inserts a Bayesian neural network layer between the GRU and the Bayesian dynamic linear model. The function of the BNN layer is to map the GRU hidden state to an observation distribution, which includes two parameters: mean and variance. The mean of the observation distribution represents the best observation estimate based on the GRU hidden state, and the variance of the observation distribution represents the degree of uncertainty of this estimate.

[0159] The difference between BNNs and traditional neural networks lies in the fact that their weights are not deterministic point estimates, but rather probability distributions. By sampling the weight distribution, BNNs can output predicted distribution parameters, thus naturally quantifying model and data uncertainty. During the inference phase, a robust estimate of the variance of the observation distribution is obtained through multiple random forward propagations (Monte Carlo Dropout or weight sampling). Specifically, BNNs output the mean and variance of the observation distribution:

[0160] ;

[0161] in and All are generated from the hidden states of the GRU through a learnable network. The mean of the observed distribution, The variance of the observed distribution.

[0162] 6.2 State-space model of BDLM:

[0163] The Bayesian dynamic linear model models the evolution of the health status index as a degradation process driven by environmental stress. The core structure of its state equation is: the health status index at the next time step equals the current health status index plus the degradation increment. Specifically, the state equation is:

[0164] ;

[0165] in Let be the health status index of node i at time t, with a value range of [0,1], where a larger value indicates a healthier node; The degradation coefficient; This is the environmental stress-accelerated degradation function; For time step; This is process noise.

[0166] The observation equation is:

[0167] ;

[0168] in For the observed values, For observation coefficients, To observe the noise variance.

[0169] Environmental stress accelerates degradation function The specific form is as follows:

[0170] ;

[0171] in For ambient temperature, Rated temperature The acceleration coefficient is the temperature. For load rate, For rated load, This is the load acceleration factor. This function reflects the accelerating effect of temperature on the degradation rate (power-law amplification) and the accelerating effect of electrical stress on the degradation rate (linear amplification).

[0172] 6.3 Coupling of process noise and Copula tail correlation coefficient:

[0173] In traditional Bayesian dynamic linear models, the process noise variance is typically set to a constant or varies slowly over time. This embodiment dynamically couples the process noise variance with the Copula tail correlation coefficient of the loop containing the node:

[0174] ;

[0175] in Let k be the variance of the basic process noise, and k be an adjustable coefficient. This is the Copula tail correlation coefficient of the loop containing this node. This represents the process noise variance. The larger the tail correlation coefficient, the larger the process noise variance.

[0176] The physical basis for this is that the process noise characterization model fails to explain random disturbances. If a node has a high tail correlation coefficient with other nodes, it indicates that the node is subjected to a greater degree of common stress disturbances (such as voltage fluctuations, communication quality fluctuations, and ambient temperature changes), and its process noise variance should increase accordingly. This coupling allows the Bayesian dynamic linear model to adaptively adjust the process noise level according to the correlation structure between devices.

[0177] 6.4 Variance Fusion Update in Particle Filtering

[0178] In the particle filter update phase, each particle represents a possible value of the health status index. The particle weight update relies on a consistency measure between the observed values ​​and the predicted particle values. In this embodiment, the observed values ​​are no longer point estimates, but rather the observation distribution output by the BNN. Therefore, the particle weight update considers two sources of variance simultaneously: the observation variance σi(t) of the BNN (characterizing the uncertainty in feature extraction and temporal modeling) and the observation noise variance σobs(t) of the BDLM (characterizing measurement noise and model observation error). The particle weight update formula is:

[0179] ;

[0180] in Let be the weight of the m-th particle at time t. Let m be the value of the health status index corresponding to the m-th particle. To observe the noise variance, This represents the observation variance output by the BNN. The sum of the two variances serves as the total variance in the particle weight update, ensuring that the final health status index and remaining lifetime probability distribution fully encompass the uncertainties across the entire process from feature extraction and temporal modeling to state estimation.

[0181] 6.5 Output of the remaining lifetime probability distribution:

[0182] The posterior distribution of the health status index is represented by a set of particles. Based on this posterior distribution, the future trajectory of the health status index distribution can be extrapolated. When the extrapolated health status index first falls below a preset failure threshold, the corresponding extrapolated time is the remaining lifespan of the device. Since the posterior distribution of the health status index is a probability distribution, the extrapolated remaining lifespan is also a probability distribution.

[0183] This probability distribution provides a quantitative basis for operational decisions: it not only tells us "what is the most likely remaining lifespan," but also "what is the probability that the remaining lifespan is less than 30 days." Operations personnel can develop preventative maintenance plans based on the 5th percentile of the remaining lifespan (the most pessimistic 5% scenario), or assess their risk tolerance for "delayed maintenance" based on the 95th percentile.

[0184] VII. Differentiated Copula Tail Correlation Correction for Fault Type Perception.

[0185] 7.1 Physical correspondence between fault types and Copula tail structure:

[0186] The joint distribution of health status indices among nodes in the same power supply circuit or the same communication subnet exhibits different types of tail-related structures, depending on the dominant failure mode.

[0187] Lower-tail correlation of electrical faults: Electrical faults in a power supply circuit (such as voltage sags, harmonic distortion, and abnormal leakage current) typically propagate within the circuit in the form of "downward impacts." When an electrical fault occurs at a node, its health status index drops sharply (entering the lower-tail region of the distribution), and other nodes in the same circuit, sharing the electrical stress, also tend to experience a synchronous decline in their health status indices. This lower-tail dependency structure, where "if one node deteriorates, the other also deteriorates," closely matches the tail characteristics of the Clayton Copula. The Clayton Copula exhibits a strictly positive lower-tail correlation coefficient and a zero upper-tail correlation coefficient.

[0188] Upper-tail correlation of communication failures: Faults within a communication subnet (such as increased packet loss rate, RSSI degradation, and disconnection) typically exhibit "collective synchronous degradation." When the performance of a gateway or base station degrades, the communication quality of all nodes within the subnet decreases simultaneously. In the distribution of health status indices, this synchronous degradation manifests as upper-tail correlation—although the absolute values ​​of the health status indices of each node differ, the probability that they are "all in a poor state" is significantly higher than the independence hypothesis. GumbelCopula exhibits characteristics of upper-tail correlation.

[0189] Symmetrical correlation of luminous flux degradation: LED luminous flux decay is a slow, gradual physical process, mainly affected by ambient temperature and operating time. There is no obvious tail dependence driven by extreme events among adjacent luminaires within the same area; instead, a symmetrical correlation structure is observed. The tail correlation coefficient of Gaussian Copula is zero, but its rank correlation coefficient can well describe the symmetrical linear correlation between gradual degradation.

[0190] Hybrid Copula for Complex Faults: In actual operation, multiple fault modes may coexist, and a single Copula function cannot fully characterize the complex tail-related structure. In this case, a convex combination of multiple Copulas is used to construct a hybrid Copula, where the weights of each component Copula reflect the dominance of the corresponding fault mode in the current period. The hybrid weights are estimated online using the EM algorithm: the E-step calculates the posterior probability of each observation sample belonging to each component Copula, and the M-step maximizes the expected log-likelihood to update the weight parameters. The weights are only used to correct the health status index after the estimation converges to a steady state.

[0191] 7.2 Handling of uncertainty in Copula selection:

[0192] When the difference in AIC values ​​between two candidate Copula functions is less than a preset threshold (usually 2), it indicates that the data is insufficient to clearly distinguish the goodness of fit between the two Copulas. In this case, a Bayesian model averaging method is used instead of hard selection: the tail correlation coefficients calculated for the two candidate Copulas are weighted and averaged according to their respective posterior model probabilities, and a comprehensive tail correlation coefficient is output.

[0193] It is also necessary to determine whether the weight estimation of the hybrid Copula has reached a steady state. If the change in the hybrid weights is less than the preset convergence threshold in two adjacent evaluation periods, it is considered to be in a steady state. If a steady state has not been reached, the weight parameters of the previous period are continued to be used, and the current state is marked as "Copula uncertain," indicating that the reliability of the current correlation analysis results of the operation and maintenance system is limited.

[0194] 7.3 Copula Tail Correction for the Health Status Index:

[0195] The results of Copula tail correlation analysis are not an independent output metric, but are directly used to adjust the health status index of each node. The adjustment rule is as follows: For node i, check all neighboring nodes j within its power supply circuit or communication subnet. If the fault type of neighboring node j is the same as that of node i, then extract the tail correlation coefficient between node i and node j. Take the largest tail correlation coefficient among all neighbors of the same type, multiply it by a preset adjustment strength coefficient, and obtain the adjustment amount. The health status index of node i is reduced by this adjustment amount. The specific adjustment formula is:

[0196] ;

[0197] in Let i be the set of other nodes within the loop or communication subnet where node i is located. For adjustable correction strength coefficient, Let be the tail correlation coefficient between node i and node j. This is an indicator function.

[0198] The physical meaning of the revised formula is: if a device has a strong tail dependency with a neighbor of the same type (such as two devices on the same circuit having the risk of electrical failure), then when the neighbor's health deteriorates, the device's own failure risk is also amplified. The larger the tail correlation coefficient, the stronger this risk transmission effect, and the greater the downward adjustment of the health status index.

[0199] Through this correction mechanism, the overall device health status assessment no longer treats each device as an independent individual, but fully considers the propagation effect of faults within the same circuit or subnet, significantly improving the early warning capability for group faults.

[0200] 8. Operation and maintenance strategy generation and closed-loop self-evolution.

[0201] 8.1 Generation of Operation and Maintenance Strategies;

[0202] Based on the revised global device health status assessment results and remaining lifespan probability distribution, differentiated operation and maintenance strategies are generated:

[0203] For equipment with a single lamp health status index below 0.6 and a remaining lifespan 5th percentile of less than 30 days, a replacement work order will be issued.

[0204] For equipment with a health status index between 0.6 and 0.75 and an excessive luminous flux attenuation rate, a cleaning and maintenance work order will be issued.

[0205] For circuits where the three-phase voltage imbalance or total harmonic distortion rate continuously exceeds the standard, a circuit maintenance work order shall be issued.

[0206] For nodes within the same communication subnet that exhibit increased packet loss rate and continuously deteriorating received signal strength, dispatch a communication module reset or gateway check work order.

[0207] For low-risk areas (where the health status index is generally above 0.85), appropriately increase the dimming depth or shorten the full-power operation time to reduce energy consumption.

[0208] For high-risk areas (where the health status index is generally below 0.7), reduce the dimming depth or limit the maximum load rate to slow down equipment degradation.

[0209] 8.2 Multidimensional maintenance effect evaluation:

[0210] After the operation and maintenance strategy is implemented, the health status index of the maintained equipment is continuously monitored to calculate the multi-dimensional maintenance effect indicators.

[0211] Health recovery indicators Defined as:

[0212] ;

[0213] in The threshold for triggering maintenance, and These are the health status indices before and after maintenance, respectively. If the health recovery index is greater than 1, it means that the health status after maintenance has not only recovered to above the trigger threshold, but the improvement has also exceeded the deviation before maintenance; if it is between 0 and 1, it means that there has been improvement but not a complete recovery; if it is less than 0, it means that the maintenance did not produce a positive effect.

[0214] in, The determination method is as follows: using historical health status indices as input and whether the equipment will malfunction within the next 30 days as a label, a receiver operating characteristic (ROC) curve is plotted, and the health status index corresponding to the point of maximum Youden index is selected as the baseline. Alternatively, provided the false alarm rate does not exceed a preset upper limit, select the point with the lowest false alarm rate as the baseline. .

[0215] Uncertainty calibration improvement index Defined as:

[0216] ;

[0217] in This represents the nominal observation variance of this type of equipment under nominal operating conditions. and These represent the observed variances before and after maintenance. Maintenance should not only improve the health status of the equipment (mean level) but also reduce the uncertainty of health status estimation (variance level). If the observed variance after maintenance fails to approach the nominal value, it indicates that there may be a problem with the selection of the maintenance method or the object being maintained.

[0218] 8.3 Closed-loop parameter adjustment:

[0219] When the uncertainty calibration improvement index is lower than expected, it indicates that there is a systematic bias in the model's estimation of uncertainty. In this case, the following parameters should be adjusted first:

[0220] Dropout rate of a Bayesian neural network layer: The dropout rate controls the inference variance of the BNN. Appropriately adjusting the dropout rate can correct the observation variance level of the BNN output, making it consistent with the true observation uncertainty.

[0221] Process noise of the Bayesian dynamic linear model: If the maintenance effect assessment indicates that the process noise setting is too low (the model is too "confident"), then the basic process noise variance should be appropriately increased; otherwise, it should be decreased.

[0222] 8.4 Cross-device knowledge transfer:

[0223] Lighting fixtures of the same batch and model exhibit similar physical degradation and degradation rates. By aggregating maintenance performance data from multiple lighting fixtures of the same model, the prior degradation coefficients for all devices of that model are updated.

[0224] For example, if the degradation rate of a certain type of LED lighting fixture in actual operation is 20% faster than initially assumed, a corrected degradation coefficient can be estimated by statistically analyzing the changes in the health status index of the maintained fixtures of that model before and after maintenance. This corrected coefficient is then applied to the degradation model of all unmaintained fixtures of that model, making their remaining life prediction more consistent with the actual degradation characteristics of that model of equipment.

[0225] This mechanism enables collective intelligence by “repairing one lamp and optimizing a whole group of lamps”: each maintenance not only solves the problem of a single lamp, but also provides valuable field degradation data for the health assessment of the same type of equipment, so that the accuracy of the health status assessment of all equipment can be continuously improved with the accumulation of maintenance activities.

[0226] The method described in this embodiment addresses the problem of assessing and predicting the health status of all devices in a smart control system for the Internet of Lights. It constructs a complete technical solution based on "physical topology constraints as the framework, uncertainty propagation throughout the entire link as the main line, tail correlation correction as the key, and closed-loop self-evolution as the guarantee." Its core principles can be summarized into the following four levels:

[0227] The first level involves a fusion modeling approach combining physical topology constraints and data-driven methods. This embodiment does not rely on purely data-driven adjacency matrix learning; instead, it explicitly embeds the power supply topology, communication topology, and spatial topology of the lighting network into the construction of the device heterogeneous graph. Power supply edge weights are determined by loop impedance, power supply distance, and load rate; communication edge weights are dynamically updated with RSSI, SNR, and packet loss rate; and spatial edge weights decay exponentially with geographical distance. Edge weights are iteratively updated within the physical constraint framework based on the aggregated output of the graph attention network. Simultaneously, coupling constraints are introduced, such as forcibly setting spatial edges to zero when power supply edges deteriorate and injecting uncertainty into the observation blind zone when communication edges deteriorate. This effectively solves the problem of erroneous aggregation caused by inconsistencies between data correlation and physical logic.

[0228] The second layer is an evaluation architecture with calibrable end-to-end uncertainty. This embodiment constructs a complete uncertainty propagation chain from the GRU hidden state to the output of the Bayesian dynamic linear model: the Bayesian neural network layer maps the GRU hidden state to an observation distribution containing mean and variance. In the particle filter update, both observation variance and observation noise variance are considered simultaneously, with the process noise variance dynamically determined by the Copula tail correlation coefficient. Simultaneously, the observation variance is fed back to the attention temperature coefficient of the graph attention network, achieving adaptive information aggregation by "listening to neighbors when uncertain about oneself." This architecture enables the final output health status index and remaining lifetime probability distribution to have credibility measurement capabilities.

[0229] The third level involves tail correlation correction for fault type perception. This embodiment uses Clayton, Gumbel, and Gaussian Copula to perform differentiated tail correlation analysis on electrical faults, communication faults, and optical degradation based on the physical matching relationship between fault types and Copula tail structures. A specific correction formula is then used to convert the tail correlation coefficient into a downward adjustment to the health status index, thereby fully capturing the fault propagation effect within the same power supply circuit or communication subnet and significantly improving the early warning capability for group faults.

[0230] The fourth level involves closed-loop self-evolution and cross-device knowledge transfer. This embodiment introduces multi-dimensional maintenance effect evaluation after the execution of the operation and maintenance strategy. It not only evaluates the degree of health recovery but also the degree of uncertainty calibration improvement, and adjusts the model parameters accordingly. At the same time, it aggregates the maintenance effect data of the same model of equipment and updates the prior degradation coefficient, realizing a swarm intelligence mechanism of "repairing one lamp and optimizing a group of lamps", so that the model can continuously improve the evaluation accuracy and adaptability in long-term operation.

[0231] Overall, the technical solution of this embodiment takes the physical characteristics of the Internet of Lights as its anchor point, and organically integrates the spatial aggregation capability of graph neural networks, the temporal modeling capability of recurrent neural networks, the uncertainty quantification capability of Bayesian methods, and the tail correlation analysis capability of Copula theory into an inseparable whole. It solves the problems of lack of physical constraints, broken uncertainty propagation, insufficient early warning of group failures, and open-loop operation and maintenance decisions in the prior art, and realizes high-precision, high-reliability, and self-evolving intelligent assessment and prediction of the health status of Internet of Lights devices across the entire domain.

[0232] The above description is merely a preferred embodiment of the present invention and is not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, and improvements made by those skilled in the art within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A project operation and maintenance intelligent assessment and prediction method based on the overall device health status, characterized in that, include: Acquire multi-source monitoring data of all network lighting devices, and construct a device heterogeneity graph based on the power supply topology, communication topology and spatial location of the network lighting devices. The device heterogeneity graph includes nodes and edges. Multi-time-scale time-frequency feature extraction is performed on the multi-source monitoring data of each node to obtain the health feature vector of each node. Then, the health feature vectors of each node are input into the graph attention network for spatial aggregation to obtain the spatial coupling features of each node. The spatial coupling features of each node are sequentially input into a gated recurrent unit network to obtain a spatiotemporal degenerate hidden state sequence for each node. The spatiotemporal degenerate hidden state sequence is then mapped to an observation distribution through a Bayesian neural network layer. The observation distribution includes the mean and variance. The observation distribution is then input into a Bayesian dynamic linear model, using the mean and variance of the observation distribution as the observation uncertainty, to estimate the health status index and remaining lifespan probability distribution of each node online. Based on a fault type-aware differentiated connection function selection strategy, tail correlation analysis is performed on the health status index of nodes in the same power supply circuit or the same communication subnet. According to the fault type, the corresponding connection function is selected from the preset connection function candidate set to calculate the tail correlation coefficient. The health status index is then corrected based on the tail correlation coefficient to obtain the overall device health status assessment result. Operation and maintenance strategies are generated based on the overall device health status assessment results and the remaining lifespan probability distribution.

2. The intelligent assessment and prediction method for project operation and maintenance based on the overall device health status as described in claim 1, characterized in that, The multi-source monitoring data includes electrical parameters, optical parameters, environmental parameters, communication parameters, and operational parameters; The heterogeneous graph of the devices includes nodes and edges. The nodes include lighting fixture nodes, single-lamp controller nodes, centralized controller nodes, gateway nodes, and power distribution loop nodes. The weight of the power supply edge is determined by the loop impedance, power supply distance, and load rate. The weight of the communication edge is dynamically updated by the received signal strength indicator, signal-to-noise ratio, and packet loss rate. The weight of the spatial edge is determined by the geographical distance at exponential decay. The edge weights of the heterogeneous graph are iteratively updated based on the feedback from the subsequent graph attention network aggregation output, forming a dynamic evolution closed loop between graph construction and aggregation. The multiple time scales include short-term windows, medium-term windows, and long-term windows; wherein, the trend features extracted by the long-term window are used as degradation priors to modulate the attention weights of the short-term window features; The graph attention network adopts a multi-head attention mechanism, in which multiple attention heads are divided according to edge type and aggregate messages from neighboring nodes of the corresponding type. During the aggregation process, the attention temperature coefficient is dynamically adjusted according to the observation variance output by the Bayesian neural network layer. When the observation variance of a node exceeds a preset threshold, the weight of that node as a message transmission source is reduced. The state equation of the Bayesian dynamic linear model includes an environmental stress-accelerated degradation term, and its process noise variance is dynamically determined by the tail correlation coefficient of the connection function of the loop in which the node is located.

3. The intelligent assessment and prediction method for project operation and maintenance based on the overall device health status as described in claim 2, characterized in that, The construction of the device heterogeneity diagram also includes the following physical constraints: The weight of the power supply side decreases with increasing loop impedance and power supply distance, and increases with increasing load rate; The weight of the communication side increases with the increase of the received signal strength indication and the signal-to-noise ratio, and decreases with the increase of the packet loss rate; Spatial edges are established only when the geographical distance is less than a set threshold, and their weight decreases as the geographical distance increases. When the weight of any power supply edge is lower than the preset power supply threshold, the spatial edge weight between the downstream load node of that power supply edge and its spatially adjacent node is forcibly reset to zero. When the weight of any communication edge is lower than the preset communication threshold, the node does not participate in neighbor message passing in the graph attention network aggregation, but it is injected into the Bayesian neural network layer as a prior uncertainty of the observation blind zone.

4. The intelligent assessment and prediction method for project operation and maintenance based on the overall device health status as described in claim 2, characterized in that, In the multi-timescale time-frequency feature extraction: The short-time window is used to capture abrupt changes and transient anomalies, and the extracted features include kurtosis and peak factor. The mid-time window is used to capture daily fluctuation patterns, and the extracted features include mean, variance, and sample entropy. The long-term window is used to capture trend degradation, and the extracted features include the slope of the trend term and the energy decay rate; Features from three time scales are weighted and fused using a time-scale attention mechanism; When the slope of the trend term in the long-term window indicates that the device is in an accelerated degradation phase, the attention weight of the short-term window features is increased; when the slope of the trend term in the long-term window indicates that the device is in a healthy and stable phase, the attention weight of the short-term window features is suppressed.

5. The intelligent assessment and prediction method for project operation and maintenance based on the overall device health status as described in claim 2, characterized in that, In the multi-head attention mechanism of the graph attention network: The attention heads are divided into three groups: power supply side aggregation heads, communication side aggregation heads, and spatial side aggregation heads, which aggregate messages from neighboring nodes of the corresponding types. During splicing, weighted fusion is performed based on the average edge weight of each edge type; The attention temperature coefficient is dynamically adjusted by the observation variance output by the Bayesian neural network layer. When the observation variance of a node exceeds a preset threshold, the attention weight of that node as a message transmission source is multiplied by a decay factor less than 1, thereby reducing its influence on neighboring nodes.

6. The intelligent assessment and prediction method for project operation and maintenance based on the overall device health status as described in claim 5, characterized in that, The Bayesian neural network layer maps the hidden states output by the gated recurrent unit network to the mean and variance of the observation distribution, including; The state equation of the Bayesian dynamic linear model includes an environmental stress accelerated degradation term, which is determined by both the ambient temperature and the load rate. During the particle filter update process, the weight of each particle simultaneously considers the variance of the observation distribution and the observation noise variance of the Bayesian dynamic linear model. The process noise variance is dynamically determined by the tail correlation coefficient of the connection function of the loop to which the node is located. The larger the tail correlation coefficient, the larger the process noise variance. The environmental stress accelerated degradation term is obtained by multiplying the ratio of ambient temperature to rated temperature by a power factor and then by a load acceleration factor determined by the ratio of load rate to rated load.

7. The intelligent assessment and prediction method for project operation and maintenance based on the overall device health status as described in claim 1, characterized in that, The specific strategy for selecting the differential join function is as follows: For electrical faults, the Clayton connection function is selected to calculate the lower-tail correlation coefficient; For communication faults, the Gombbel connection function is used to calculate the upper tail correlation coefficient; The rank correlation coefficient is calculated by selecting a Gaussian connection function for light degradation. For complex faults, a hybrid connection function is constructed by a convex combination of multiple connection functions, and its weights are estimated online by the EM algorithm. When the difference in AIC values ​​between two candidate connectivity functions is less than a preset threshold, the Bayesian model averaging is used instead of hard selection, and the weighted average tail correlation coefficient is output. The weights of the hybrid connection function are only used to correct the health status index after they reach a steady state; otherwise, the weights of the previous cycle are used and marked as an uncertain state of the connection function. The correction is performed according to the following rules: the health status index of a node is reduced based on the maximum tail correlation coefficient of its neighboring nodes with the same fault type in its loop or communication subnet, and the reduction range is controlled by a preset correction strength coefficient.

8. The intelligent assessment and prediction method for project operation and maintenance based on the overall device health status as described in claim 1, characterized in that, After the operation and maintenance strategy is executed, it also includes: Calculate multidimensional maintenance effectiveness indicators, adjust model parameters in a closed loop based on the maintenance effectiveness indicators, and aggregate maintenance effectiveness data into the prior degradation coefficient of the same type of equipment to achieve cross-equipment knowledge transfer; The multidimensional maintenance effect indicators include health recovery indicators and uncertainty calibration improvement indicators; The health recovery indicators reflect the degree of improvement in the health status index before and after maintenance; The uncertainty calibration improvement index reflects the degree of improvement in the observed variance before and after maintenance relative to the nominal observed variance; When the uncertainty calibration improvement index is lower than expected, priority should be given to adjusting the dropout rate of the Bayesian neural network layer or the process noise of the Bayesian dynamic linear model. The maintenance performance data of the same batch and model of lamps are aggregated and used to update the prior degradation coefficient of all equipment of that model.

Images