Array type sensing health evaluation method for power metering terminal integrated environment

Abstract

This invention discloses an array-based sensing health assessment method for an integrated environment of an energy metering terminal. The steps are as follows: establishing a hierarchical monitoring space within the terminal's integrated environment to obtain real-time multi-source environmental monitoring data streams; deploying a dynamic sliding time window for real-time analysis, triggering a high-speed caching mechanism to capture transient adverse operating conditions when the absolute humidity gradient change rate continuously exceeds the critical threshold for micro-condensation; using an equidistant mapping manifold algorithm to reduce the dimensionality of high-dimensional parameters to obtain a comprehensive stress vector of the physical field; using the electrical connections and heat dissipation conduction paths of the printed circuit board as physical topological edges to complete spatial mapping and construct a spatiotemporal graph neural network model; introducing the Arrhenius thermodynamic formula as a physical prior penalty term into the loss function to obtain the lifespan loss curves of key components within the terminal. This invention can comprehensively perceive the multi-physical field environmental state within the terminal, effectively improving the accuracy of health assessment and the timeliness of early warning.

Description

Technical Field

[0001] This invention belongs to the field of power equipment condition monitoring and health assessment technology, specifically, it relates to an array-type sensor health assessment method for an integrated environment of power metering terminals. Background Technology

[0002] Electricity metering terminals are core equipment in smart grid electricity information collection systems, widely deployed in substations, distribution cabinets, and outdoor metering boxes. These terminals integrate metering chips, communication modules, power management units, and related printed circuit boards, undertaking critical tasks such as electricity data collection, rate calculation, and remote communication. With the advancement of new power system construction, the operational reliability of these terminals directly impacts the accuracy of electricity metering and the quality of power supply services; therefore, accurately assessing the health status of the terminal integration environment is of paramount importance.

[0003] In existing technologies, health assessments of the integrated environment of electricity metering terminals mainly employ a combination of periodic manual inspections and single-threshold alarms. Manual inspections involve spot checks of the terminal's casing surface and interior using infrared thermal imagers or portable thermometers and hygrometers to obtain environmental parameters such as temperature and humidity. Single-threshold alarms utilize a small number of temperature and humidity sensors installed inside the terminal; when detected values ​​exceed preset upper or lower limits, an alarm signal is triggered. Some technical solutions incorporate predictive methods based on statistical analysis or simple machine learning models, using historical operating data to build regression models to predict future trends in environmental parameters.

[0004] However, the above methods have the following shortcomings. First, single-threshold alarms cannot identify the slow deterioration trend of environmental parameters, often triggering alarms only after a fault has occurred, lacking early warning capabilities. Second, the sparse layout of a small number of sensors makes it difficult to capture the spatial unevenness of the internal microenvironment of the terminal, especially in areas with severe monitoring blind spots, such as heat dissipation dead zones and areas prone to condensation. Third, existing methods usually only focus on the single physical field of temperature and humidity, ignoring the coupled impact of mechanical vibration and electromagnetic radiation on component reliability. Fourth, prediction models based on pure data are prone to producing prediction results that violate physical laws when there are insufficient samples, resulting in poor model interpretability. Summary of the Invention

[0005] Current health assessment methods for integrated environments of electricity metering terminals mainly combine periodic manual inspections with single threshold alarms. This leads to problems such as a lack of early warning capabilities, serious monitoring blind spots, insufficient sensor coverage, and poor model interpretability. To address these issues, the following invention is proposed: An array-based sensor health assessment method for an integrated environment of an energy metering terminal, comprising the following steps: Step S01: Establish a layered monitoring space within the integrated environment of the power metering terminal, set the top layer as the core acquisition area for temperature and humidity, set the bottom layer as the auxiliary acquisition area for electromagnetic and structural vibration, and perform real-time synchronous monitoring of each acquisition area to obtain a layered multi-source environment real-time monitoring data stream. Step S02: Receive the layered multi-source environment real-time monitoring data stream, deploy a dynamic sliding time window for real-time analysis, and when the absolute humidity gradient change rate of the core collection area exceeds the micro condensation critical threshold for a continuous preset number of cycles, trigger a high-speed caching mechanism to capture transient adverse working conditions and obtain transient full multi-source heterogeneous environment cache data. Step S03: Receive the transient full multi-source heterogeneous environment cache data, use the equidistant mapping manifold algorithm to perform background noise stripping and dimensionality reduction on the high-dimensional parameters, extract information characterizing the physical field abrupt change characteristics, and obtain the physical field comprehensive stress vector. Step S04: Receive the physical field comprehensive stress vector and use it as a node feature attribute. Use the electrical connection and heat dissipation conduction path of the printed circuit board inside the terminal as physical topology edges to complete spatial mapping and obtain a spatiotemporal graph neural network model with node feature attributes. Step S05: Receive the spatiotemporal graph neural network model with node feature attributes, introduce the Arrhenius thermodynamic formula as a physical prior penalty term into the model's loss function, perform forward propagation calculation, and obtain the lifetime loss curve of key components in the terminal.

[0006] Furthermore, in step S01, the layered monitoring space adopts an array-type sensor layout strategy. In the core temperature and humidity acquisition area, the temperature and humidity sensor array is arranged in a grid with equal spacing along the horizontal direction. The distance between adjacent sensors does not exceed half the diagonal length of the maximum heat dissipation area of ​​the printed circuit board, so as to ensure that the spatial resolution of the temperature and humidity gradient field meets the requirements for micro condensation detection.

[0007] Furthermore, in step S01, a triaxial accelerometer and an electromagnetic field strength probe are deployed in the electromagnetic and structural vibration assisted acquisition area. The triaxial accelerometer is installed at the fixed support point of the printed circuit board to acquire the mechanical resonance spectrum, and the electromagnetic field strength probe is installed in the near-field region of the main circuit conductor to capture the power frequency and harmonic electromagnetic radiation components.

[0008] Furthermore, in step S02, the window width of the dynamic sliding time window is adaptively adjusted according to the rate of change of ambient temperature; when the rate of change of temperature is lower than the first rate threshold, the window width takes the maximum value, and when the rate of change of temperature is higher than the second rate threshold, the window width shrinks to the minimum value. The window width is determined by linear interpolation between the first rate threshold and the second rate threshold.

[0009] Further, in step S02, the critical threshold for micro-condensation is calculated based on the Magnus approximation formula; the corresponding saturated vapor pressure is calculated by real-time collected temperature values, the partial derivative of the saturated vapor pressure with respect to spatial coordinates is used as the absolute humidity gradient, and the time derivative of the gradient is obtained to obtain the rate of change of the absolute humidity gradient. When the rate of change exceeds the critical threshold for micro-condensation for three consecutive cycles, it is determined to be a precursor event of condensation.

[0010] Furthermore, in step S03, the isometric mapping manifold algorithm first constructs the geodesic distance matrix of the high-dimensional parametric samples before performing dimensionality reduction, uses the k-nearest neighbor graph connection strategy to determine the local neighborhood, approximates the manifold geodesic distance between any two samples through the shortest path algorithm, and then embeds the geodesic distance matrix into the low-dimensional Euclidean space by multidimensional scaling to maintain the global geometric structure invariance.

[0011] Further, in step S03, each component of the physical field integrated stress vector corresponds to the thermal stress component, humidity stress component, vibration stress component and electromagnetic stress component, respectively. After normalization, the four components are weighted and combined according to their respective weight coefficients for contributing to the failure of the component. The weight coefficients are determined by the analytic hierarchy process.

[0012] Further, in step S04, the weight of the physical topology edge is determined by the thermal conductivity coefficient and electrical coupling strength between adjacent nodes. The thermal conductivity coefficient is calculated based on the thermal conductivity of the printed circuit board substrate and the copper foil coverage, and the electrical coupling strength is calculated based on the trace impedance and parasitic capacitance between adjacent nodes. The two are linearly superimposed after normalization as the edge weight.

[0013] Furthermore, in step S04, the spatiotemporal graph neural network model adopts a spatiotemporal alternating update mechanism; in the spatial dimension, it aggregates the feature information of neighboring nodes through a graph attention layer, and in the temporal dimension, it captures the temporal evolution trend of node features through a gated loop unit. After the spatial aggregation and temporal update are alternately executed for a preset number of rounds, the spatiotemporal fusion features of each node are output.

[0014] Furthermore, in step S05, the Arrhenius thermodynamic formula, as a physical prior penalty term, is specifically expressed as an exponential function of temperature, where the numerator of the exponential term is the apparent activation energy of the element material, and the denominator is the product of the Boltzmann constant and the absolute temperature; this penalty term constrains the lifetime loss rate output by the model to satisfy the physical law that it increases monotonically with increasing temperature.

[0015] Compared with existing technologies, this invention has the following advantages: First, the hierarchical array sensor layout eliminates the spatial blind spots of traditional sparse monitoring, achieving full coverage perception of temperature, humidity, vibration, and electromagnetic fields inside the terminal; Second, the dynamic sliding time window and condensation precursor detection mechanism can capture transient abnormal data in the early stages of environmental deterioration, improving the timeliness of early warning; Third, the multi-physics joint dimensionality reduction based on the equidistant mapping manifold algorithm effectively solves the curse of dimensionality problem of high-dimensional heterogeneous data; Fourth, the spatiotemporal graph neural network is constructed based on the physical topology of the printed circuit board, making full use of the thermal conduction and electrical coupling relationships between components; Fifth, the Arrhenius physical prior penalty term ensures that the lifetime loss prediction results conform to thermodynamic laws, improving the reliability and interpretability of the model under small sample conditions. Attached Figure Description

[0016] Figure 1 This is a flowchart of the overall method of the present invention; Figure 2 This is a flowchart of step S02 of the present invention. Detailed Implementation

[0017] The present invention will now be described in further detail with reference to the accompanying drawings, through specific embodiments, so that those skilled in the art can more clearly understand the present invention. The steps are as follows: This paper proposes an array-based sensor health assessment method for an integrated environment of power metering terminals.

[0018] Step S01: Establish a hierarchical monitoring space and perform real-time synchronous data acquisition from multiple environmental sources. The integrated environment for electricity metering terminals (hereinafter referred to as the terminal environment) is typically a sealed or semi-sealed metal shell structure, which integrates electricity metering modules, communication modules, power supply modules, and related printed circuit boards (PCBs). Since the terminal environment operates in outdoor distribution cabinets or metering boxes for extended periods, its internal microenvironment is affected by a combination of external temperature fluctuations, humidity intrusion, mechanical vibration, and its own electromagnetic radiation. Therefore, it is necessary to establish a layered monitoring space to achieve comprehensive perception of multiple physical fields.

[0019] Specifically, the monitoring space within the terminal environment is divided into two acquisition zones: a top layer and a bottom layer. The top layer is designated as the core temperature and humidity acquisition zone, located above the PCB, a high-risk area for heat accumulation and moisture condensation. Within this core zone, an array-style sensor layout strategy is employed, arranging temperature and humidity sensors in a horizontally spaced grid. The spacing between adjacent sensors does not exceed half the diagonal length of the PCB's largest heat dissipation area. For example, when the diagonal length of the PCB's main heat dissipation area is 80 mm, the sensor array spacing is set to no more than 40 mm. This dense layout strategy ensures that the spatial resolution of the temperature and humidity gradient field meets the requirements for microscopic condensation detection, avoiding monitoring blind spots.

[0020] The bottom layer is designated as an electromagnetic and structural vibration-assisted acquisition area, primarily covering the bottom of the terminal environment and the vicinity of the PCB fixed support structure. A triaxial accelerometer and an electromagnetic field strength probe are deployed within this acquisition area. The triaxial accelerometer is mounted at the PCB fixed support point to acquire the mechanical resonance spectrum caused by external environmental vibrations or internal relay operation. The electromagnetic field strength probe is installed in the near-field region of the main circuit conductor to capture power frequency and harmonic electromagnetic radiation components. All sensors are synchronized in real-time using a unified clock synchronization protocol to ensure strict alignment of temperature, humidity, vibration acceleration, and electromagnetic field strength data on the time axis. The data collected by all sensors is aggregated to form a layered, multi-source real-time environmental monitoring data stream. This data stream includes timestamps, spatial location identifiers, sensor type identifiers, and corresponding physical quantity measurements, serving as input for subsequent step S02.

[0021] Step S02: Threshold triggering and transient data capture based on dynamic sliding time window After receiving the hierarchical multi-source real-time environmental monitoring data stream output in step S01, a dynamic sliding time window is deployed to analyze the data stream in real time. The core function of the dynamic sliding time window is to identify statistically significant environmental deterioration trends from the continuous monitoring data stream, rather than simply providing instantaneous limit-crossing alarms.

[0022] The width of the dynamic sliding time window is adaptively adjusted according to the rate of change of ambient temperature. The rate of temperature change is defined as the first-order difference mean of the temperature sequence within the window. When the rate of temperature change is lower than a first rate threshold (e.g., ...), the window width is adjusted accordingly. When the temperature change rate exceeds the second rate threshold (e.g., 30 minutes), it indicates that the environment is in a steady state, and the window width is set to the maximum value to improve statistical stability. When the time frame indicates a rapid change in the environment, the window width shrinks to a minimum (e.g., 5 minutes) to improve temporal resolution. The window width is determined by linear interpolation between the first and second rate thresholds, calculated using the following formula:

[0023] in The current window width. Maximum window width Minimum window width, The current rate of temperature change. The first rate threshold, This is the second rate threshold.

[0024] Within each sliding window period, the focus is on the rate of change of the absolute humidity gradient in the core collection area. The critical threshold for micro-condensation is calculated based on the Magnus approximation formula. The Magnus approximation formula uses saturated water vapor pressure... Represented as temperature Functions:

[0025] in The unit is Celsius. The unit is hectopascal (hPa). The saturated vapor pressure is calculated from the real-time collected temperature values, and the partial derivative of the saturated vapor pressure with respect to spatial coordinates is used as the absolute humidity gradient. Then, the time derivative of the gradient is taken to obtain the absolute humidity gradient change rate. When the rate of change exceeds the micro-condensation critical threshold for three consecutive cycles, it is determined to be a condensation precursor event, and the system automatically triggers the high-speed cache mechanism.

[0026] Once the caching mechanism is activated, the raw sampling data from all sensors within the current sliding window and one window period before and after it are completely cached at the highest sampling rate, forming transient full-volume multi-source heterogeneous environment cache data. This cached data not only includes temperature and humidity information, but also synchronously includes vibration acceleration and electromagnetic field intensity data, providing a complete data foundation for the subsequent multiphysics joint analysis in step S03.

[0027] Step S03: Dimensionality reduction and physical field abrupt change feature extraction based on the isometric mapping manifold algorithm After receiving the transient full-volume multi-source heterogeneous environment cache data output from step S02, it is necessary to transform the high-dimensional multi-source parameters into low-dimensional feature representations that can characterize the degree of environmental degradation. Since the cache data contains four types of physical quantities—temperature field, humidity field, vibration field, and electromagnetic field—and each physical quantity has a large number of sampling channels, direct processing would face the curse of dimensionality. Therefore, the isometric mapping (ISOMAP) manifold algorithm is used for dimensionality reduction.

[0028] Before performing dimensionality reduction, the isometric mapping manifold algorithm first constructs a high-dimensional sample point from all sensor readings at each sampling time in the transient buffer data. Let the total number of sensor channels be... Then each sample is A high-dimensional parametric vector is then constructed. The geodesic distance matrix for the high-dimensional parametric samples is then used. A k-nearest neighbor graph connection strategy is employed to determine the local neighborhood, i.e., for each sample point, the nearest Euclidean distance is found. Each sample point is considered as its neighbor, and the local distance between neighbors is represented by Euclidean distance. Then, a shortest path algorithm (such as Dijkstra's algorithm) is used to approximate the manifold geodesic distance between any two samples, thus accumulating the local Euclidean distances along the manifold path to obtain the global geodesic distance. Finally, the geodesic distance matrix is ​​embedded into a low-dimensional Euclidean space using the classical multidimensional scaling (MDS) method, preserving the invariance of the global geometric structure. The dimension of the reduced space is determined by the cumulative variance contribution rate. Dimensionality reduction stops when the cumulative variance contribution rate of each dimension reaches a preset threshold (e.g., 95%).

[0029] After ISOMAP dimensionality reduction, information characterizing the abrupt changes in the physical field is extracted from the low-dimensional embedding space. Each component of the physical field composite stress vector corresponds to a thermal stress component. Humidity stress component Vibration stress components With electromagnetic stress components After normalization, the four components are weighted and combined according to their respective contribution coefficients to component failure. These weighting coefficients are determined using the Analytic Hierarchy Process (AHP). Specifically, a pairwise importance judgment matrix is ​​constructed between the four stress components. The weight vector is obtained by calculating and normalizing the eigenvector corresponding to the largest eigenvalue of the judgment matrix. Physical field combined stress vector The calculation formula is:

[0030] The physical field combined stress vector It comprehensively reflects the equivalent stress level exerted on the components by various physical fields under transient adverse working conditions in the terminal environment, and serves as the node feature attribute input for the spatiotemporal graph neural network model in the subsequent step S04.

[0031] Step S04: Graph network mapping and spatiotemporal modeling based on printed circuit board physical topology After receiving the physical field integrated stress vector output in step S03, it needs to be combined with the physical spatial structure inside the terminal to establish an analytical model that can capture spatial coupling effects and temporal evolution trends. Since there are physical interactions such as heat conduction and electromagnetic coupling between the components inside the terminal, analyzing the stress level of a single component in isolation cannot accurately assess the overall health status. Therefore, a graph neural network (GNN) is used to model the topological relationships between the components.

[0032] Key components on the PCB inside the terminal (including metering chips, power management chips, communication chips, power devices, crystal oscillators, etc.) are defined as nodes in the graph. The initial characteristic attributes of each node are the physical field comprehensive stress vector obtained in step S03. An undirected weighted graph is constructed using the actual electrical connections and heat dissipation paths of the PCB as physical topology edges. ,in For a set of nodes, Let be the set of edges. It is an adjacency matrix.

[0033] The weights of the physical topology edges are determined by the thermal conductivity and electrical coupling strength between adjacent nodes. Thermal conductivity Based on the thermal conductivity of the PCB substrate With copper foil coverage Calculation yielded:

[0034] in The thermal conductivity enhancement coefficient of copper foil, This represents the physical distance between adjacent nodes. Electrical coupling strength. Based on the trace impedance between adjacent nodes With parasitic capacitance Calculation yielded:

[0035] in The operating frequency is used. Both are normalized and then linearly superimposed as the edge weights. ,in It is a balancing factor.

[0036] After completing the spatial mapping of the graph, a spatiotemporal graph neural network model is constructed. This model employs a spatiotemporal alternating update mechanism. In the spatial dimension, the feature information of neighboring nodes is aggregated through a graph attention layer. The graph attention layer dynamically calculates the attention weight for each edge, so that neighboring nodes with closer physical interaction with the current node receive higher aggregation weights. In the temporal dimension, the temporal evolution trend of node features is captured through a gated recurrent unit (GRU). The hidden states of the GRU record the historical change trajectory of the stress level of each node. After the spatial aggregation and temporal update are alternately executed for a preset number of rounds (e.g., 3 rounds), the spatiotemporal fusion features of each node are output. Finally, a spatiotemporal graph neural network model with node feature attributes is obtained, which serves as the input for the lifetime depreciation assessment in step S05.

[0037] Step S05: Lifetime Depletion Assessment Introducing Arrhenius Physical Prior Constraints After receiving the spatiotemporal graph neural network model with node feature attributes output from step S04, the model output needs to be converted into a physically meaningful component lifespan degradation assessment result. Traditional pure data-driven models are prone to producing prediction results that violate basic physical laws when training samples are insufficient, such as predicting a decrease in the lifespan degradation rate under high-temperature conditions. To solve this problem, the Arrhenius thermodynamic formula is introduced into the model's loss function as a physical prior penalty term.

[0038] The Arrhenius equation describes the exponential relationship between chemical reaction rates and temperature, and is widely used in electronic component reliability engineering to describe the accelerating effect of temperature on component aging rates. This involves using component acceleration factors... Represented as an exponential function of temperature:

[0039] in For frequency factors, It is the apparent activation energy of the component material (in electron volts). Boltzmann constant ( ), This refers to absolute temperature (in Kelvin). Different types of component materials have different apparent activation energies; for example, the activation energy of an aluminum electrolytic capacitor is approximately... The activation energy of silicon-based semiconductor devices is approximately .

[0040] The above Arrhenius formula is embedded into the training process of a spatiotemporal graph neural network model. The model's total loss function... Loss term from data fitting Physical prior penalty terms It consists of two parts:

[0041] in This is the penalty coefficient. The data fitting loss term. Mean squared error is used to measure the deviation between the model output and the measured life loss data. Physical prior penalty term. The lifetime attrition rate output by the constraint model must satisfy the physical law that increases monotonically with increasing temperature, specifically in the form of:

[0042] in For the model to the first The predicted lifetime at each node Let be the temperature of that node. This penalty term applies a positive penalty when the model's predicted depreciation rate decreases with increasing temperature, forcing the model to learn a temperature-depreciation mapping that conforms to Arrhenius's laws of physics.

[0043] In actual training, the penalty coefficient The value of directly affects the balance between the model's accuracy in fitting data and the degree to which physical constraints are met. If the value is too small, the effect of physical constraints will be insignificant, and the model may still produce non-physical predictions in sparse data regions. If the value is too large, the model becomes overly biased towards physical priors, ignoring individual differences in the actual data. In practical applications, the optimal comprehensive evaluation metric can be achieved by searching on the validation set. The value is used to determine this hyperparameter. Furthermore, the apparent activation energy of different types of elements... The activation energy can be obtained from the component datasheet or accelerated life test report. For components with unknown activation energies, empirical values ​​from similar material systems can be used as initial estimates and fine-tuned during the training process.

[0044] After training and convergence, the model performs forward propagation calculations on each key component node within the terminal, outputting the lifespan degradation rate of each node under the current transient adverse operating conditions. Integrating the degradation rate along the time axis yields the lifespan degradation curve of the key components within the terminal. This curve, with time on the horizontal axis and cumulative lifespan degradation percentage on the vertical axis, visually displays the health degradation trend of each component under different operating conditions, providing quantitative basis for maintenance personnel to formulate preventative maintenance strategies. When the cumulative lifespan degradation percentage of a component reaches a preset health warning line, the system automatically generates a maintenance recommendation work order, indicating the remaining expected lifespan of the component and the recommended replacement time window, thereby realizing the transformation from passive fault response to proactive preventative maintenance.

Claims

1. An array-type sensor health assessment method for an integrated environment of an electricity metering terminal, characterized in that, Includes the following steps: Step S01: Establish a layered monitoring space within the integrated environment of the power metering terminal, set the top layer as the core acquisition area for temperature and humidity, set the bottom layer as the auxiliary acquisition area for electromagnetic and structural vibration, and perform real-time synchronous monitoring of each acquisition area to obtain a layered multi-source environment real-time monitoring data stream. Step S02: Receive the layered multi-source environment real-time monitoring data stream, deploy a dynamic sliding time window for real-time analysis, and when the absolute humidity gradient change rate of the core collection area exceeds the micro condensation critical threshold for a continuous preset number of cycles, trigger a high-speed caching mechanism to capture transient adverse working conditions and obtain transient full multi-source heterogeneous environment cache data. Step S03: Receive the transient full multi-source heterogeneous environment cache data, use the equidistant mapping manifold algorithm to perform background noise stripping and dimensionality reduction on the high-dimensional parameters, extract information characterizing the physical field abrupt change characteristics, and obtain the physical field comprehensive stress vector. Step S04: Receive the physical field comprehensive stress vector and use it as a node feature attribute. Use the electrical connection and heat dissipation conduction path of the printed circuit board inside the terminal as physical topology edges to complete spatial mapping and obtain a spatiotemporal graph neural network model with node feature attributes. Step S05: Receive the spatiotemporal graph neural network model with node feature attributes, introduce the Arrhenius thermodynamic formula as a physical prior penalty term into the model's loss function, perform forward propagation calculation, and obtain the lifetime loss curve of key components in the terminal.

2. The array-type sensor health assessment method for an integrated environment of an energy metering terminal according to claim 1, characterized in that: In step S01, the layered monitoring space adopts an array-type sensor layout strategy. In the core temperature and humidity acquisition area, the temperature and humidity sensor array is arranged in a grid with equal spacing along the horizontal direction. The distance between adjacent sensors does not exceed half the diagonal length of the maximum heat dissipation area of ​​the printed circuit board, so as to ensure that the spatial resolution of the temperature and humidity gradient field meets the requirements for micro condensation detection.

3. The array-type sensor health assessment method for an integrated environment of an energy metering terminal according to claim 1, characterized in that: In step S01, a triaxial accelerometer and an electromagnetic field strength probe are deployed in the electromagnetic and structural vibration assisted acquisition area. The triaxial accelerometer is installed at the fixed support point of the printed circuit board to acquire the mechanical resonance spectrum, and the electromagnetic field strength probe is installed in the near field region of the main circuit conductor to capture the power frequency and harmonic electromagnetic radiation components.

4. The array-type sensor health assessment method for an integrated environment of an energy metering terminal according to claim 1, characterized in that: In step S02, the window width of the dynamic sliding time window is adaptively adjusted according to the rate of change of ambient temperature; when the rate of change of temperature is lower than the first rate threshold, the window width takes the maximum value, and when the rate of change of temperature is higher than the second rate threshold, the window width shrinks to the minimum value. The window width is determined by linear interpolation between the first rate threshold and the second rate threshold.

5. The array-type sensor health assessment method for an integrated environment of an energy metering terminal according to claim 1, characterized in that: In step S02, the critical threshold for micro-condensation is calculated based on the Magnus approximation formula; the corresponding saturated vapor pressure is calculated by real-time collected temperature values, the partial derivative of the saturated vapor pressure with respect to spatial coordinates is used as the absolute humidity gradient, and the time derivative of the gradient is obtained to obtain the rate of change of the absolute humidity gradient. When the rate of change exceeds the critical threshold for micro-condensation for three consecutive cycles, it is determined to be a precursor event of condensation.

6. The array-type sensor health assessment method for an integrated environment of an energy metering terminal according to claim 1, characterized in that: In step S03, the isometric mapping manifold algorithm first constructs the geodesic distance matrix of high-dimensional parametric samples before performing dimensionality reduction, uses the k-nearest neighbor graph connection strategy to determine the local neighborhood, approximates the manifold geodesic distance between any two samples through the shortest path algorithm, and then embeds the geodesic distance matrix into the low-dimensional Euclidean space by multidimensional scaling to maintain the invariance of the global geometric structure.

7. The array-type sensor health assessment method for an integrated environment of an energy metering terminal according to claim 1, characterized in that: In step S03, each component of the physical field integrated stress vector corresponds to the thermal stress component, humidity stress component, vibration stress component, and electromagnetic stress component, respectively. After normalization, the four components are weighted and combined according to their respective weight coefficients for contributing to component failure. The weight coefficients are determined by the analytic hierarchy process.

8. The array-type sensor health assessment method for an integrated environment of an energy metering terminal according to claim 1, characterized in that: In step S04, the weight of the physical topology edge is determined by the thermal conductivity coefficient and electrical coupling strength between adjacent nodes. The thermal conductivity coefficient is calculated based on the thermal conductivity of the printed circuit board substrate and the copper foil coverage. The electrical coupling strength is calculated based on the trace impedance and parasitic capacitance between adjacent nodes. The two are normalized and then linearly superimposed as the edge weight.

9. The array-type sensor health assessment method for an integrated environment of an energy metering terminal according to claim 1, characterized in that: In step S04, the spatiotemporal graph neural network model adopts a spatiotemporal alternating update mechanism; in the spatial dimension, it aggregates the feature information of neighboring nodes through the graph attention layer, and in the temporal dimension, it captures the temporal evolution trend of node features through the gated recurrent unit. After the spatial aggregation and temporal update are alternately executed for a preset number of rounds, the spatiotemporal fusion features of each node are output.

10. The array-type sensor health assessment method for an integrated environment of an energy metering terminal according to claim 1, characterized in that: In step S05, the Arrhenius thermodynamic formula, as a physical prior penalty term, is specifically expressed as an exponential function of temperature, where the numerator of the exponential term is the apparent activation energy of the element material, and the denominator is the product of the Boltzmann constant and the absolute temperature. This penalty term constrains the lifetime loss rate output by the model to satisfy the physical law that it increases monotonically with increasing temperature.

Images