Remote monitoring and diagnosis system for working state of hot rod in frozen soil region based on edge calculation

Abstract

The invention relates to the technical field of engineering monitoring in a frozen soil region, and particularly discloses a remote monitoring and diagnosis system for a working state of a hot rod in a frozen soil region based on edge calculation, which comprises the following steps of: constructing a thermodynamic state matrix with a space-time correlation characteristic by deploying a multi-source sensor to synchronously collect temperature, pressure and environment wind speed data of a key part of the hot rod; a feature extraction method combining multi-scale decomposition and physical constraint is adopted to generate nonlinear dynamic response features capable of distinguishing environmental interference and real faults; a multi-dimensional diagnosis vector is constructed based on phase-space reconstruction and trajectory analysis, and comprehensive health degree parameters are generated through feature level fusion in combination with real-time environment parameters; a health state classification model with a fuzzy boundary is established, and four-level health level judgment considering the duration effect is achieved; through correlation analysis of environment disturbance and state evolution, an edge-cloud collaborative adaptive optimization mechanism is constructed.

Description

Technical Field

[0001] This invention relates to the field of engineering monitoring technology in permafrost regions, specifically to a remote monitoring and diagnostic system for the working status of heat pipes in permafrost regions based on edge computing. Background Technology

[0002] In the field of remote monitoring of heat pipes in permafrost regions, existing technologies mostly rely on general data-driven models for condition diagnosis. However, when dealing with special locations with unique microclimate characteristics (such as mountain passes with frequent strong winds or valleys with drastic temperature variations), these models often misjudge normal physical phenomena such as strong winds and heat dissipation as faults because the training data does not fully cover local extreme conditions, leading to frequent false alarms. Furthermore, traditional models experience performance degradation over long-term operation due to environmental changes and equipment aging, and lack self-optimization capabilities, making it difficult to maintain diagnostic accuracy in harsh environments. These problems have been rarely explored in depth in existing research, yet they seriously affect the reliability and availability of the system.

[0003] The problem addressed by this invention is that when heat pipes are located in mountain passes with frequent strong winds or valleys with drastic temperature changes, traditional models, lacking an understanding of the underlying physical mechanisms, may incorrectly identify normal heat dissipation caused by strong winds as "internal blockage" or "performance degradation," generating numerous false alarms. This not only wastes operational resources but may also mask genuine faults.

[0004] This invention constructs a hybrid diagnostic mechanism driven by both data and physics by embedding thermodynamic physical equations as constraints into the feature extraction process. This enables the system to accurately distinguish between environmental interference and real faults, thereby improving diagnostic accuracy in special local environments. Summary of the Invention

[0005] The purpose of this invention is to provide a remote monitoring and diagnostic system for the working status of heat pipes in permafrost regions based on edge computing, so as to solve the problems mentioned above.

[0006] The objective of this invention can be achieved through the following technical solutions: A remote monitoring and diagnostic system for the operating status of heat pipes in permafrost regions based on edge computing includes: The data acquisition and state matrix construction module is used to acquire time-series monitoring data of the heat pipe under the monitoring of multiple source sensors, and construct a thermodynamic state matrix reflecting the internal thermodynamic state of the heat pipe based on the time-series monitoring data. The multi-scale feature extraction and fusion module is used to perform multi-scale tensor decomposition on the thermodynamic state matrix, extract spatiotemporal feature modes at different time scales, and generate nonlinear dynamic response features characterizing the working state of the heat pipe through an adaptive feature fusion algorithm. The health parameter inference and generation module constructs a multi-dimensional diagnostic vector based on nonlinear dynamic response characteristics, combines the real-time detection signal from the environmental wind speed sensor, and generates comprehensive health parameters through feature-level fusion and inference calculation. The health status classification and judgment module takes the comprehensive health parameters and inputs them into the preset health status classification model, and outputs the corresponding health level judgment result. The online adaptive optimization module is used to track the changes in multi-source time-series signals of the heat pipe's working status and the evolution trend of environmental disturbance intensity in real time during continuous monitoring, and to optimize the health status classification model online based on the tracking results.

[0007] As a further aspect of the present invention: the process of constructing the thermodynamic state matrix is ​​as follows: The system acquires multi-point temperature sequences and ambient wind speed sequences for the evaporation and condensation sections of the heat pipe, and uses hardware synchronization pulse signals to ensure accurate alignment of timestamps for each channel's data. The collected multi-source time-series data are reorganized into three-dimensional data according to a preset time window to construct a three-dimensional data tensor with time axis, sensor spatial distribution axis and physical quantity type axis as dimensions; Tensor expansion and dimensionality reduction are performed on the three-dimensional data tensor. A thermodynamic state matrix with spatiotemporal correlation characteristics is generated by a feature extraction method that preserves the main feature patterns.

[0008] As a further aspect of the present invention: the extraction process of the spatiotemporal feature modes is as follows: By setting multiple time windows of different lengths to perform sliding segmentation on the thermodynamic state matrix, three sets of time-scale data fragments at the second, minute, and hour levels are obtained. Spatiotemporal feature decoupling is performed on data segments at each time scale to separate and extract spatial sensor distribution features from temporal evolution features; The separated spatiotemporal features are input into a deep feature encoder to generate scale-specific spatiotemporal feature modes. Cross-scale correlation analysis was performed on spatiotemporal feature modes at different time scales to screen out core feature modes that showed responses at multiple scales, which were denoted as spatiotemporal feature modes.

[0009] As a further aspect of the present invention: the generation of nonlinear dynamic response characteristics characterizing the working state of the heat pipe specifically includes: Based on real-time wind speed and temperature gradient, spatiotemporal feature modes are dynamically grouped and selected. Feature selection criteria are established through dual verification of feature reconstruction fidelity and diagnostic consistency. State evolution analysis was performed on the selected feature sequences to identify the transient response stage and steady-state operation stage during the operation of the heat pipe; The transient response features and steady-state operation features are dynamically coupled and encoded to generate a nonlinear dynamic response feature vector that simultaneously contains fast response characteristics and long-term operation characteristics.

[0010] As a further aspect of the present invention: the construction of the multidimensional diagnostic vector specifically includes: Phase space reconstruction is performed based on nonlinear dynamic response characteristics, which expands the one-dimensional feature sequence into a multi-dimensional state space representation. Cluster analysis of the system dynamic trajectory in the reconstructed phase space is performed to identify motion patterns that characterize different operating states; Extract the trajectory curvature, motion velocity, and distribution density features of each motion mode to form a set of geometric features in the state space; By encoding the geometric feature set and the nonlinear dynamic response features in parallel, a multidimensional diagnostic vector with physical interpretability is generated.

[0011] As a further aspect of the present invention: the specific process for generating the comprehensive health parameters is as follows: Dynamically filter and reorganize multidimensional diagnostic vectors using real-time wind speed signals; Time-series alignment and interactive analysis were performed between wind speed change patterns and diagnostic vector features; By coordinating and integrating diagnostic results from multiple time scales, the characteristic differences between transient disturbances and actual faults can be identified. An adaptive weight allocation based on information value assessment generates a comprehensive health parameter with confidence level labels.

[0012] As a further aspect of the present invention: the construction process of the health status classification model is as follows: Based on the comprehensive health parameter sequence in historical operational data, a characteristic spatial distribution map of health status is constructed. In the feature space, the boundary regions of different health states are fuzzily divided to establish a health state classification boundary with a transition zone; Perform state persistence analysis on the dynamic change patterns of health parameters; The feature spatial distribution and time series analysis results are integrated to form the final health level determination result.

[0013] As a further aspect of the present invention: the decision process of the health status classification model includes: Based on the preset four-level health status classification standard, the comprehensive health parameters are mapped to the corresponding health levels; When the overall health parameter is greater than the first preset threshold, it is determined to be in normal working condition; When the overall health parameter is between the second preset threshold and the first preset threshold and the duration exceeds 24 hours, it is determined to be a state of mild performance degradation. When the overall health parameter is between the third preset threshold and the second preset threshold and the duration exceeds 12 hours, it is determined to be a state of moderate performance degradation. When the overall health parameter is less than the third preset threshold and the duration exceeds 6 hours, it is judged as a serious fault state. The instantaneous state, short-term state, and continuous state are classified and confirmed separately. The final output is a health level determination result that has both time persistence and state stability.

[0014] As a further aspect of the present invention: the real-time tracking of the multi-source time-series signal changes and the evolution trend of environmental disturbance intensity in the operating state of the heat pipe specifically includes: An evolution curve of environmental disturbance intensity is constructed. By continuously collecting wind speed and ambient temperature data and calculating their change gradient, the dynamic impact of the external environment is quantified. Simultaneously monitor the temporal variation patterns of evaporation section temperature, condensation section temperature, and operating pressure; Extract early feature patterns characterizing heat rod performance degradation from multi-source time-series signals; By analyzing the correlation between environmental disturbances and heat pipe response, the environmental sensitivity characteristics of abnormal operating conditions can be identified.

[0015] As a further aspect of the present invention: the online optimization of the health status classification model specifically includes: Establish a mapping relationship between the parameters of the health status classification model and environmental disturbances, and dynamically adjust the decision boundary of the health status classification model according to environmental characteristics; When the tracking results show that the system status is undergoing continuous changes, the health status classification model will be updated. While maintaining the qualitative nature of the health status classification model, the classification threshold is gradually optimized. The optimization effect was verified by using historical state sequences.

[0016] The beneficial effects of this invention are: (1) This invention constructs a diagnostic mechanism driven by both physical laws and data features by embedding the physical equations of the heat pipe operation as intrinsic constraints into the feature extraction process. This enables the system to accurately identify the difference between the strong wind heat dissipation effect and the actual fault at special locations such as the Kunlun Mountain Pass, significantly reducing the misjudgment rate in complex local environments and solving the industry problem of adaptability of general models in special scenarios.

[0017] (2) This invention achieves online adaptive optimization capability by tracking the correlation features between environmental disturbances and state evolution in real time at the edge, combined with incremental learning and dynamic threshold adjustment in the cloud. This mechanism enables the system to maintain diagnostic accuracy continuously under unattended conditions, overcomes the technical bottleneck of model performance degradation due to operating condition drift in extreme environments, and realizes autonomous optimization throughout the entire life cycle of the system. Attached Figure Description

[0018] The invention will now be further described with reference to the accompanying drawings.

[0019] Figure 1 This is a flowchart of the system of the present invention. Detailed Implementation

[0020] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0021] Please see Figure 1 As shown, this invention is a remote monitoring and diagnostic system for the operating status of heat pipes in permafrost regions based on edge computing, comprising: The data acquisition and state matrix construction module is used to acquire time-series monitoring data of the heat pipe under the monitoring of multiple source sensors, and construct a thermodynamic state matrix reflecting the internal thermodynamic state of the heat pipe based on the time-series monitoring data. The multi-scale feature extraction and fusion module is used to perform multi-scale tensor decomposition on the thermodynamic state matrix, extract spatiotemporal feature modes at different time scales, and generate nonlinear dynamic response features characterizing the working state of the heat pipe through an adaptive feature fusion algorithm. The health parameter inference and generation module constructs a multi-dimensional diagnostic vector based on nonlinear dynamic response characteristics, combines the real-time detection signal from the environmental wind speed sensor, and generates comprehensive health parameters through feature-level fusion and inference calculation. The health status classification and judgment module takes the comprehensive health parameters and inputs them into the preset health status classification model, and outputs the corresponding health level judgment result. The online adaptive optimization module is used to track the changes in multi-source time-series signals of the heat pipe's working status and the evolution trend of environmental disturbance intensity in real time during continuous monitoring, and to optimize the health status classification model online based on the tracking results.

[0022] In the data acquisition and state matrix construction module, data acquisition is accomplished through sensor arrays installed at key locations on the heat pipe. Three temperature sensors are positioned at the top, middle, and bottom of the evaporation section; two temperature sensors are positioned at the top and middle of the condensation section; and an ambient wind speed sensor is also included. These sensors are connected via a multi-channel synchronous acquisition device with a built-in hardware synchronization pulse generator that produces one synchronization pulse signal per second. When the synchronization pulse signal is triggered, all sensors simultaneously acquire data, ensuring that the data from each channel has identical timestamps. The acquired data is transmitted to the edge computing node via wired transmission, with cyclic redundancy check (CRC) used during transmission to ensure data integrity.

[0023] After obtaining multi-source time-series data, three-dimensional data reconstruction processing is performed. The data from 1800 consecutively acquired time points are treated as a single processing unit, corresponding to a 30-minute monitoring period. Within each processing unit, temperature sensor data are arranged according to spatial location, forming a sensor spatial distribution axis; different physical quantities are arranged according to type, forming a physical quantity type axis; and the time series is arranged according to the order of acquisition, forming a time axis. This arrangement constructs a three-dimensional data tensor. The first dimension of this tensor contains 1800 time points, the second dimension contains 6 sensor location points, and the third dimension contains 2 physical quantity types. This three-dimensional data structure can completely preserve the spatiotemporal correlation characteristics of the data.

[0024] The three-dimensional data tensor is expanded and its dimensionality reduced. First, the three-dimensional data tensor is expanded along the time axis to form a two-dimensional matrix. The rows of this matrix correspond to different time points, and the columns correspond to combinations of sensor locations and physical quantity types. Eigenvalue decomposition is performed on this two-dimensional matrix to calculate its eigenvalues ​​and eigenvectors. The eigenvectors corresponding to the first k eigenvalues ​​are selected in descending order of eigenvalues, where the value of k is chosen such that the sum of the retained eigenvalues ​​accounts for more than 95% of the sum of all eigenvalues. These eigenvectors are used to construct a projection matrix, projecting the original data onto a new feature space. Through this process, while preserving the main feature patterns, the dimensionality of the data is reduced from the original high-dimensional space to k dimensions, resulting in a thermodynamic state matrix with better feature representation capabilities and computational efficiency.

[0025] In the multi-scale feature extraction and fusion module, the first step is to divide the time window into multiple scales. Three different time window lengths are set: a second-level window containing 60 consecutive data points, corresponding to a 1-minute monitoring duration, with a window sliding step of 10 data points; a minute-level window containing 600 consecutive data points, corresponding to a 10-minute monitoring duration, with a window sliding step of 60 data points; and an hour-level window containing 3600 consecutive data points, corresponding to a 1-hour monitoring duration, with a window sliding step of 600 data points. This setup allows for the extraction of data segments with different temporal resolutions from the thermodynamic state matrix, each data segment containing complete feature information at its corresponding time scale.

[0026] Spatiotemporal feature decoupling was performed on data segments at each time scale. When processing second-level data segments, the time dimension was first fixed, and the spatial correlation between the six sensor locations was analyzed. The covariance matrix of the sensor data was calculated, and spatial principal components were obtained through eigenvalue decomposition. These spatial principal components reflect the correlation patterns between different sensors. Then, the spatial dimension was fixed, and the variation of each sensor's data over time was analyzed. Frequency domain analysis was performed on the time series to extract the main frequency components. Through this decoupling process, the original data was decomposed into components representing spatial distribution characteristics and components representing temporal evolution characteristics. The same processing method was used for data segments at other time scales.

[0027] The decoupled spatiotemporal features are input into a feature encoding network. This network consists of three fully connected layers: the first layer has 64 nodes and uses the hyperbolic tangent activation function; the second layer has 32 nodes and uses the linear activation function; and the third layer has 16 nodes and uses the hyperbolic tangent activation function. During network training, mean squared error is used as the loss function, and the network parameters are optimized using gradient descent. After training, the output of the third layer is used as the spatiotemporal feature mode for that time scale. A set of 16-dimensional feature vectors is generated for each time scale, representing the specific operating mode of the heat pipe at that time scale.

[0028] Cross-scale correlation analysis was performed on spatiotemporal feature modes at different time scales. The correlation coefficients between second-level and minute-level features, and between minute-level and hour-level features, were calculated. A correlation coefficient threshold of 0.7 was set; a feature was considered a multi-scale core feature when the correlation coefficients at two adjacent time scales both exceeded this threshold. These core features were marked as spatiotemporal feature modes for subsequent feature fusion processing. This selection method ensures that the final features used have good stability and representativeness.

[0029] The spatiotemporal characteristic modes are dynamically grouped based on real-time environmental parameters. Real-time wind speed data is updated every 10 seconds, and the temperature gradient is obtained by calculating the temperature difference between adjacent sensors. When the wind speed is greater than 5 m / s and the temperature gradient is less than 2 degrees Celsius, the characteristic mode related to rapid response is selected; when the wind speed is less than 2 m / s and the temperature gradient is greater than 5 degrees Celsius, the characteristic mode related to steady-state operation is selected. The specific implementation of feature grouping is accomplished by querying a predefined feature-environment mapping table, which contains feature combination schemes that should be used under different environmental conditions.

[0030] Feature selection criteria are established through dual validation. Feature reconstruction fidelity validation is achieved by calculating the ability of selected features to reconstruct the original data, requiring a reconstruction error of less than 0.1. Diagnostic consistency validation is achieved by comparing the consistency of the current feature selection results with the historical usage of features in the same period, requiring a consistency coefficient greater than 0.8. Only feature combinations that meet both conditions will be adopted. The specific process of feature selection is as follows: first, features are initially screened based on environmental parameters; then, the reconstruction error and consistency coefficient are calculated; and finally, the final feature combinations to be used are determined based on the validation results.

[0031] State evolution analysis is performed on the filtered feature sequences. Feature data from 120 consecutive time points are collected, and the rate of change of the feature amplitude at each time point is calculated. When the rate of change exceeds the threshold of 0.05 for three consecutive time points, it is considered to have entered the transient response stage; when the rate of change is below the threshold of 0.01 for ten consecutive time points, it is considered to have entered the steady-state operation stage. The specific steps for state determination are: first, calculate the first difference of the feature sequence; then, perform moving average filtering; and finally, determine the current operating stage based on the set threshold.

[0032] Transient response features and steady-state operating features are dynamically coupled and encoded. Transient response features are derived from data at five time points before and after the state transition, calculating the difference between the maximum and minimum values. Steady-state operating features are derived from data at 20 consecutive time points during the stable phase, calculating their average value. These two feature values ​​are then weighted in a 6:4 ratio, with the transient feature having a weight of 0.6 and the steady-state feature having a weight of 0.4. The resulting feature vector contains 32 dimensions: the first 16 dimensions represent fast response characteristics, and the last 16 dimensions represent long-term operating characteristics. This feature vector is the final nonlinear dynamic response feature.

[0033] In the health parameter inference and generation module, phase space reconstruction is performed based on nonlinear dynamic response features. A sequence of nonlinear dynamic response features at 100 consecutive time points is used as the processing object. The delay time is set to 5 data points, and the embedding dimension is 3. The one-dimensional feature sequence is expanded into a three-dimensional state space representation using the delayed coordinate method. Specifically, the 1st, 6th, and 11th data points are used as the first state point, the 2nd, 7th, and 12th data points as the second state point, and so on, until the 90th, 95th, and 100th data points are used as the last state point. Through this reconstruction, 31 three-dimensional state points are obtained, which constitute the system's trajectory in phase space.

[0034] Cluster analysis was performed on the system dynamic trajectory in the reconstructed phase space. A density-based clustering method was adopted, with a neighborhood radius of 0.1 and a minimum number of points to be included (5). First, the distance between each state point and its nearest neighbor was calculated, and points whose distance was less than the neighborhood radius and which were mutually reachable were grouped into the same cluster. This method can divide the trajectory points in the phase space into 3 to 5 different clusters, each representing a typical operating state pattern. The silhouette coefficient was used to evaluate the clustering effect during the clustering process, requiring a silhouette coefficient greater than 0.6; otherwise, the neighborhood radius was adjusted and clustering was performed again.

[0035] Geometric features of each motion mode are extracted. For each cluster, its trajectory curvature, velocity, and distribution density are calculated. Trajectory curvature is obtained by calculating the rate of change of the angle of the line connecting adjacent state points, and the average curvature of all state points is taken as the trajectory curvature feature of the cluster. Velocity is obtained by calculating the Euclidean distance between adjacent state points divided by the time interval, and the average velocity of all state points is taken as the velocity feature of the cluster. Distribution density is obtained by calculating the ratio of the number of state points within the cluster to the cluster volume. Finally, a set of state-space geometric features containing 9 to 15 geometric features is formed.

[0036] The geometric feature set and the nonlinear dynamic response feature are encoded in parallel. The geometric feature set contains 9 to 15 feature values, and the nonlinear dynamic response feature contains 32 feature values. These two sets of features are concatenated sequentially to form a 41 to 47-dimensional feature vector. This feature vector is then standardized so that the values ​​of each dimension are between 0 and 1. The standardization method involves taking the maximum and minimum values ​​of each dimension across all samples, linearly mapping the original values ​​to the 0-1 interval. The standardized feature vector is the final multidimensional diagnostic vector.

[0037] Multidimensional diagnostic vectors are dynamically filtered and recombined using real-time wind speed signals. Real-time wind speed data is updated every 10 seconds, and the operating status is divided into three intervals based on wind speed: low wind speed (less than 3 m / s), medium wind speed (3 to 6 m / s), and high wind speed (greater than 6 m / s). Each interval corresponds to a specific set of diagnostic vector filtering rules. The low wind speed interval primarily selects diagnostic vector dimensions related to heat transfer efficiency, the medium wind speed interval primarily selects dimensions related to convection cooling, and the high wind speed interval primarily selects dimensions related to stability. The number of filtered diagnostic vector dimensions is 60% to 80% of the original dimensions.

[0038] The wind speed change pattern and diagnostic vector features are time-series aligned and interactively analyzed. The wind speed data sequence and diagnostic vector sequence for the most recent 30 minutes are taken and time-series aligned at 10-second intervals. The correlation coefficient between the wind speed change rate and the diagnostic vector change rate is calculated; a significant correlation is considered to exist when the absolute value of the correlation coefficient is greater than 0.7. For correlated feature pairs, their phase relationship is further analyzed to calculate the time difference between wind speed change leading or lagging the diagnostic vector change. Through this analysis, the degree of influence and response time of wind speed change on the heat pipe's operating status can be determined.

[0039] By coordinating and integrating diagnostic results across multiple time scales, the characteristic differences between transient disturbances and true faults are identified. Three time scales are set: a second-level scale to observe transient response, a minute-level scale to observe short-term changes, and an hour-level scale to observe long-term trends. Anomaly indices of the diagnostic vector are calculated at each time scale: the second-level anomaly index is obtained by calculating the deviation between the current value and the historical mean; the minute-level anomaly index is obtained by calculating the coefficient of variation within a sliding window; and the hour-level anomaly index is obtained by calculating the trend slope. When the anomaly indices at all three scales simultaneously exceed a threshold, it is determined to be a true fault; when only the anomaly indices at some scales exceed the threshold, it is determined to be a transient disturbance.

[0040] An adaptive weighting system based on information value assessment is used. Information value is determined by calculating the discriminative power of each diagnostic vector dimension across different health states. First, health states are divided into three categories: normal, slightly deteriorated, and severely deteriorated. Then, the information gain ratio (IVR) of each feature dimension across different categories is calculated. Feature dimensions with an IVR greater than 0.3 are considered to have high information value and assigned a weight of 0.7; feature dimensions with an IVR between 0.1 and 0.3 are assigned a weight of 0.3; and feature dimensions with an IVR less than 0.1 are assigned a weight of 0.1. Finally, the weighted average of all features is used as the overall health score parameter, ranging from 0 to 1. The number of features used in the calculation and the total information value are recorded as confidence level markers.

[0041] In the health status classification and decision module, a feature space distribution map of health status is constructed based on the comprehensive health parameter sequence from historical operational data. At least one year of historical operational data is collected, including comprehensive health parameters under different seasons and environmental conditions. Health status is initially divided into four categories: normal working state, slight performance degradation state, moderate performance degradation state, and severe failure state. Density clustering is used to analyze the distribution of each status category in the feature space, calculating the center point and distribution range of each category. The feature space distribution map is represented by a two-dimensional grid, with the horizontal axis representing the health parameter values ​​and the vertical axis representing the stability index of the health parameters. Each grid cell records the frequency of different health statuses within that region.

[0042] In the feature space, the boundary regions of different health states are fuzzily partitioned to establish health state classification boundaries with transition bands. A transition band with a width of 0.05 is set at the boundary region between adjacent health state categories. Within the transition band, the affiliation of a health state is determined by a membership function. The membership function adopts a trapezoidal function, with a membership degree of 0 at the beginning of the transition band and a membership degree of 1 at the end. For example, the transition band between the normal working state and the slightly degraded performance state is set between the health parameter 0.75 and 0.8. When the health parameter falls within this range, the membership degrees belonging to both states are calculated simultaneously, and the state with the higher membership degree is taken as the current state. This fuzzy partitioning method can better handle uncertainties near the state boundaries.

[0043] A persistence analysis was conducted on the dynamic change patterns of health parameters. A 24-hour continuous sequence of health parameters was taken, and its changing trends and fluctuation characteristics were calculated. First, the sequence was subjected to first-order differencing to obtain the rate of change of the parameters. Then, the autocorrelation coefficient of the sequence was calculated to analyze the persistence characteristics of the parameters. The persistence criteria were set as follows: when the autocorrelation coefficient was greater than 0.6 and the rate of change was less than 0.01 per hour, the state was considered persistent; when the autocorrelation coefficient was less than 0.3 or the rate of change was greater than 0.05 per hour, the state was considered non-persistent. The results of the persistence analysis serve as an important reference for state determination.

[0044] The feature spatial distribution and time series analysis results are fused to form the final health level determination result. Decision rules for state determination are established: when the feature spatial distribution shows that the health parameter clearly falls within a certain state region, and the time series analysis shows that the state is persistent, it is directly determined to be in that state; when the feature spatial distribution shows that the health parameter falls within a transition zone, or the time series analysis shows that the state is not persistent, a review procedure is initiated. The review procedure includes steps such as extending the observation period, checking the influence of environmental factors, and comparing with historical data from the same period. This fusion method ensures that the health level determination result considers both the current parameter value and the trend of parameter changes.

[0045] Based on a pre-defined four-level health status classification standard, the comprehensive health parameter is mapped to the corresponding health level. The first pre-defined threshold is 0.8; when the comprehensive health parameter is greater than 0.8, the system enters the normal working state determination process. The second pre-defined threshold is 0.6; when the comprehensive health parameter is between 0.6 and 0.8, the system enters the mild performance degradation state determination process. The third pre-defined threshold is 0.4; when the comprehensive health parameter is between 0.4 and 0.6, the system enters the moderate performance degradation state determination process. When the comprehensive health parameter is less than 0.4, the system enters the severe fault state determination process. These thresholds were determined by analyzing a large amount of historical data, comprehensively considering the performance characteristics of the heat pipes and maintenance experience.

[0046] To determine a normal operating state, a health parameter greater than 0.8 is required, with no duration requirement. To determine a state of slight performance degradation, two conditions must be met simultaneously: the health parameter must be between 0.6 and 0.8, and this state must persist for more than 24 hours. The duration is calculated from the first time the health parameter falls into this range, allowing for brief fluctuations of no more than 2 hours, but the fluctuation amplitude cannot exceed 0.1. If the fluctuation exceeds this range, the duration needs to be recalculated.

[0047] To determine a moderate performance degradation state, the health parameter must be between 0.4 and 0.6 for more than 12 hours. To determine a severe fault state, the health parameter must be less than 0.4 for more than 6 hours. For both performance degradation and fault states, the state parameters must remain relatively stable when calculating the duration, i.e., the rate of change per hour must not exceed 0.05. If the rate of change exceeds this limit, the state is not yet stable and further observation is required.

[0048] Transient states, short-term states, and persistent states are classified and confirmed separately. Transient states refer to state changes lasting less than 1 hour; these changes are usually caused by environmental disturbances and do not change the health level determination. Short-term states refer to state changes lasting between 1 and 6 hours; these changes need to be recorded but do not require immediate adjustment of the health level. Persistent states refer to state changes lasting more than 6 hours; these changes will trigger a reassessment of the health level. When confirming persistent states, interference factors such as environmental anomalies and sensor malfunctions need to be ruled out.

[0049] The final output provides a health level assessment result that is both time-dependent and state-stable. The health level assessment result comprises four parts: current health level, level confidence score, state duration, and recent trend. The current health level is one of four levels, represented by numbers 1 to 4. The level confidence score is calculated by analyzing parameter stability, environmental consistency, and historical compliance, ranging from 0 to 1. The state duration records the length of time the current level has been maintained. The recent trend is represented by rising, falling, or stable. This output format includes both current state information and historical state change information, providing a complete basis for maintenance decisions.

[0050] In the online adaptive optimization module of the model, wind speed and ambient temperature data are collected every 30 seconds when constructing the evolution curve of environmental disturbance intensity. The difference between data points at adjacent time points is calculated to obtain the gradients of wind speed change and temperature change. The absolute value of the wind speed change gradient is taken, while the sign of the temperature change gradient is retained. The two gradient values ​​are weighted and summed in a 7:3 ratio, with the weighting coefficients determined through historical data analysis. The result of the weighted summation is the environmental disturbance intensity value. The environmental disturbance intensity values ​​are continuously recorded for 24 hours, and the points are connected by a smooth curve to form the evolution curve. The vertical axis of the curve represents the disturbance intensity, ranging from 0 to 1, and the horizontal axis represents time. This curve allows for a visual observation of the change process and intensity characteristics of the environmental disturbance.

[0051] When simultaneously monitoring the temporal changes of evaporator section temperature, condenser section temperature, and operating pressure, all three parameters are acquired at the same frequency, recording data every 10 seconds. Temperature data is obtained from sensors installed at different locations on the heat pipe, and pressure data is obtained from a pressure transmitter installed in the pipeline. During monitoring, the current value, trend, and fluctuation characteristics of each parameter are recorded. The trend is obtained by calculating the slope of a linear regression of the data from the most recent 5 minutes, and the fluctuation characteristics are obtained by calculating the standard deviation of the data from the most recent 30 minutes. The monitoring results of the three parameters are displayed using different color curves for easy comparison and analysis. When an abnormal change is detected, the data acquisition frequency is automatically increased to once per second, and observation continues for 10 minutes.

[0052] When extracting early characteristic patterns representing heat pipe performance degradation from multi-source time-series signals, three types of features are emphasized: trend features, periodic features, and abrupt change features. Trend features are obtained by calculating the rate of change of the average value of data for the same time period each day, observing whether there is a continuous upward or downward trend. Periodic features are obtained through spectral analysis, examining changes in the amplitude of daily and yearly periodic features. Abrupt change features are obtained by calculating the first difference of the data, identifying abnormal peaks and troughs. These features are extracted using a sliding time window with a window length of 7 days, with feature values ​​updated daily. When feature values ​​are found to change in the same direction for three consecutive days, performance degradation is considered likely.

[0053] To identify environmental sensitivity characteristics of abnormal operating conditions through correlation analysis between environmental disturbances and heat pipe response, a correspondence table between environmental parameters and heat pipe operating parameters is first established. This table includes the normal operating parameter ranges corresponding to environmental temperature and wind speed ranges. When environmental conditions change, the heat pipe operating parameters are checked to see if they remain within the expected range. If they deviate from the expected range, the degree and duration of the deviation are recorded. The sensitivity of the heat pipe response under different environmental conditions is analyzed, and the ratio of the change in environmental parameters to the change in heat pipe operating parameters is calculated. A ratio exceeding three times the normal range is considered to indicate environmental sensitivity. Identifying these characteristics helps distinguish between environmental impacts and actual faults.

[0054] When establishing the mapping relationship between health status classification parameters and environmental disturbances, the intensity of environmental disturbances is first divided into four levels: 0-0.25 for slight disturbances, 0.25-0.5 for moderate disturbances, 0.5-0.75 for severe disturbances, and 0.75-1 for extreme disturbances. Each disturbance level corresponds to a set of classification parameter adjustment schemes. The adjustment schemes are obtained by analyzing the optimal classification results under different environmental conditions in historical data. The mapping relationship is stored in the form of a lookup table with four rows, each containing the recommended classification threshold and judgment rule for that disturbance level. When the environmental disturbance level changes, the corresponding parameter settings are automatically invoked.

[0055] When dynamically adjusting the decision boundary of the health status classification method based on environmental characteristics, three key parameters are primarily adjusted: the threshold for determining the duration of a state, the tolerance for fluctuations in health parameters, and the confirmation conditions for state transitions. Under conditions of strong environmental disturbances, the threshold for determining the duration of a state is appropriately extended, the tolerance for fluctuations in health parameters is increased, and the confirmation conditions for state transitions are strengthened. The specific adjustment range is determined experimentally. For example, under conditions of severe disturbances, the threshold for the duration of a state is increased by 20%, the tolerance for fluctuations is increased by 0.05, and state transitions require confirmation at three consecutive time points. These adjustments can avoid misjudgments caused by environmental fluctuations.

[0056] When tracking results show a continuous change in system status, the health status classification method is updated. The criteria for continuous change are: the health parameter maintains the same trend for 7 consecutive days, with daily changes exceeding 0.01; or the health status distribution characteristics deviate from the historical normal range for 14 consecutive days. When these conditions are met, the update procedure is automatically initiated. The update procedure first checks data quality to rule out the influence of factors such as sensor malfunction, then analyzes the change characteristics to determine whether to adjust the classification threshold or modify the judgment rules. The update process is carried out in steps, adjusting only one parameter at a time, observing the effect for 3 days before deciding whether to continue adjusting.

[0057] When gradually optimizing the classification threshold while maintaining the stability of the health status classification method, a small-step, incremental adjustment strategy is adopted. Each adjustment does not exceed 5% of the original value, and the time interval between two adjacent adjustments is no less than 7 days. During the optimization process, the adjustment history of each threshold is recorded, including the adjustment time, adjustment magnitude, and post-adjustment effect evaluation. Effect evaluation is performed by calculating the adjusted classification accuracy and false positive rate. If the adjusted accuracy improves and the false positive rate decreases, the adjustment is retained; if the effect deteriorates, the previous version is reverted. This conservative optimization strategy ensures that the stability of the classification method is not affected.

[0058] When validating the optimization effect using historical state sequences, data from two time periods were selected for comparative analysis: 30 days before optimization and 30 days after optimization. The comparison included the rationality of the distribution of healthy states, the smoothness of state transitions, and the accuracy of anomaly identification. Rationality was judged by checking whether the proportion of each state conformed to historical patterns; smoothness was evaluated by analyzing the frequency of state transitions; and accuracy was verified by comparing with manual inspection results. Simultaneously, the adaptability of the optimized method under different environmental conditions was examined to ensure good classification performance under various typical operating conditions. The validation period was 30 days, during which the changing trends of various indicators were closely monitored.

[0059] The working principle of this invention is as follows: Multi-source sensors deployed at key locations on the heat pipe synchronously collect time-series monitoring data such as temperature, pressure, and ambient wind speed to construct a thermodynamic state matrix with spatiotemporal correlation characteristics. This matrix is ​​then decomposed at multiple scales to extract characteristic modal components at different time scales, and nonlinear dynamic response features are generated through a feature recombination mechanism based on environmental parameter perception. A multi-dimensional diagnostic vector is constructed based on phase space reconstruction and trajectory analysis, and combined with real-time wind speed signals, a comprehensive health parameter is generated through feature-level fusion and inference calculation. According to a preset four-level health state classification standard, the health parameter is combined with duration requirements, and the health level determination result is output through a fusion of feature space distribution and time series analysis. During continuous monitoring, by tracking the correlation between environmental disturbance intensity and heat pipe state evolution, the classification threshold is dynamically adjusted, and the health state determination method is optimized online, forming a complete closed-loop diagnostic system from data acquisition to model self-optimization.

[0060] The foregoing has provided a detailed description of one embodiment of the present invention, but this description is merely a preferred embodiment and should not be construed as limiting the scope of the invention. All equivalent variations and modifications made within the scope of the claims of this invention should still fall within the patent coverage of this invention.

Claims

1. A remote monitoring and diagnosing system for working condition of a thermal probe in permafrost region based on edge computing, characterized in that, include: The data acquisition and state matrix construction module is used to acquire time-series monitoring data of the heat pipe under the monitoring of multiple source sensors, and construct a thermodynamic state matrix reflecting the internal thermodynamic state of the heat pipe based on the time-series monitoring data. The multi-scale feature extraction and fusion module is used to perform multi-scale tensor decomposition on the thermodynamic state matrix, extract spatiotemporal feature modes at different time scales, and generate nonlinear dynamic response features characterizing the working state of the heat pipe through an adaptive feature fusion algorithm. The health parameter inference and generation module constructs a multi-dimensional diagnostic vector based on nonlinear dynamic response characteristics, combines the real-time detection signal from the environmental wind speed sensor, and generates comprehensive health parameters through feature-level fusion and inference calculation. The health status classification and judgment module takes the comprehensive health parameters and inputs them into the preset health status classification model, and outputs the corresponding health level judgment result. The online adaptive optimization module is used to track the changes in multi-source time-series signals of the heat pipe's working status and the evolution trend of environmental disturbance intensity in real time during continuous monitoring, and to optimize the health status classification model online based on the tracking results.

2. The edge-computing-based remote monitoring and diagnosing system for working conditions of a thermal rod in permafrost regions according to claim 1, characterized in that, The process of constructing the thermodynamic state matrix is ​​as follows: The system acquires multi-point temperature sequences and ambient wind speed sequences for the evaporation and condensation sections of the heat pipe, and uses hardware synchronization pulse signals to ensure accurate alignment of timestamps for each channel's data. The collected multi-source time-series data are reorganized into three-dimensional data according to a preset time window to construct a three-dimensional data tensor with time axis, sensor spatial distribution axis and physical quantity type axis as dimensions; Tensor expansion and dimensionality reduction are performed on the three-dimensional data tensor. A thermodynamic state matrix with spatiotemporal correlation characteristics is generated by a feature extraction method that preserves the main feature patterns.

3. The remote monitoring and diagnostic system for the working status of heat pipes in permafrost regions based on edge computing as described in claim 1, characterized in that, The extraction process of the spatiotemporal feature modes is as follows: By setting multiple time windows of different lengths to perform sliding segmentation on the thermodynamic state matrix, three sets of time-scale data fragments at the second, minute, and hour levels are obtained. Spatiotemporal feature decoupling is performed on data segments at each time scale to separate and extract spatial sensor distribution features from temporal evolution features; The separated spatiotemporal features are input into a deep feature encoder to generate scale-specific spatiotemporal feature modes. Cross-scale correlation analysis was performed on spatiotemporal feature modes at different time scales to screen out core feature modes that showed responses at multiple scales, which were denoted as spatiotemporal feature modes.

4. The remote monitoring and diagnostic system for the working status of heat pipes in permafrost regions based on edge computing as described in claim 1, characterized in that, The nonlinear dynamic response characteristics that characterize the working state of the heat pipe specifically include: Based on real-time wind speed and temperature gradient, spatiotemporal feature modes are dynamically grouped and selected. Feature selection criteria are established through dual verification of feature reconstruction fidelity and diagnostic consistency. State evolution analysis was performed on the selected feature sequences to identify the transient response stage and steady-state operation stage during the operation of the heat pipe; The transient response features and steady-state operation features are dynamically coupled and encoded to generate a nonlinear dynamic response feature vector that simultaneously contains fast response characteristics and long-term operation characteristics.

5. The remote monitoring and diagnostic system for the working status of heat pipes in permafrost regions based on edge computing as described in claim 1, characterized in that, The construction of the multidimensional diagnostic vector specifically includes: Phase space reconstruction is performed based on nonlinear dynamic response characteristics, which expands the one-dimensional feature sequence into a multi-dimensional state space representation. Cluster analysis of the system dynamic trajectory in the reconstructed phase space is performed to identify motion patterns that characterize different operating states; Extract the trajectory curvature, motion velocity, and distribution density features of each motion mode to form a set of geometric features in the state space; By encoding the geometric feature set and the nonlinear dynamic response features in parallel, a multidimensional diagnostic vector with physical interpretability is generated.

6. The remote monitoring and diagnostic system for the working status of heat pipes in permafrost regions based on edge computing as described in claim 1, characterized in that, The specific process for generating the comprehensive health parameters is as follows: Dynamically filter and reorganize multidimensional diagnostic vectors using real-time wind speed signals; Time-series alignment and interactive analysis were performed between wind speed change patterns and diagnostic vector features; By coordinating and integrating diagnostic results from multiple time scales, the characteristic differences between transient disturbances and actual faults can be identified. An adaptive weight allocation based on information value assessment generates a comprehensive health parameter with confidence level labels.

7. The remote monitoring and diagnostic system for the working status of heat pipes in permafrost regions based on edge computing as described in claim 1, characterized in that, The construction process of the health status classification model is as follows: Based on the comprehensive health parameter sequence in historical operational data, a characteristic spatial distribution map of health status is constructed. In the feature space, the boundary regions of different health states are fuzzily divided to establish a health state classification boundary with a transition zone; Perform state persistence analysis on the dynamic change patterns of health parameters; The feature spatial distribution and time series analysis results are integrated to form the final health level determination result.

8. The remote monitoring and diagnostic system for the working status of heat pipes in permafrost regions based on edge computing as described in claim 7, characterized in that, The decision process of the health status classification model includes: Based on the preset four-level health status classification standard, the comprehensive health parameters are mapped to the corresponding health levels; When the overall health parameter is greater than the first preset threshold, it is determined to be in normal working condition; When the overall health parameter is between the second preset threshold and the first preset threshold and the duration exceeds 24 hours, it is determined to be a state of mild performance degradation. When the overall health parameter is between the third preset threshold and the second preset threshold and the duration exceeds 12 hours, it is determined to be a state of moderate performance degradation. When the overall health parameter is less than the third preset threshold and the duration exceeds 6 hours, it is judged as a serious fault state. The instantaneous state, short-term state, and continuous state are classified and confirmed separately. The final output is a health level determination result that has both time persistence and state stability.

9. The remote monitoring and diagnostic system for the working status of heat pipes in permafrost regions based on edge computing as described in claim 1, characterized in that, The real-time tracking of the multi-source time-series signal changes and the evolution trend of environmental disturbance intensity in the operating state of the heat pipe specifically includes: An evolution curve of environmental disturbance intensity is constructed. By continuously collecting wind speed and ambient temperature data and calculating their change gradient, the dynamic impact of the external environment is quantified. Simultaneously monitor the temporal variation patterns of evaporation section temperature, condensation section temperature, and operating pressure; Extract early feature patterns characterizing heat rod performance degradation from multi-source time-series signals; By analyzing the correlation between environmental disturbances and heat pipe response, the environmental sensitivity characteristics of abnormal operating conditions can be identified.

10. The remote monitoring and diagnostic system for the working status of heat pipes in permafrost regions based on edge computing according to claim 1, characterized in that, The online optimization of the health status classification model specifically includes: Establish a mapping relationship between the parameters of the health status classification model and environmental disturbances, and dynamically adjust the decision boundary of the health status classification model according to environmental characteristics; When the tracking results show that the system status is undergoing continuous changes, the health status classification model will be updated. While maintaining the qualitative nature of the health status classification model, the classification threshold is gradually optimized. The optimization effect was verified by using historical state sequences.

Images