A distributed edge computing system and method based on internet of things

Abstract

The application discloses a kind of based on Internet of Things distributed edge computing system and method, belong to edge computing and intelligent sensing technical field, by constructing node level three-domain coupling model, realize the collaborative modeling of original sampling sequence, clock signal and electromagnetic environment disturbance data;Further identify pseudo-synchronous interval and generate phase distortion indicator function, reconstruct the phase of data;On this basis, extract phase collapse entropy gradient parameter and cross-node phase hysteresis loop area parameter, construct topological mapping matrix to depict phase collapse propagation path;Task dynamic decomposition and path avoidance are realized in combination with the topological mapping matrix, and phase distortion compensation is introduced in asynchronous computing process;Finally, output high-confidence state evaluation results through reverse consistency check and fusion;The application can effectively inhibit the influence of phase collapse, improve the accuracy and stability of distributed edge computing.

Description

Technical Field

[0001] This invention relates to the fields of edge computing and intelligent sensing technology, specifically to a distributed edge computing system and method based on the Internet of Things. Background Technology

[0002] In IoT applications for micro-damage monitoring of bridges in high-altitude and cold mountainous areas, distributed edge nodes face not only clock nonlinear drift caused by low temperatures but also electromagnetic reflection anomalies caused by ice and snow cover. This leads to a "pseudo-synchronization" phenomenon in local nodes within a very short time window, where the timestamps appear consistent but the physical sampling phases are misaligned. This rare problem, coupled with micro-power jitter, creates an unobservable data phase collapse region at the edge, systematically masking fine-grained damage characteristics. Existing technologies mostly rely on single time calibration or simple redundancy mechanisms, making it difficult to identify and repair such hidden and locally occurring phase-level errors. Summary of the Invention

[0003] The purpose of this invention is to provide a distributed edge computing system and method based on the Internet of Things to address the shortcomings of the prior art.

[0004] To achieve the above objectives, the present invention provides the following technical solution: a distributed edge computing method based on the Internet of Things, comprising: S100 collects the original sampling sequences, clock signals and electromagnetic environmental disturbance data of each IoT node to construct a node-level three-domain coupling model C; S200 identifies the pseudo-synchronization intervals of each node based on C and generates the corresponding phase distortion indicator function Φ. It then performs phase reconstruction on the original sampling sequence to obtain the phase correction dataset D1. S300, using D1 to extract the phase collapse entropy gradient parameters and the phase hysteresis loop area parameters across nodes, constructs a topological mapping matrix T that reflects the local phase collapse propagation path; S400 inputs T and D1 into the edge-side adaptive splitting calculation model to generate a dynamic task decomposition strategy Q that avoids the phase collapse path; S500 performs asynchronous sharding computation at each edge node based on Q, and introduces a computation process compensation based on Φ to obtain the compensation result set D2; S600 performs reverse consistency verification and fusion on D2 based on topology mapping T, and outputs a highly reliable state assessment result that eliminates the effects of phase collapse.

[0005] Preferably, the construction of the node-level three-domain coupling model includes the following steps: S110, decompose the original sampling sequence of each IoT node into a time window, extract the signal phase features and amplitude features at different time scales, and form an initial signal feature set; S120, perform frequency stability analysis and nonlinear drift fitting on the clock signal of each IoT node to obtain the corresponding clock drift function, and perform time alignment processing on the clock drift function and the initial signal feature set to obtain the clock correction feature set; S130, Perform spectral decomposition and spatial correlation modeling on electromagnetic environment disturbance data, extract electromagnetic disturbance intensity distribution features, and map them onto the time axis corresponding to the clock correction feature set to form a disturbance correlation feature set; S140, the initial signal feature set, clock correction feature set and disturbance correlation feature set are coupled and fused to construct a node-level three-domain coupling model that uniformly represents the correlation between the time domain, phase domain and energy domain.

[0006] Preferably, the step of identifying the pseudo-synchronization intervals of each node based on C includes the following steps: S210, based on a node-level three-domain coupling model, extracts the degree of coupling deviation between the time domain, phase domain and energy domain, constructs a coupling consistency judgment index, and calculates the continuous change sequence of the coupling consistency judgment index through a sliding time window; S220, based on the abrupt change magnitude and duration of the coupling consistency judgment index, and combined with the preset judgment threshold, identify pseudo-synchronization intervals, and perform boundary marking on each pseudo-synchronization interval to obtain a set of pseudo-synchronization intervals.

[0007] Preferably, obtaining the phase correction dataset D1 includes: S230, constructing a phase distortion indicator function based on the phase evolution trajectory within the pseudo-synchronization interval set and the phase reference relationship with the normal interval, and mapping the phase distortion indicator function to the corresponding time axis; S240, performing point-by-point compensation and continuity reconstruction of the phase components in the original sampling sequence according to the phase distortion indicator function to obtain the phase correction dataset D1.

[0008] Preferably, the method for obtaining the phase collapse entropy gradient parameter is as follows: S310, Delayed embedding is performed on the phase sequence in the phase correction dataset to construct a high-dimensional phase space trajectory, and the nearest neighbor search method is used to determine the local neighborhood point set based on the high-dimensional phase space trajectory; S320, construct the Riemannian manifold structure based on the geodesic distance between the local neighborhood point set, and estimate the probability density of the phase distribution in the local neighborhood based on the geodesic distance to obtain the manifold probability distribution function; S330, calculate the corresponding manifold information entropy based on the manifold probability distribution function, and perform a difference operation on the manifold information entropy in the spatial neighborhood direction to obtain the local information entropy change rate; S340, the local information entropy change rate is spatially weighted and accumulated to generate phase collapse entropy gradient parameters.

[0009] Preferably, the method for obtaining the cross-node phase hysteresis loop area parameter is as follows: S350: Select the phase sequences of any two IoT nodes from the phase correction dataset, construct a two-dimensional phase trajectory point set according to a unified reference time axis, and form phase evolution point cloud data based on the two-dimensional phase trajectory point set; S360, construct a simple complex structure for the phase evolution point cloud data according to the distance threshold increment method, and generate corresponding topological connection relationships at different distance scales; S370, calculate one-dimensional homology features based on the simple complex structure, extract closed loop structures that persist during scale changes, and record their generation and extinction scales; S380, calculate the persistence metric based on the generation and destruction scales of the closed loop structure, and map the persistence metric to the corresponding cross-node phase hysteresis loop area parameter.

[0010] Preferably, the generation of the dynamic task decomposition strategy Q that avoids phase collapse paths includes the following steps: S410: Based on the topology mapping matrix, the phase collapse propagation path strength between each IoT node is extracted, and combined with the data stability characteristics of each node in the phase correction dataset, a node risk weight sequence is constructed. S420, the phase correction dataset is spatially partitioned according to the node risk weight sequence to generate an initial task partition set, and the data regions corresponding to high-risk nodes are isolated and marked. S430, Based on the propagation path constraint relationship between the initial task partition set and the topology mapping matrix, a task splitting criterion is constructed, and the initial task partition is recursively decomposed to form a multi-level task subset; S440 reorganizes and schedules the task subsets according to the node risk weights and data temporal continuity corresponding to each task subset, generating a dynamic task decomposition strategy that avoids the phase collapse propagation path.

[0011] Preferably, asynchronous sharding computation is performed on each edge node according to Q, including the following steps: S510 divides the phase correction dataset into multiple time-continuous and spatially isolated task slices according to the dynamic task decomposition strategy, and assigns each task slice to the corresponding edge node to perform asynchronous computation. S520 extracts the phase distortion indicator function value at the corresponding time point in real time during the calculation process at each edge node, constructs the phase distortion compensation coefficient, and embeds the compensation coefficient into the feature weight update stage of the calculation process. S530 corrects the intermediate calculation results of each task segment on a time-by-time basis according to the phase distortion compensation coefficient to obtain the compensated local calculation results. S540 aligns and stitches the compensated local calculation results output by each edge node according to a unified reference time axis to form a compensation result set.

[0012] Preferably, the reverse consistency verification and fusion of D2 based on topology mapping T includes the following steps: S610: Based on the topology mapping matrix, the back propagation path between each IoT node in the compensation result set is extracted, the back consistency constraint relationship between nodes is constructed, and the back consistency weight corresponding to each node is calculated. S620, based on the reverse consistency weight, performs consistency deviation detection on the results of each node in the compensation result set, obtains the deviation value sequence at each time point, and compares it with the preset consistency threshold; S630: For node results whose deviation values ​​exceed the consistency threshold, backtrack and correct them step by step according to the backpropagation path to generate a corrected node result sequence. S640, the corrected node result sequence is weighted and fused according to a unified reference time axis to obtain a highly reliable state assessment result that eliminates the effect of phase collapse.

[0013] The present invention also provides a distributed edge computing system based on the Internet of Things, comprising: Three-domain coupling modeling module: Collects the original sampling sequence, clock signal and electromagnetic environment disturbance data of each IoT node, and constructs a node-level three-domain coupling model C; Phase reconstruction module: Based on C, identify the pseudo-synchronization interval of each node and generate the corresponding phase distortion indicator function Φ. Perform phase reconstruction on the original sampling sequence to obtain the phase correction dataset D1. Topology modeling module: Use D1 to extract the phase collapse entropy gradient parameters and phase hysteresis loop area parameters across nodes, and construct a topology mapping matrix T that reflects the local phase collapse propagation path; Adaptive Task Decomposition Module: Input T and D1 into the edge-side adaptive splitting calculation model to generate a dynamic task decomposition strategy Q that avoids the phase collapse path; Compensation module: Performs asynchronous sharding calculations at each edge node based on Q, and introduces a calculation process based on Φ for compensation, to obtain the compensation result set D2; Fusion output module: Performs reverse consistency verification and fusion on D2 based on topology mapping T, and outputs a highly reliable state assessment result that eliminates the effects of phase collapse.

[0014] The technical effects and advantages provided by the present invention in the above technical solution are as follows: 1. This invention constructs a node-level three-domain coupling model to uniformly model the original sampling sequence, clock signal, and electromagnetic environment disturbance data, achieving deep correlation analysis between the time domain, phase domain, and energy domain. Based on this, it introduces methods for identifying pseudo-synchronization intervals and generating phase distortion indicator functions, enabling the identification of hidden anomalies such as "pseudo-synchronization but actual distortion" that are difficult to detect by traditional methods under extreme environments. Furthermore, it effectively eliminates the phase misalignment problem caused by the superposition of nonlinear clock drift and electromagnetic disturbances through phase reconstruction, improving data quality from the source. Simultaneously, by introducing the phase collapse entropy gradient parameter based on Riemannian manifold information entropy and the cross-node phase hysteresis loop area parameter based on topological persistent cohomology, a topological mapping matrix is ​​constructed, which can characterize the propagation path and structural features of phase collapse in space. Compared with existing technologies that rely on simple correlation or time delay analysis, it has stronger structural expression capabilities and anomaly identification accuracy.

[0015] 2. This invention utilizes an edge-side adaptive splitting computation model, which synergistically integrates the topology mapping matrix and the phase correction dataset to achieve dynamic task decomposition and path avoidance. This allows the computation process to proactively avoid phase collapse propagation paths. Furthermore, a full-process compensation mechanism based on a phase distortion indicator function is introduced during asynchronous slice computation to effectively suppress error propagation in the computation chain. Simultaneously, a reverse consistency check and step-by-step backtracking correction method based on the topology mapping matrix are used to globally constrain and fuse the compensation results, ultimately outputting a highly reliable state assessment result. The overall technical solution not only improves computational stability and result reliability in complex environments but also achieves highly sensitive identification of minor local anomalies. Attached Figure Description

[0016] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments recorded in this invention. For those skilled in the art, other drawings can be obtained based on these drawings.

[0017] Figure 1 This is a flowchart of the method of the present invention.

[0018] Figure 2 This is a flowchart of the system modules of the present invention. Detailed Implementation

[0019] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, 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, 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.

[0020] Example 1, please refer to Figure 1 As shown in this embodiment, a distributed edge computing method based on the Internet of Things includes: S100 collects the original sampling sequences, clock signals, and electromagnetic environmental disturbance data of each IoT node to construct a node-level three-domain coupling model C.

[0021] In this embodiment, the construction of the node-level three-domain coupling model includes steps S110 to S140.

[0022] For step S110, the original sampling sequences of each IoT node are decomposed into multi-scale time windows to extract signal phase and amplitude features at different time scales, forming an initial signal feature set. Specifically, each IoT node is deployed at different monitoring locations on a bridge in a high-altitude, cold mountainous area to continuously collect the original sampling sequences corresponding to vibration, strain, or displacement. To avoid smoothing out local anomalies by directly analyzing based on a single sampling scale, the original sampling sequences are first preprocessed. The preprocessing includes removing missing sampling points, using cubic spline interpolation to fill in discontinuous sampling segments, and using sliding median filtering to eliminate isolated impulse noise. The sliding median filtering window length is set to five consecutive sampling points. After preprocessing, the original sampling sequences are decomposed into time windows according to multiple time scales. The multiple time scales are preferably short-time windows, medium-time windows, and long-time windows. The short-time window length is set to 0.5 seconds, the medium-time window length is set to 2 seconds, and the long-time window length is set to 8 seconds. Adjacent time windows have a 50% overlap rate to ensure that both short-term abrupt changes and long-term evolution trends are preserved.

[0023] Within each time window, the original sampling sequence is first bandpass filtered. The bandpass range is determined based on the main response frequency band of the bridge structure, preferably set to 0.5 Hz to 50 Hz, to remove low-frequency drift and high-frequency random noise. Then, a Hilbert transform is used to construct an analytic signal, and the instantaneous phase and instantaneous amplitude are calculated based on the analytic signal. The instantaneous phase is defined as the arctangent of the imaginary and real parts of the analytic signal, and the instantaneous amplitude is defined as the magnitude of the analytic signal. To ensure comparability of phase information, the instantaneous phase is expanded to eliminate the double-circumference discontinuity caused by phase jumps. Subsequently, within each time window, the phase mean, phase variance, phase change rate, amplitude mean, amplitude variance, and amplitude kurtosis are calculated. The phase change rate is defined as the ratio of the instantaneous phase difference between adjacent sampling times to the sampling time interval, used to describe the phase evolution speed. The amplitude kurtosis is used to characterize the local impact response characteristics. Finally, all phase and amplitude features extracted from the short, medium, and long time windows are concatenated using a unified time index to form the initial signal feature set corresponding to each IoT node. This initial signal feature set preserves both micro-scale disturbances and evolution trends over longer time scales, providing a foundation for subsequent clock correction.

[0024] For step S120, frequency stability analysis and nonlinear drift fitting are performed on the clock signals of each IoT node to obtain the corresponding clock drift function. The clock drift function is then time-aligned with the initial signal feature set to obtain a clock correction feature set. Specifically, each IoT node records its local clock signal while acquiring the original sampling sequence. The clock signal includes continuous timing pulses and a local timestamp corresponding to each sampling point. Due to the influence of temperature changes, power supply fluctuations, and device aging on the crystal oscillator frequency in cold environments, the local clock will experience nonlinear drift. Relying solely on fixed offset compensation cannot accurately reflect the time error. Therefore, the error sequence between the local clock signal and the reference time signal is first extracted according to a preset calibration period. The reference time signal can be a unified timing signal from the edge side, and the calibration period is preferably set to 10 seconds. The change in time error within adjacent calibration periods is divided by the corresponding time interval to obtain the clock frequency offset sequence. Then, a frequency stability index is calculated based on the clock frequency offset sequence. The frequency stability index is defined as the standard deviation of the frequency offset sequence within three consecutive calibration periods, used to reflect the degree of short-term clock jitter.

[0025] To describe the nonlinear changes in clock drift, the time error sequence is piecewise fitted. Specifically, a third-order polynomial fitting method is used to construct the clock drift function. Within each calibration period, the clock drift function is represented as a cubic polynomial of the reference time, where the polynomial coefficients are solved using the least squares method. If the fitting residual exceeds a preset residual threshold within a certain calibration period, the calibration period is further split into two sub-periods, and refitted for each. The residual threshold is defined as 1.5 times the root mean square value of the time error sequence within that calibration period, to avoid situations where a single polynomial cannot describe abrupt drifts. After obtaining the clock drift function corresponding to each calibration period, time inversion correction is performed on each phase and amplitude feature according to the time index in the initial signal feature set, that is, mapping the local time to the corrected time under the reference time axis. Subsequently, the feature values ​​on the corrected time axis are resampled using linear interpolation, so that the feature data of different IoT nodes correspond to a unified reference time scale, finally obtaining the clock correction feature set. The clock correction feature set is essentially a unified time base feature set formed after eliminating the influence of nonlinear clock drift, which can be used to subsequently introduce the influence of electromagnetic environmental disturbances.

[0026] For step S130, spectral decomposition and spatial correlation modeling are performed on the electromagnetic environment disturbance data to extract the electromagnetic disturbance intensity distribution features, which are then mapped onto the time axis corresponding to the clock correction feature set to form a disturbance correlation feature set. Specifically, the electromagnetic environment disturbance data is synchronously collected by electromagnetic sensing devices deployed near each IoT node, and the collected content includes the sequence values ​​of electromagnetic field intensity changing over time. Since ice and snow cover, metal component reflection, and local ice shedding can change the electromagnetic propagation path, thereby inducing pseudo-synchronization phenomena, the electromagnetic environment disturbance data needs to be analyzed in detail. First, the electromagnetic environment disturbance data corresponding to each IoT node is segmented, with the segment length consistent with the mid-time window in step S110, both set to 2 seconds, to facilitate subsequent time sequence correspondence. A fast Fourier transform is performed on each segment of electromagnetic environment disturbance data to obtain the frequency domain energy distribution. The proportions of low-frequency, mid-frequency, and high-frequency disturbance energy are calculated based on the frequency domain energy distribution. The low-frequency range is set to 0 Hz to 5 Hz, the mid-frequency range to 5 Hz to 30 Hz, and the high-frequency range above 30 Hz. The electromagnetic disturbance intensity distribution characteristics include total energy, dominant frequency, spectral centroid, and frequency band energy proportion. The spectral centroid is defined as the sum of the products of each frequency component and its corresponding amplitude, divided by the sum of all amplitudes, and is used to characterize the location of concentrated electromagnetic disturbance energy.

[0027] After obtaining the electromagnetic disturbance intensity distribution characteristics of each node, spatial correlation is further constructed. Specifically, the spatial distance between any two IoT nodes is calculated based on their actual installation locations on the bridge. A spatial correlation function is then established using the spatial distance as the independent variable and the correlation between the electromagnetic disturbance intensity distribution characteristics as the dependent variable. The correlation is calculated using the Pearson correlation coefficient, and the spatial correlation function uses a Gaussian decay form, meaning the closer the spatial distance, the greater the correlation. This method yields a weighted result of the electromagnetic disturbance propagation influence of neighboring nodes on each IoT node at the current moment. Then, the electromagnetic disturbance intensity distribution characteristics and their spatial correlation results are mapped according to the unified reference time axis established in step S120. During mapping, for each correction time point, the electromagnetic disturbance segment feature value with the smallest time difference is selected as the base value, and then a weighted average compensation is applied to the adjacent two segment feature values, with the weight determined by the reciprocal of the time difference. After mapping, a disturbance correlation feature set corresponding to each time point of the clock correction feature set is obtained. The disturbance correlation feature set reflects the state of an IoT node under the combined influence of local and neighboring electromagnetic environment disturbances at a unified time scale.

[0028] For step S140, the initial signal feature set, clock correction feature set, and disturbance correlation feature set are coupled and fused to construct a node-level three-domain coupling model that uniformly represents the correlation between the time domain, phase domain, and energy domain. Specifically, in order for the node-level three-domain coupling model to simultaneously represent the evolution of the original signal, the clock drift correction result, and the impact of electromagnetic environment disturbances, the initial signal feature set, clock correction feature set, and disturbance correlation feature set are first normalized. The normalization adopts the maximum-minimum value normalization method, mapping each feature value to the interval between zero and one to eliminate dimensional differences. Subsequently, using each time point on the unified reference time axis as the basic modeling unit, a coupling vector between the three types of features is constructed. The coupling vector includes at least the phase mean, phase change rate, amplitude mean, amplitude kurtosis corresponding to the original sampling sequence, the drift amount and drift change rate of the clock drift function at that time point, and the total energy, spectral centroid, and spatial correlation weighted value in the electromagnetic disturbance intensity distribution characteristics.

[0029] Based on this, a node-level three-domain coupling model is constructed using a constrained weighted coupling modeling approach. Specifically, the mutual information value between any two features is first calculated, representing the degree of nonlinear correlation between them. This mutual information value is then normalized and used as candidate coupling weights. To avoid excessive noise introduced by weakly correlated features, a coupling screening threshold is set, retaining only feature connections corresponding to mutual information values ​​greater than 0.35 after normalization. The coupling screening threshold of 0.35 is determined by testing six thresholds (0.20, 0.25, 0.30, 0.35, 0.40, and 0.45) on historical monitoring data, comparing the anomaly identification accuracy with the false alarm rate, and selecting the threshold that significantly improves the identification accuracy without significantly increasing the false alarm rate as the final value. For the retained feature connections, the correlation strength between the three types of features is represented in a weighted adjacency form, thus forming the basic structure of the node-level three-domain coupling model.

[0030] Furthermore, to reflect the dynamic influence between the time domain, phase domain, and energy domain, a time-recursive update process is introduced on the aforementioned basic structure. For the node-level three-domain coupling model C at the current time point, not only is the coupling vector at the current time point used, but the coupling vectors from the previous two time points are also introduced for weighted updates, where the weight of the current time point is set to 0.6, the weight of the previous time point is set to 0.3, and the weight of the previous two time points is set to 0.1. This recursive method preserves short-term memory characteristics, enabling the model to respond to sudden disturbances without experiencing drastic fluctuations due to single-point anomalies. The final node-level three-domain coupling model is a dynamic correlation representation result indexed by a unified reference time axis and with feature coupling strength as its core content. This model can specifically reflect the temporal coupling state of a certain IoT node under the combined effects of original sampling changes, clock nonlinear drift, and electromagnetic environmental disturbances, and provides an input basis for subsequent pseudo-synchronization interval identification, phase distortion indicator function construction, and phase reconstruction processing.

[0031] S200 identifies the pseudo-synchronization intervals of each node based on C and generates the corresponding phase distortion indicator function Φ. It then performs phase reconstruction on the original sampling sequence to obtain the phase correction dataset D1.

[0032] In this embodiment, a cable-stayed bridge in a high-altitude, cold mountainous area is used as the application object. A total of 24 IoT nodes are deployed in the main beam and cable area of ​​the bridge deck. Each IoT node is equipped with a vibration sensor and an electromagnetic environment sensing device. The sampling frequency is set to 200 Hz to continuously collect bridge structural response data and electromagnetic environment disturbance data. Based on the node-level three-domain coupling model constructed in the aforementioned steps, steps S210 to S240 are further executed to achieve pseudo-synchronization interval identification and phase correction processing.

[0033] First, step S210 is executed. Based on the node-level three-domain coupling model, the degree of coupling deviation between the time domain, phase domain, and energy domain is extracted to construct a coupling consistency judgment index. The continuous change sequence of this index is then calculated using a sliding time window. Specifically, for each IoT node, calculations are performed step-by-step on a unified reference time axis using a sliding time window of 2 seconds, with a sliding step size of 0.5 seconds. Within each sliding time window, the coupling vector in the node-level three-domain coupling model is extracted. This coupling vector includes the phase mean, phase change rate, amplitude mean, clock drift, and the total energy and spectral centroid from the electromagnetic disturbance intensity distribution characteristics. Subsequently, the Euclidean distance between this coupling vector in the current time window and the previous time window is calculated to characterize the degree of change in the coupling relationship. Simultaneously, to enhance sensitivity to abnormal changes, the Euclidean distance is normalized by dividing the Euclidean distance in the current time window by the average of the Euclidean distances of the past 10 time windows. Finally, the normalized Euclidean distance is defined as the coupling consistency judgment index. This index is calculated over all sliding time windows to form a continuous sequence of coupling consistency judgment indices.

[0034] Next, step S220 is executed. Based on the abrupt change magnitude and duration of the coupling consistency judgment index, and combined with a preset judgment threshold, pseudo-synchronization intervals are identified, and the boundaries of each pseudo-synchronization interval are defined to obtain a set of pseudo-synchronization intervals. Specifically, the judgment threshold for the coupling consistency judgment index is first set to 1.8. This threshold is obtained through statistical analysis of historical monitoring data. That is, under normal operating conditions, the mean of the coupling consistency judgment index is approximately 1.0, and the standard deviation is approximately 0.3. Therefore, the mean plus twice the standard deviation is selected as the judgment threshold. When the coupling consistency judgment index exceeds the judgment threshold for two or more consecutive sliding time windows, this time period is initially marked as an abnormal interval. Furthermore, to avoid misjudgment caused by short-term noise, a minimum duration constraint is set, that is, the duration of the abnormal interval must be greater than 1 second. Abnormal intervals that meet the above conditions are identified as pseudo-synchronization intervals. Subsequently, the boundaries of each pseudo-synchronization interval are finely defined. Specifically, the method is as follows: backtracking from the initially identified interval start position until the coupling consistency judgment index is below 1.2 as the starting boundary; extending backward from the interval end position until the index is below 1.2 as the ending boundary, thereby obtaining a more accurate range of pseudo-synchronization intervals. Finally, a set of pseudo-synchronization intervals is formed.

[0035] Then, step S230 is executed. Based on the phase evolution trajectory within the pseudo-synchronization interval set and the phase reference relationship with the normal interval, a phase distortion indicator function is constructed, and the phase distortion indicator function is mapped to the corresponding time axis. Specifically, in each IoT node, a normal interval 5 seconds before and after the pseudo-synchronization interval is selected as a reference interval. Within the reference interval, the trend curve of the phase mean changing with time is calculated, and the phase reference function is obtained by fitting a quadratic polynomial. Subsequently, the actual phase evolution trajectory is extracted within the pseudo-synchronization interval, and the deviation value between the trajectory and the phase reference function is calculated. The deviation value is defined as the difference between the actual phase and the reference phase. To avoid the influence of local fluctuations, a moving average method is used to smooth the deviation value, and the sliding window length is set to 0.2 seconds. Afterward, the smoothed deviation value is normalized to obtain a function with a value range between -1 and +1, which is the phase distortion indicator function. The larger the absolute value of the phase distortion indicator function, the higher the degree of phase distortion; the sign indicates the direction of phase offset. Finally, the phase distortion indicator function is mapped according to a unified reference time axis so that it corresponds point-by-point with the original sampling sequence.

[0036] Finally, step S240 is executed, where the phase components in the original sampling sequence are compensated point-by-point and their continuity reconstructed based on the phase distortion indicator function to obtain a phase-corrected dataset. Specifically, the original sampling sequence is first re-performed with a Hilbert transform to extract the instantaneous phase. For each sampling point within the pseudo-synchronization interval, the instantaneous phase is compensated according to the phase distortion indicator function value at the corresponding time point. The compensation method is as follows: the phase distortion indicator function value is multiplied by the maximum phase compensation amplitude, and then this compensation value is subtracted from the original instantaneous phase. The maximum phase compensation amplitude is determined by the phase fluctuation range within the reference interval, specifically 1.5 times the standard deviation of the reference interval phase. After point-by-point compensation, to ensure the continuity of the phase sequence, the compensated phase sequence is subjected to phase expansion processing, and cubic spline interpolation is used to smooth the transition between adjacent sampling points. Subsequently, the analytical signal is reconstructed based on the compensated phase and the original amplitude, and the corresponding time-domain signal is extracted as the phase-corrected sampling result. The above process is repeated for all IoT nodes to obtain a phase-corrected dataset under a unified time reference.

[0037] Through the above steps, this embodiment can effectively identify pseudo-synchronization intervals and perform phase reconstruction processing under complex electromagnetic environments and power supply fluctuations in high-altitude and cold mountainous areas. Verified by actual monitoring data, during a sudden drop in temperature accompanied by strong winds and snow, a pseudo-synchronization phenomenon lasting approximately 3 seconds occurred in a cable area of ​​the bridge. After processing with this method, the phase deviation was reduced from approximately 0.8 radians to within 0.15 radians, significantly improving the accuracy of subsequent structural anomaly identification.

[0038] S300 uses D1 to extract the phase collapse entropy gradient parameters and the phase hysteresis loop area parameters across nodes, and constructs a topological mapping matrix T that reflects the local phase collapse propagation path.

[0039] In this embodiment, step S310 is first executed to perform delayed embedding processing on the phase sequence in the phase correction dataset to construct a high-dimensional phase space trajectory. Based on this high-dimensional phase space trajectory, a nearest neighbor search method is used to determine the local neighborhood point set. Specifically, for the phase sequence of each IoT node, it is assumed to be a discrete sequence arranged along a unified reference time axis. An embedding dimension of 5 and a time delay of 3 sampling intervals are selected. The one-dimensional phase sequence is converted into a five-dimensional vector sequence using the delayed embedding method, where each vector consists of the current phase value and its four subsequent delayed phase values, thus forming a high-dimensional phase space trajectory. Subsequently, in the high-dimensional phase space trajectory, a nearest neighbor search method based on Euclidean distance is used for each vector point to determine its local neighborhood point set. Specifically, the number of nearest neighbors is set to 10. For each vector point, the Euclidean distance between it and all other vector points is calculated, and the 10 points with the smallest distance are selected as the local neighborhood point set of that vector point, thus providing a foundation for subsequent manifold construction.

[0040] Next, step S320 is executed. A Riemannian manifold structure is constructed based on the geodesic distances between local neighborhood points, and the probability density of the phase distribution within the local neighborhood is estimated based on the geodesic distances to obtain the manifold probability distribution function. Specifically, firstly, an adjacency graph is constructed based on the local neighborhood points obtained in step S310, where each vector point is a node in the graph. If two nodes are neighbors, a connecting edge is established between them, and the weight of the edge is defined as the corresponding Euclidean distance. Subsequently, the shortest path algorithm is used to calculate the shortest path distance between any two neighborhood points on this adjacency graph, and this shortest path distance is used as the geodesic distance, thus obtaining an approximate expression of the local Riemannian manifold structure. Based on the geodesic distances, probability density is estimated for each local neighborhood point set, specifically using a Gaussian kernel function, where the kernel function bandwidth is taken as the average value of the geodesic distances within the neighborhood. For any neighborhood point, its probability density value is defined as the sum and normalization of the geodesic distances between that point and other points in the neighborhood calculated using the Gaussian kernel function. Through the above process, the probability distribution function of the manifold corresponding to each local neighborhood is obtained.

[0041] Then, step S330 is executed, calculating the corresponding manifold information entropy based on the manifold probability distribution function, and performing a difference operation on the manifold information entropy in the spatial neighborhood direction to obtain the local information entropy change rate. Specifically, for each local neighborhood point set, the information entropy is calculated according to its manifold probability distribution function. The information entropy is defined as the negative sum of the products of all probability density values ​​and their logarithms. To characterize the spatial change trend, for each vector point, its corresponding information entropy is calculated by difference with the information entropy of its nearest neighbors, and the average of the difference results is taken as the local information entropy change rate of that point. This change rate is used to reflect the change in the complexity of the phase structure in the manifold space. When the local information entropy decreases rapidly, it indicates that the phase structure is collapsing.

[0042] Then, step S340 is executed to perform spatial weighted accumulation processing on the local information entropy change rate to generate the phase collapse entropy gradient parameter. Specifically, for each vector point, its local information entropy change rate is weighted and summed using the reciprocal of the geodesic distance between it and all points in its neighborhood, with closer distances having larger weights. The weighted result is then normalized to obtain the phase collapse entropy gradient parameter corresponding to that point. The normalization method is to subtract the minimum value from the weighted result of all vector points and divide by the difference between the maximum and minimum values, so that the result is distributed between zero and one. This parameter is used to characterize the spatial gradient strength of phase collapse.

[0043] Next, step S350 is executed, selecting the phase sequences of any two IoT nodes from the phase correction dataset, constructing a two-dimensional phase trajectory point set according to a unified reference time axis, and forming phase evolution point cloud data based on the two-dimensional phase trajectory point set. Specifically, the phase values ​​of two IoT nodes at the same time point are selected as the horizontal and vertical coordinates, respectively, and the coordinates corresponding to all time points are combined to form a two-dimensional trajectory. This trajectory is then discretized into point cloud data, with each point representing the cross-node phase relationship at a time sampling moment.

[0044] Then, step S360 is executed, constructing a simplex structure on the phase evolution point cloud data using an incremental distance threshold, and generating corresponding topological connections at different distance scales. Specifically, the initial distance threshold is set to 0.01 radians, and the distance threshold is gradually increased in steps of 0.01 radians. At each distance threshold, for any two points, if their Euclidean distance is less than the current threshold, a connection edge is established; if all three points satisfy the connection conditions pairwise, a two-dimensional simplex is constructed. By continuously increasing the distance threshold, a series of nested simplex structures are gradually formed.

[0045] Then, step S370 is executed to calculate one-dimensional homology features based on the simple complex structure, extract closed loop structures that persist during scale changes, and record their generation and disappearance scales. Specifically, closed loop structures are identified at each distance threshold. When a closed loop first appears, that threshold is recorded as the generation scale; when the closed loop disappears at a larger threshold, that threshold is recorded as the disappearance scale. Only closed loop structures with a persistent scale difference greater than 0.02 radians are retained to exclude transient loops caused by noise.

[0046] Finally, step S380 is executed, calculating the persistence metric based on the generation and destruction scales of the closed-loop structure, and mapping the persistence metric to the corresponding cross-node phase hysteresis loop area parameter. Specifically, for each closed-loop structure, its persistence metric is defined as the destruction scale minus the generation scale. The persistence metrics of all closed-loop structures are weighted and summed, where the weights are determined by the number of points covered by the closed loop in the point cloud. The final result is then normalized to obtain the cross-node phase hysteresis loop area parameter, which characterizes the intensity of the cross-node phase hysteresis effect.

[0047] In this embodiment, after obtaining the phase collapse entropy gradient parameters corresponding to each IoT node and the cross-node phase hysteresis loop area parameters between any node pairs, a topology mapping matrix is ​​further constructed to characterize the local phase collapse propagation path.

[0048] Specifically, all IoT nodes are first numbered, with a total number of nodes of N. The topology mapping matrix is ​​then an N×N two-dimensional matrix, where each element represents the phase collapse propagation correlation strength between the corresponding nodes. In the implementation process, the phase collapse entropy gradient parameter of each IoT node is first standardized. The standardization process uses a max-min normalization method to map the phase collapse entropy gradient parameters of all nodes to between zero and one, thereby eliminating the dimensional differences between different nodes. The standardized phase collapse entropy gradient parameter is used to characterize the intensity level of phase structure collapse at that node. Subsequently, for any two IoT nodes, the cross-node phase hysteresis loop area parameter obtained in step S380 is used as the coupling characteristic quantity between the node pair. To improve numerical stability, the cross-node phase hysteresis loop area parameter is also normalized, using the same normalization method as the phase collapse entropy gradient parameter. After completing the above processing, the element values ​​in the topology mapping matrix are constructed. The element in the i-th row and j-th column of the matrix is ​​calculated as follows: First, the phase collapse entropy gradient parameter of the i-th IoT node is taken as the source node weight. Then, the cross-node phase hysteresis loop area parameter between the i-th and j-th nodes is taken as the coupling weight. The two are multiplied to obtain the initial correlation value. This initial correlation value is used to represent the potential strength of the phase collapse effect propagating from the i-th node to the j-th node.

[0049] To avoid interference from weak correlation noise on the topology, a correlation screening threshold is introduced. This threshold is obtained through historical data statistics and is specifically calculated as the mean of all nodes' initial correlation values ​​plus one standard deviation. When an initial correlation value is lower than this threshold, the corresponding matrix element is set to 0; when it is higher, the value is retained as an effective connection weight.

[0050] Furthermore, to reflect the directionality of propagation, the topology mapping matrix is ​​subjected to asymmetric constraints. Specifically, the matrix elements are retained only when the phase collapse entropy gradient parameter of the source node is greater than that of the target node; otherwise, they are set to 0. This processing enables the topology mapping matrix to reflect the propagation trend of phase collapse from high-intensity nodes to low-intensity nodes.

[0051] After obtaining the initial topological mapping matrix, the matrix is ​​row-normalized so that the sum of the elements in each row is 1, thereby transforming the matrix into a probabilistic propagation weight distribution. The normalization method is to divide each element of each row by the sum of all elements in that row; if the sum of the elements in a row is 0, then that row is kept entirely zero.

[0052] The resulting topology mapping matrix describes the phase collapse propagation paths and intensity distributions among IoT nodes at the current moment. Non-zero elements in the matrix correspond to potential propagation paths, and element values ​​represent propagation intensity. This topology mapping matrix not only preserves the local collapse characteristics of nodes but also incorporates cross-node hysteresis relationships, thus accurately characterizing the spatial propagation structure of local phase collapse and providing fundamental data support for subsequent task scheduling and anomaly assessment.

[0053] S400 inputs T and D1 into the edge-side adaptive splitting computation model to generate a dynamic task decomposition strategy Q that avoids the phase collapse path.

[0054] In this embodiment, step S410 is first executed to extract the phase collapse propagation path strength between each IoT node based on the topology mapping matrix, and to construct a node risk weight sequence by combining the data stability characteristics of each node in the phase correction dataset. Specifically, for each row in the topology mapping matrix, the sum of all non-zero elements in that row is calculated and defined as the propagation output strength of the corresponding IoT node; simultaneously, for each column, the sum of all non-zero elements in that column is calculated and defined as the propagation input strength of the corresponding IoT node. Subsequently, the propagation output strength and the propagation input strength are weighted and summed to obtain the propagation impact value of the node, where the weight of the propagation output strength is set to 0.6 and the weight of the propagation input strength is set to 0.4. Based on this, the phase fluctuation standard deviation and the mean of the phase change rate of each node are extracted from the phase correction dataset, and the average of the two is normalized and used as a data stability index. Finally, the propagation impact value and the data stability index are multiplied and the result is subjected to max-min normalization to obtain the node risk weight sequence. The larger the value, the more likely the node is to be in the phase collapse propagation path.

[0055] Next, step S420 is executed, where the phase correction dataset is spatially partitioned according to the node risk weight sequence to generate an initial task partition set, and the data regions corresponding to high-risk nodes are isolated and marked. Specifically, a risk threshold is first set as the mean of the node risk weight sequence plus 0.5 standard deviations; nodes exceeding this threshold are defined as high-risk nodes. Subsequently, based on the actual spatial location of the IoT nodes in the bridge structure, all nodes are divided into several continuous spatial regions, each corresponding to an initial task partition. The partitioning rule is: spatially adjacent nodes with a risk weight difference of less than 0.2 are grouped into the same region. For regions containing high-risk nodes, these regions are marked as isolated regions, and their data interaction with other regions is restricted in subsequent task processing, allowing only unidirectional data output to avoid phase collapse propagation.

[0056] Then, step S430 is executed. Based on the propagation path constraints in the initial task partition set and the topology mapping matrix, a task splitting criterion is constructed, and the initial task partitions are recursively decomposed to form multi-level task subsets. Specifically, for each initial task partition, it is first checked whether there are path connections with a weight greater than 0.3 between nodes in the partition in the topology mapping matrix. If so, it indicates that there is a significant risk of phase collapse propagation within the partition, and further splitting is required. The splitting method is as follows: nodes are sorted from high to low risk weight, high-risk nodes are divided into a separate subset, and the remaining nodes are further divided into two subsets according to spatial continuity. The above checking process is repeated for the newly generated subsets until the connection weight between all nodes in each subset is less than 0.3. Through this recursive decomposition process, multiple task subsets that satisfy the propagation constraints are formed.

[0057] Finally, step S440 is executed. Based on the node risk weights and data temporal continuity of each task subset, the task subsets are reorganized and scheduled to generate a dynamic task decomposition strategy that avoids the phase collapse propagation path. Specifically, for each task subset, the average value of its internal node risk weights is calculated, and the subsets are sorted from low to high according to the average value, prioritizing the processing of low-risk task subsets. During scheduling, the temporal continuity constraints in the phase correction dataset are also considered. Data segments that are temporally adjacent and belong to the same physical event are assigned to the same task subset for processing as much as possible to avoid analysis errors caused by data fragmentation. For task subsets marked as isolated regions, their processing results are only allowed to participate in the calculation in the final fusion stage and do not participate in the interaction of intermediate results. Finally, the task execution order and allocation scheme are generated based on the sorting results and constraints, which constitutes the dynamic task decomposition strategy.

[0058] Through the above steps, the constructed edge-side adaptive splitting calculation model can effectively avoid the interference of phase collapse propagation path on the calculation results while ensuring computational efficiency.

[0059] S500 performs asynchronous sharding computation at each edge node based on Q, and introduces a computation process compensation based on Φ to obtain the compensation result set D2.

[0060] In this embodiment, step S510 is first executed, dividing the phase correction dataset into multiple temporally continuous and spatially isolated task slices according to a dynamic task decomposition strategy, and then assigning each task slice to a corresponding edge node for asynchronous computation. Specifically, based on the task subsets determined in the dynamic task decomposition strategy, the phase correction dataset is divided according to a unified reference time axis. Each task slice contains data within a continuous time interval, preferably with a time length between 1 and 3 seconds to ensure a balance between computational granularity and real-time performance. Simultaneously, each task slice contains only IoT node data from the same task subset to meet spatial isolation requirements. Subsequently, each task slice is distributed according to a pre-defined edge node allocation rule: task slices with low node risk weights are preferentially allocated to edge nodes with higher computing power, while task slices with high node risk weights are allocated to relatively independent edge nodes to reduce the impact of phase collapse propagation. After distribution, each edge node independently starts the computation process asynchronously, meaning each edge node processes its received task slice independently without waiting for other nodes to complete, thereby improving overall computational efficiency.

[0061] Next, step S520 is executed. During the calculation process at each edge node, the phase distortion indicator function value at the corresponding time point is extracted in real time, and a phase distortion compensation coefficient is constructed. This compensation coefficient is then embedded into the feature weight update stage of the calculation process. Specifically, within each edge node, for the currently processed task slice, the corresponding phase distortion indicator function value is read point by point in chronological order. A phase distortion compensation coefficient is constructed based on the phase distortion indicator function value. The compensation coefficient is defined as the ratio of the absolute value of the phase distortion indicator function to a preset maximum distortion threshold, where the maximum distortion threshold is set to 1 to ensure that the compensation coefficient is between zero and one. When the absolute value of the phase distortion indicator function is close to 1, it indicates that the phase distortion at that time point is severe, and the compensation coefficient is close to zero; when the phase distortion is small, the compensation coefficient is close to 1. Subsequently, during specific calculations, such as weighted statistics or feature fusion of phase features, the compensation coefficient is introduced as a weight adjustment factor, that is, each feature value is multiplied by the compensation coefficient at the corresponding time point, thereby reducing the impact of large distortion data on the calculation results.

[0062] Then, step S530 is executed, whereby the intermediate calculation results of each task segment are corrected point-by-point according to the phase distortion compensation coefficient to obtain the compensated local calculation results. Specifically, after completing the initial calculation of the task segments, the obtained intermediate result sequence is corrected point-by-point again according to time points. The correction method is as follows: for the calculation result at each time point, it is weighted and fused with the corresponding compensation coefficient. Specifically, the original calculation result is multiplied by the compensation coefficient, and then smoothly fused with the calculation results of adjacent time points using a linear interpolation method, where the interpolation weight is determined by the compensation coefficient. When the compensation coefficient is low, the correction relies more on the results of adjacent time points to eliminate the influence of abrupt changes. In this way, local abnormal fluctuations caused by phase distortion can be effectively suppressed, and stable compensated local calculation results can be obtained.

[0063] Finally, step S540 is executed to align and stitch the compensated local calculation results output by each edge node according to a unified reference time axis, forming a compensation result set. Specifically, the local calculation results output by all edge nodes are sorted according to their corresponding time labels. For overlapping time portions, a weighted average method is used for fusion, with the weight determined by the reciprocal of the corresponding node's risk weight; that is, the lower the node's risk weight, the higher its result weight. For time discontinuities, a linear interpolation method is used to fill in the gaps, ensuring the continuity of the results on the time axis. Finally, all processed results are stitched together in chronological order to form a complete compensation result set.

[0064] Through the above steps, a phase distortion indicator function was introduced to compensate for the entire process in an asynchronous distributed computing environment. This not only ensured computational efficiency but also effectively reduced the error propagation caused by phase collapse.

[0065] S600 performs reverse consistency verification and fusion on D2 based on topology mapping T, and outputs a highly reliable state assessment result that eliminates the effects of phase collapse.

[0066] In this embodiment, step S610 is first executed to extract the backpropagation paths between IoT nodes in the compensation result set based on the topology mapping matrix, construct the backpropagation consistency constraint relationship between nodes, and calculate the backpropagation consistency weight corresponding to each node. Specifically, each element in the topology mapping matrix represents the phase collapse propagation intensity from the source node to the target node. To construct the backpropagation path, the topology mapping matrix is ​​first transposed to reverse the original propagation direction. Subsequently, for each row of the transposed matrix, the node connection relationships corresponding to all non-zero elements are extracted to form a set of backpropagation paths. For each IoT node, the weight values ​​of all its backpropagation input paths are summed, and the summation result is defined as the backpropagation influence value of that node. Further, to make different nodes comparable, the backpropagation influence values ​​of all nodes are subjected to max-min normalization. The normalized result is the backpropagation consistency weight; the larger the value, the stronger the constraint on the node in the backpropagation relationship.

[0067] Next, step S620 is executed, whereby consistency deviation detection is performed on the results of each node in the compensation result set according to the reverse consistency weight, resulting in a sequence of deviation values ​​at each time point, which is then compared with a preset consistency threshold. Specifically, for each IoT node in the compensation result set, the difference between its result value and the weighted average of its reverse neighboring node results is calculated point by point on a unified reference time axis. The weighted average is calculated using the reverse consistency weight as the weight; that is, for a given node, the results of all its neighboring nodes on the reverse propagation path are taken at the current time point, and then weighted, summed, and normalized according to the corresponding reverse consistency weight to obtain a reference consistency value. Subsequently, the absolute difference between the current result value of the node and the reference consistency value is defined as the deviation value. By performing the above calculation on all time points, a sequence of deviation values ​​is formed. To determine whether there is a consistency anomaly, a consistency threshold is set as the mean of all deviation values ​​plus 0.8 times the standard deviation. When the deviation value at a certain time point exceeds this consistency threshold, it is determined that there is a consistency deviation at that time point.

[0068] Then, step S630 is executed. For node results with deviation values ​​exceeding the consistency threshold, a step-by-step backpropagation correction is performed according to the backpropagation path to generate a corrected node result sequence. Specifically, for each time point with consistency deviation, the corresponding node and its upstream nodes on the backpropagation path are first located. Following the hierarchical relationship of the backpropagation path, the process proceeds upwards from the upstream node closest to the current node. For each level of node, its result value is weighted and fused with the result value of its next-level node, where the weight is determined by the corresponding back-consistency weight. The fusion calculation method is: the corrected value of the current node equals the sum of the original result value multiplied by its own weight and the upstream node's result multiplied by its upstream weight, then divided by the sum of the weights. Through step-by-step backpropagation and fusion, local anomalies can be gradually and smoothly diffused, thereby eliminating sudden deviations. After completing the backpropagation, the corrected result corresponding to that time point is obtained. The above process is repeated for all time points with deviations to form a complete corrected node result sequence.

[0069] Finally, step S640 is executed, where the corrected node result sequence is weighted and fused according to a unified reference time axis to obtain a highly reliable state assessment result that eliminates the impact of phase collapse. Specifically, the corrected results of all IoT nodes at the same time point are fused. The fusion weight is determined by the product of the reciprocal of the node risk weight and the reverse consistency weight; that is, the lower the node risk weight and the higher the reverse consistency weight, the greater its fusion weight. The results of each node are weighted, summed, and normalized to obtain the overall state assessment value at that time point. This fusion process is repeated for all time points to form a complete highly reliable state assessment result sequence.

[0070] Through the above steps, reverse consistency verification and step-by-step backtracking correction based on the topology mapping matrix are realized. This not only identifies and repairs local anomalies, but also improves the reliability of the overall results by utilizing the propagation relationship between nodes.

[0071] Example 2, please refer to Figure 2 As shown in this embodiment, a distributed edge computing system based on the Internet of Things includes: Three-domain coupling modeling module: Collects the original sampling sequence, clock signal and electromagnetic environment disturbance data of each IoT node, and constructs a node-level three-domain coupling model C; Phase reconstruction module: Based on C, identify the pseudo-synchronization interval of each node and generate the corresponding phase distortion indicator function Φ. Perform phase reconstruction on the original sampling sequence to obtain the phase correction dataset D1. Topology modeling module: Use D1 to extract the phase collapse entropy gradient parameters and phase hysteresis loop area parameters across nodes, and construct a topology mapping matrix T that reflects the local phase collapse propagation path; Adaptive Task Decomposition Module: Input T and D1 into the edge-side adaptive splitting calculation model to generate a dynamic task decomposition strategy Q that avoids the phase collapse path; Compensation module: Performs asynchronous sharding calculations at each edge node based on Q, and introduces a calculation process based on Φ for compensation, to obtain the compensation result set D2; Fusion output module: Performs reverse consistency verification and fusion on D2 based on topology mapping T, and outputs a highly reliable state assessment result that eliminates the effects of phase collapse.

[0072] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application.

Claims

1. A distributed edge computing method based on the Internet of Things, characterized in that: include: S100 collects the original sampling sequences, clock signals and electromagnetic environmental disturbance data of each IoT node to construct a node-level three-domain coupling model C; S200 identifies the pseudo-synchronization intervals of each node based on C and generates the corresponding phase distortion indicator function Φ. It then performs phase reconstruction on the original sampling sequence to obtain the phase correction dataset D1. S300, using D1 to extract the phase collapse entropy gradient parameters and the phase hysteresis loop area parameters across nodes, constructs a topological mapping matrix T that reflects the local phase collapse propagation path; S400 inputs T and D1 into the edge-side adaptive splitting calculation model to generate a dynamic task decomposition strategy Q that avoids the phase collapse path; S500 performs asynchronous sharding computation at each edge node based on Q, and introduces a computation process compensation based on Φ to obtain the compensation result set D2; S600 performs reverse consistency verification and fusion on D2 based on topology mapping T, and outputs a highly reliable state assessment result that eliminates the effects of phase collapse.

2. The distributed edge computing method based on the Internet of Things according to claim 1, characterized in that: The construction of the node-level three-domain coupling model includes the following steps: S110, decompose the original sampling sequence of each IoT node into a time window, extract the signal phase features and amplitude features at different time scales, and form an initial signal feature set; S120, perform frequency stability analysis and nonlinear drift fitting on the clock signal of each IoT node to obtain the corresponding clock drift function, and perform time alignment processing on the clock drift function and the initial signal feature set to obtain the clock correction feature set; S130, Perform spectral decomposition and spatial correlation modeling on electromagnetic environment disturbance data, extract electromagnetic disturbance intensity distribution features, and map them onto the time axis corresponding to the clock correction feature set to form a disturbance correlation feature set; S140, the initial signal feature set, clock correction feature set and disturbance correlation feature set are coupled and fused to construct a node-level three-domain coupling model that uniformly represents the correlation between the time domain, phase domain and energy domain.

3. The distributed edge computing method based on the Internet of Things according to claim 1, characterized in that: The method of identifying the pseudo-synchronization intervals of each node based on C includes the following steps: S210, based on a node-level three-domain coupling model, extracts the degree of coupling deviation between the time domain, phase domain and energy domain, constructs a coupling consistency judgment index, and calculates the continuous change sequence of the coupling consistency judgment index through a sliding time window; S220, based on the abrupt change magnitude and duration of the coupling consistency judgment index, and combined with the preset judgment threshold, identify pseudo-synchronization intervals, and perform boundary marking on each pseudo-synchronization interval to obtain a set of pseudo-synchronization intervals.

4. The distributed edge computing method based on the Internet of Things according to claim 3, characterized in that: The phase correction dataset D1 is obtained, including: S230, constructing a phase distortion indicator function based on the phase evolution trajectory within the pseudo-synchronization interval set and the phase reference relationship with the normal interval, and mapping the phase distortion indicator function to the corresponding time axis; S240, performing point-by-point compensation and continuity reconstruction of the phase components in the original sampling sequence according to the phase distortion indicator function to obtain the phase correction dataset D1.

5. A distributed edge computing method based on the Internet of Things according to claim 1, characterized in that: The method for obtaining the phase collapse entropy gradient parameter is as follows: S310, Delayed embedding is performed on the phase sequence in the phase correction dataset to construct a high-dimensional phase space trajectory, and the nearest neighbor search method is used to determine the local neighborhood point set based on the high-dimensional phase space trajectory; S320, construct the Riemannian manifold structure based on the geodesic distance between the local neighborhood point set, and estimate the probability density of the phase distribution in the local neighborhood based on the geodesic distance to obtain the manifold probability distribution function; S330, calculate the corresponding manifold information entropy based on the manifold probability distribution function, and perform a difference operation on the manifold information entropy in the spatial neighborhood direction to obtain the local information entropy change rate; S340, the local information entropy change rate is spatially weighted and accumulated to generate phase collapse entropy gradient parameters.

6. The distributed edge computing method based on the Internet of Things according to claim 1, characterized in that: The method for obtaining the cross-node phase hysteresis loop area parameter is as follows: S350: Select the phase sequences of any two IoT nodes from the phase correction dataset, construct a two-dimensional phase trajectory point set according to a unified reference time axis, and form phase evolution point cloud data based on the two-dimensional phase trajectory point set; S360, construct a simple complex structure for the phase evolution point cloud data according to the distance threshold increment method, and generate corresponding topological connection relationships at different distance scales; S370, calculate one-dimensional homology features based on the simple complex structure, extract closed loop structures that persist during scale changes, and record their generation and extinction scales; S380, calculate the persistence metric based on the generation and destruction scales of the closed loop structure, and map the persistence metric to the corresponding cross-node phase hysteresis loop area parameter.

7. The distributed edge computing method based on the Internet of Things according to claim 1, characterized in that: The dynamic task decomposition strategy Q that avoids phase collapse paths includes the following steps: S410: Based on the topology mapping matrix, the phase collapse propagation path strength between each IoT node is extracted, and combined with the data stability characteristics of each node in the phase correction dataset, a node risk weight sequence is constructed. S420, the phase correction dataset is spatially partitioned according to the node risk weight sequence to generate an initial task partition set, and the data regions corresponding to high-risk nodes are isolated and marked. S430, Based on the propagation path constraint relationship between the initial task partition set and the topology mapping matrix, a task splitting criterion is constructed, and the initial task partition is recursively decomposed to form a multi-level task subset; S440 reorganizes and schedules the task subsets according to the node risk weights and data temporal continuity corresponding to each task subset, generating a dynamic task decomposition strategy that avoids the phase collapse propagation path.

8. A distributed edge computing method based on the Internet of Things according to claim 1, characterized in that: Asynchronous sharding computation is performed on each edge node based on Q, including the following steps: S510 divides the phase correction dataset into multiple time-continuous and spatially isolated task slices according to the dynamic task decomposition strategy, and assigns each task slice to the corresponding edge node to perform asynchronous computation. S520 extracts the phase distortion indicator function value at the corresponding time point in real time during the calculation process at each edge node, constructs the phase distortion compensation coefficient, and embeds the compensation coefficient into the feature weight update stage of the calculation process. S530 corrects the intermediate calculation results of each task segment on a time-by-time basis according to the phase distortion compensation coefficient to obtain the compensated local calculation results. S540 aligns and stitches the compensated local calculation results output by each edge node according to a unified reference time axis to form a compensation result set.

9. A distributed edge computing method based on the Internet of Things according to claim 1, characterized in that: The reverse consistency verification and fusion of D2 based on topology mapping T includes the following steps: S610: Based on the topology mapping matrix, the back propagation path between each IoT node in the compensation result set is extracted, the back consistency constraint relationship between nodes is constructed, and the back consistency weight corresponding to each node is calculated. S620, based on the reverse consistency weight, performs consistency deviation detection on the results of each node in the compensation result set, obtains the deviation value sequence at each time point, and compares it with the preset consistency threshold; S630: For node results whose deviation values ​​exceed the consistency threshold, backtrack and correct them step by step according to the backpropagation path to generate a corrected node result sequence. S640, the corrected node result sequence is weighted and fused according to a unified reference time axis to obtain a highly reliable state assessment result that eliminates the effect of phase collapse.

10. A distributed edge computing system based on the Internet of Things (IoT), used to implement the distributed edge computing method based on the Internet of Things as described in any one of claims 1-9, characterized in that: include: Three-domain coupling modeling module: Collects the original sampling sequence, clock signal and electromagnetic environment disturbance data of each IoT node, and constructs a node-level three-domain coupling model C; Phase reconstruction module: Based on C, identify the pseudo-synchronization interval of each node and generate the corresponding phase distortion indicator function Φ. Perform phase reconstruction on the original sampling sequence to obtain the phase correction dataset D1. Topology modeling module: Use D1 to extract the phase collapse entropy gradient parameters and phase hysteresis loop area parameters across nodes, and construct a topology mapping matrix T that reflects the local phase collapse propagation path; Adaptive Task Decomposition Module: Input T and D1 into the edge-side adaptive splitting calculation model to generate a dynamic task decomposition strategy Q that avoids the phase collapse path; Compensation module: Performs asynchronous sharding calculations at each edge node based on Q, and introduces a calculation process based on Φ for compensation, to obtain the compensation result set D2; Fusion output module: Performs reverse consistency verification and fusion on D2 based on topology mapping T, and outputs a highly reliable state assessment result that eliminates the effects of phase collapse.

Images