Inspection data automatic acquisition and space-time correlation method under Beidou space-time reference

Abstract

The application relates to the field of Beidou inspection data processing technology and discloses a method for automatically collecting and correlating inspection data under a Beidou space-time reference. The method comprises the following steps: collecting historical working condition data of an inspection device based on nanosecond-level time synchronization and centimeter-level space positioning information of a Beidou system, constructing a space-time dynamic prediction model, obtaining device sensor measurement values, a moving trajectory point set and a time stamp sequence in real time, outputting a theoretical inspection value through the model, comparing the theoretical value with an actual value in a space-time multi-dimensional manner, analyzing time cumulative error, space deviation and data flow form consistency, generating a difference feature tensor, inputting the tensor into a geographic coding topological reasoning network, combining terrain map node attributes and device distribution, simulating abnormal diffusion, generating an abnormal probability cloud map, and delineating an abnormal range, and dynamically configuring an inspection parameter according to the cloud map, enabling enhanced collection for a high-probability abnormal area, and performing data credibility verification on surrounding devices.

Description

Technical Field

[0001] This invention relates to the field of BeiDou inspection data processing technology, specifically a method for automatic acquisition and spatiotemporal correlation of inspection data under the BeiDou spatiotemporal reference. Background Technology

[0002] In fields such as industrial equipment operation and maintenance and infrastructure monitoring, routine inspections are an important means to ensure stable equipment operation and timely detection of potential faults. As equipment complexity increases and the monitoring scope expands, traditional inspection methods are gradually revealing many limitations, making it difficult to meet the needs of high-precision and high-efficiency inspections.

[0003] Traditional inspection data collection relies heavily on manual recording or single positioning technology. Time synchronization accuracy is typically at the millisecond or even second level, and spatial positioning errors are often above the meter level, making it impossible to accurately bind data with spatiotemporal information. This lack of precision results in a lack of unified spatiotemporal benchmarks for data collected from different inspection devices at different times. The correlation between data is weak, making it difficult to form a complete picture of equipment operating status. Subsequent analysis is prone to problems such as data misalignment and logical contradictions, affecting the accurate judgment of equipment status.

[0004] Traditional inspections primarily rely on post-event analysis of real-time data, lacking in-depth utilization of historical operating data and failing to establish an effective dynamic prediction mechanism. In practice, anomalies are often only identified when actual inspection values ​​exceed preset thresholds, making it impossible to detect subtle changes in equipment operating trends in advance. This results in a significant lag in anomaly detection, potentially missing the optimal window for early fault intervention and increasing equipment damage and maintenance costs.

[0005] In the data comparison and analysis phase, traditional methods often focus on numerical differences in a single dimension, ignoring the cumulative effect of errors over time, spatial coordinate deviations, and the consistency of data stream patterns. For example, comparing sensor values ​​at a single moment without considering the cumulative deviation between that value and historical data from the same period, without analyzing whether the device's movement trajectory deviates from the preset path, and without verifying whether the fluctuation patterns of the data stream conform to normal operating characteristics, leads to one-sided difference identification, making it difficult to capture potential, multi-dimensionally related abnormal signals from the device, and easily resulting in missed or false detections.

[0006] Traditional methods for delineating anomaly ranges and configuring inspection parameters lack dynamic adjustment capabilities. During anomaly detection, the range is often directly defined based on anomaly data from a single device, without considering topographic map node attributes and the distribution relationship with surrounding devices to simulate the anomaly diffusion process. This results in significant discrepancies between the delineated anomaly range and the actual diffusion, leading to insufficient targeting in subsequent inspections. Furthermore, inspection parameters are often fixed, using the same collection frequency and verification standards regardless of the anomaly probability in a given area. This results in insufficient data collection in high-probability anomaly areas and wasted resources in low-probability areas. Moreover, the reliability of data from surrounding devices in the anomaly area is not specifically verified, further impacting the reliability of the inspection results. Summary of the Invention

[0007] The purpose of this invention is to provide an automatic data acquisition and spatiotemporal correlation method for inspection data under the BeiDou spatiotemporal reference, so as to solve the problems mentioned in the background art.

[0008] To achieve the above objectives, this invention provides a method for automatic acquisition and spatiotemporal correlation of inspection data under the BeiDou spatiotemporal reference, the method comprising:

[0009] Based on the nanosecond-level time synchronization and centimeter-level spatial positioning information provided by the BeiDou Navigation Satellite System, historical operating condition data of inspection equipment are collected to construct a spatiotemporal dynamic prediction model.

[0010] The sensor measurement values, movement trajectory point set and timestamp sequence of the inspection equipment are acquired in real time, and the theoretical inspection value is output through the spatiotemporal dynamic prediction model.

[0011] The theoretical inspection values ​​and actual inspection values ​​are compared and analyzed in a spatiotemporal multidimensional manner. The analysis includes the cumulative error on the time axis, the deviation on the spatial coordinates, and the morphological consistency on the data stream, generating a difference feature tensor for each inspection device.

[0012] The difference feature tensor is input into the geocoded topology inference network. Combined with the topographic map node attributes and device distribution relationship, the abnormal data diffusion process is simulated to generate an anomaly probability cloud map and delineate the anomaly inspection range.

[0013] Based on the aforementioned anomaly probability cloud map, inspection parameters are dynamically configured, including enabling enhanced acquisition mode for high-probability anomaly areas and performing data reliability checks on surrounding equipment.

[0014] Preferably, the construction of the spatiotemporal dynamic prediction model includes: performing multi-resolution analysis on historical working condition data, using ensemble empirical mode decomposition to extract the energy ratio of long-term trend and short-term fluctuation components of sensor measurements, establishing a mapping relationship between the set of mobile trajectory points and the inspection road segment through trajectory segment clustering, and using dynamic time warping algorithm to calibrate the timestamp sequence pattern under different weather conditions.

[0015] The spatiotemporal dynamic prediction model construction includes inputting preprocessed historical operating condition data into an integrated learning framework. The integrated learning framework includes a gated recurrent unit prediction component based on an equipment degradation model to generate baseline inspection prediction values, a spatial network integrating a geographic attention mechanism to correct prediction deviations caused by positioning errors, and dynamically adjusting prediction coefficients based on a waveform adapter according to the real-time collected mobile trajectory point set. The model also synchronously acquires the signal-to-noise ratio and reception strength indicators of sensor measurements, the latitude and longitude accuracy and time synchronization residual of the mobile trajectory, and the clock drift and transmission delay parameters of the timestamp sequence through a BeiDou time synchronization acquisition module.

[0016] The theoretical inspection value calculation involves inputting real-time collected data into the spatiotemporal dynamic prediction model to obtain the theoretical inspection value.

[0017] Preferably, the calculation of the theoretical inspection value includes performing adaptive smoothing processing based on meteorological conditions to eliminate observation noise caused by wind speed and sunshine; fusing the correlation attributes of sensor measurement values, movement trajectory point sets, and timestamp sequences through a spatiotemporal encoder; and outputting a theoretical inspection value that includes a normal fluctuation range, which is adaptively updated according to the health status of the equipment.

[0018] Preferably, the spatiotemporal multidimensional comparative analysis specifically includes:

[0019] Time cumulative error calculation: The theoretical inspection value and the actual inspection value are compared in a rolling window at a set time period. The asynchronously sampled data stream is aligned using a dynamic time warping algorithm. The cumulative error in each period is calculated to form a time error vector.

[0020] Spatial deviation detection: Perform spatial transformation on the set of moving trajectory points of theoretical and actual values, calculate the Euclidean distance ratio of coordinate points, extract the deviation metric of each location point, and construct a spatial deviation vector;

[0021] Data stream morphology consistency assessment: Based on the shape context algorithm, the contour distribution of theoretical values ​​and actual data streams is matched, the delay difference of data peak points is calculated, the Jason divergence of the stream interval distribution is quantified, and a morphology consistency vector is generated.

[0022] Difference feature tensor generation: The time error vector, spatial deviation vector, and morphological consistency vector are superimposed as tensors, and the scale difference is eliminated by feature weight normalization to output the difference feature tensor.

[0023] Preferably, the spatial deviation detection specifically includes: in the spatial comparison stage, extracting the corresponding movement trajectory coordinates from the theoretical inspection value and the actual inspection value, using a density clustering algorithm to divide each group of trajectory points into multi-scale grids, extracting the location density features within a preset key geographical area, wherein the key geographical area is selected to cover the movement range of common inspection areas, after extraction, quantitatively evaluating the distribution intensity of theoretical and actual data within the key geographical area, and extracting the deviation index of each path segment based on the relative difference between the two, and collecting the deviation results of all segments to construct a spatial deviation vector.

[0024] Preferably, in the data stream morphology consistency assessment, the contour matching algorithm uses the shape context distance algorithm to perform shape matching on the set of extreme points in the two data streams and identify the time delay difference.

[0025] Preferably, the abnormal region association and location step specifically includes:

[0026] Geographic topology modeling: Construct a map grid connection topology based on the location information of the inspection equipment, label the path length parameters between each grid, and overlay the propagation loss conditions caused by terrain undulations on the topology map to generate a topology model containing a path matrix and a grid reachability matrix.

[0027] Anomaly diffusion simulation: The difference feature tensor is mapped to the corresponding grid of the topology model; anomaly diffusion inference is performed based on graph convolutional network. The calculation of anomaly diffusion inference includes calculating the decay coefficient of the anomalous data according to the path length parameter, capturing cross-grid anomaly association patterns through self-attention mechanism, and simulating the migration path of anomalous data in the topology network using random walk method.

[0028] Probability cloud map generation: Statistically analyze the frequency of abnormal data occurrences for each path in the simulation diffusion, calculate the probability value of abnormal data dwelling based on the path length parameter, generate an abnormal probability cloud map covering the entire map, and mark suspicious path groups whose probability values ​​exceed the set threshold;

[0029] Physical scope delineation: Perform density clustering analysis on the anomaly probability cloud map to identify areas with high anomaly probability; delineate the physical boundaries of anomaly inspections based on the location of the inspection equipment and its geographical topological connectivity.

[0030] Preferably, the abnormal area association and location step further includes outputting a suspicious device identifier and an abnormal spread main path. The suspicious device identifier is based on the map grid connected to the suspicious path group, associating the map grid with the real inspection equipment to form a list of suspicious device identifiers, indicating possible abnormal data sources or infected equipment. The abnormal spread main path is obtained by recording the grid paths and their order traversed during each migration in the abnormal migration process simulated by random walk. Among all simulated paths, the occurrence frequency of each path is counted, and the path sequence with the most occurrences is selected as the abnormal spread main path. The output abnormal spread main path sequence is an ordered grid list, reflecting the main propagation route of abnormal data in geospatial space.

[0031] Preferably, the step of dynamically configuring inspection parameters specifically includes: when the anomaly probability value of a certain area in the anomaly probability cloud map exceeds a set threshold, sending an acquisition strategy update command to the acquisition terminal to which the area belongs, executing the increase of the data sampling rate to the initial multiple, simultaneously starting the high-frequency mode of trajectory monitoring, capturing coordinate jump events of the moving trajectory, installing an anomaly detector on the terminal side, and recording data anomaly fragments; data credibility verification is performed: applying multiple types of data stimuli to the adjacent grid devices with the most drastic changes in anomaly probability in the anomaly probability cloud map.

[0032] Preferably, the multi-type data incentives include:

[0033] Inject a sweep frequency test waveform using a programmable signal generator;

[0034] Based on the topology model and path matrix, the theoretical response curve is calculated; the measured response is collected and recorded after the test waveform is injected into each grid to obtain the measured response curve; the difference between the theoretical response curve and the measured response curve is calculated by calculating the abnormal deviation; the difference between the measured response curve and the theoretical response curve is compared to calculate the abnormal grid offset rate; when the abnormal grid offset rate exceeds the limit, it is marked as a data abnormality associated device.

[0035] Compared with the prior art, the beneficial effects of the present invention are:

[0036] Relying on the nanosecond-level time synchronization and centimeter-level spatial positioning information provided by the BeiDou Navigation Satellite System, a high-precision spatiotemporal benchmark is established for the collection of inspection data. This enables the historical operating data of the collected inspection equipment to be accurately bound to time and space tags. Data from different equipment and different time periods have a unified correlation dimension, effectively solving the problems of spatiotemporal misalignment and weak correlation of data in traditional inspections. This allows the spatiotemporal dynamic prediction model built on historical data to more realistically reflect the operating rules of the equipment, and the theoretical inspection values ​​output by the model are more in line with the actual operating status of the equipment.

[0037] In the real-time data processing stage, by acquiring sensor measurements, movement trajectory point sets, and timestamp sequences of the inspection equipment, and combining them with a spatiotemporal dynamic prediction model to output theoretical inspection values, dynamic prediction of the equipment's operating status is achieved. Compared to traditional inspection methods that rely solely on post-event analysis of actual values, this method can detect deviations in equipment operating trends earlier by comparing theoretical and actual values ​​in advance. It can detect potential anomalies without waiting for actual values ​​to exceed thresholds, thus gaining time for early intervention and reducing the risk of damage to equipment caused by the escalation of anomalies.

[0038] The spatiotemporal multidimensional comparative analysis incorporates cumulative errors on the time axis, deviations in spatial coordinates, and consistency in the data stream's shape into the analysis scope, breaking through the limitations of traditional single-value comparison. This multidimensional analysis can comprehensively capture the characteristics of equipment anomalies, not only identifying obvious numerical deviations but also discovering implicit errors caused by time accumulation, the impact of changes in the operating environment due to spatial location offsets, and fluctuations in the equipment's internal operating mechanisms reflected by data stream shape anomalies. The generated difference feature tensor can more accurately and comprehensively characterize the abnormal state of each inspected device, reducing the probability of missed and false detections.

[0039] In anomaly identification and delineation, differential feature tensors are input into a geocoded topological inference network, and the anomaly data diffusion process is simulated by combining topographic map node attributes and equipment distribution relationships. This makes the simulation of anomaly diffusion more consistent with the actual geographical environment and equipment layout characteristics. The resulting anomaly probability cloud map can intuitively present the anomaly risk level of different areas, and the delineated anomaly inspection range is more accurate. This avoids the range deviation caused by traditional methods ignoring geographical and equipment correlations, allowing subsequent inspection work to focus on high-risk areas and improve the targeting of inspections.

[0040] The design of dynamically configuring inspection parameters based on anomaly probability cloud maps optimizes the allocation of inspection resources. Enabling an enhanced acquisition mode for high-probability anomaly areas increases the data acquisition frequency and data dimensions, ensuring sufficient detailed anomaly information is obtained. Performing data reliability checks on surrounding devices helps determine whether anomalies have spread to nearby equipment or whether data from surrounding devices has been affected by the anomaly area. This dynamic adjustment method avoids the resource waste and insufficient data collection problems caused by traditional fixed-parameter inspections, improving overall inspection efficiency while ensuring inspection quality and reducing unnecessary maintenance costs. Attached Figure Description

[0041] Figure 1 This is a schematic diagram illustrating the working principle of the automatic collection and spatiotemporal correlation method for inspection data under the BeiDou spatiotemporal reference described in this invention.

[0042] Figure 2 A flowchart illustrating the adaptive processing and output in the calculation of theoretical inspection values;

[0043] Figure 3 This is a flowchart for spatial deviation detection. Detailed Implementation

[0044] 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.

[0045] Please see Figure 1 This invention provides a method for automatic acquisition and spatiotemporal correlation of inspection data under the BeiDou spatiotemporal reference. The method includes: constructing a spatiotemporal dynamic prediction model by collecting historical operating data of inspection equipment; acquiring sensor measurements, movement trajectory point sets, and timestamp sequences of the inspection equipment in real time and inputting them into the model to output theoretical inspection values; performing spatiotemporal multidimensional comparative analysis between the theoretical and actual inspection values, covering cumulative errors on the time axis, deviations on spatial coordinates, and morphological consistency in the data stream, generating a difference feature tensor for each inspection equipment; inputting the difference feature tensor into a geocoded topological inference network, combining topographic map node attributes and equipment distribution relationships to simulate the abnormal data diffusion process, generating an anomaly probability cloud map and delineating the abnormal inspection range; and dynamically configuring inspection parameters based on the anomaly probability cloud map, including enabling enhanced acquisition mode for high-probability anomaly areas and performing data credibility checks on surrounding equipment.

[0046] Example 1: See Figure 2The construction of a spatiotemporal dynamic prediction model begins with multi-resolution analysis of historical operating data of inspection equipment. An empirical mode decomposition algorithm is applied to the sensor measurement sequence to decompose the complex signal into a series of intrinsic mode functions (IMFs). Each IMF represents a fluctuation component at a specific time scale. Long-term trend components are obtained by reconstructing low-frequency IMFs, while short-term fluctuation components are characterized by the superposition of high-frequency IMFs. Energy ratio calculation is achieved by taking the ratio of the variance of each component signal to the total variance. Trajectory segmentation and clustering processes the set of moving trajectory points. Density clustering algorithms identify dense regions within the trajectory points and divide consecutive dense points into trajectory segments. Each trajectory segment is mapped to a pre-defined inspection route segment. This mapping is established by calculating the spatial distance between the centroid of the trajectory segment and the center point of the route segment; the route with the closest distance is identified as the route segment to which the segment belongs. A dynamic time warping algorithm is used to calibrate timestamp sequences under different weather conditions. Weather condition data such as temperature, humidity, and air pressure serve as auxiliary inputs. The algorithm aligns time series under different weather patterns by stretching or compressing the time axis, ensuring consistency in comparisons between sequences. The preprocessed historical operating data is fed into an ensemble learning framework for model training. This framework comprises three core components working together. A gated recurrent unit (GRU) prediction component based on an equipment degradation model receives time series of historical sensor measurements. In the GRU network structure, the update gate controls the amount of historical information transmitted to the current state, while the reset gate determines the degree to which past information is forgotten. The equipment degradation model uses a Wiener process to describe the slow decay of equipment performance over time. The output of this component is the baseline inspection prediction value. A spatial network integrating a geographic attention mechanism processes the spatial information of the moving trajectory point set. The attention mechanism calculates the spatial correlation weight between each trajectory point and a set of predefined geographic reference points. The weight values ​​reflect the degree to which the trajectory point is influenced by the reference points. Prediction bias caused by positioning errors is compensated by weighted summation of the correction values ​​of these reference points. A waveform adapter-based component dynamically adjusts the prediction coefficients based on the real-time acquired moving trajectory point set. The curvature and velocity variation characteristics of the moving trajectory are extracted as input to the adapter. The adapter outputs a set of scaling factors to adjust the final output of the GRU prediction component.

[0047] The acquisition module, synchronized by BeiDou time synchronization, plays a crucial role in the data acquisition process, synchronously acquiring three key parameters. The signal-to-noise ratio (SNR) and received signal strength (RSS) of the sensor measurements are directly calculated from the analog-to-digital converted signal; high SNR and RSS indicate good signal quality. The latitude and longitude accuracy of the movement trajectory is evaluated by reading the accuracy factor from the BeiDou positioning data, and the time synchronization residual is obtained by comparing the high-precision time provided by the BeiDou time synchronization module with the timestamp difference of the device's local clock. The clock drift parameter of the timestamp sequence is estimated by observing the changing trend of consecutive timestamp intervals, and the transmission delay parameter is determined by the transmission time of data packets from the acquisition point to the processing center. The calculation process of theoretical inspection values ​​involves inputting the real-time acquired sensor measurements, movement trajectory point set, and timestamp sequence into a trained spatiotemporal dynamic prediction model. The calculation process includes an adaptive smoothing process based on meteorological conditions. Real-time meteorological data such as wind speed and solar radiation intensity serve as observation noise parameters for the Kalman filter. The filter continuously optimizes the estimation of the true signal state through recursive calculations, effectively eliminating the influence of random observation noise. The spatiotemporal encoder is responsible for fusing the correlation attributes of multi-source data. The encoder employs a deep neural network structure. Its input layer receives sensor data points, trajectory coordinates, and timestamps. The hidden layer extracts features through convolution and pooling operations. The output layer generates a fused high-dimensional feature vector, which encodes the spatiotemporal correlation characteristics between the data. The spatiotemporal dynamic prediction model ultimately outputs a theoretical inspection value. This theoretical inspection value is not a single numerical value but a range. The normal fluctuation range is determined based on the statistical quantiles of historical data. The theoretical inspection value adaptively adjusts as the device degradation model in the ensemble learning framework is updated, reflecting the current health status of the device. The theoretical inspection value output by the model serves as a benchmark reference for subsequent spatiotemporal multidimensional comparative analysis. The entire calculation process relies on the high-precision time synchronization and spatial positioning provided by the BeiDou system, ensuring data consistency across the spatiotemporal dimensions.

[0048] The specific execution of ensemble empirical mode decomposition involves multiple iterative screening processes. When decomposing historical sensor measurement signals, the algorithm first identifies all local extrema in the signal, and then connects all local maxima and local minima using cubic spline interpolation to form upper and lower envelopes. The mean of the upper and lower envelopes is calculated and subtracted from the unprocessed signal. This process is repeated until the definition conditions of the intrinsic mode function are met, thus obtaining an intrinsic mode function. The remaining signal is used as a new input, and the above steps are repeated until no valid intrinsic mode function can be extracted. The final remaining signal is the long-term trend component. The density clustering algorithm used in trajectory segmentation clustering requires setting a neighborhood radius and a minimum point threshold. The algorithm starts from any trajectory point and searches for all points within its neighborhood radius. If the number of points exceeds the threshold, a core point is formed and the cluster is expanded; otherwise, the point is marked as a noise point. This process traverses all points and divides the trajectory into multiple density-connected clusters, each cluster being a trajectory segment. The dynamic time warping algorithm constructs a cumulative cost matrix when calibrating timestamp sequences. Each element of the matrix represents the minimum alignment cost between two sequence prefixes. It uses dynamic programming to backtrack and find an optimal path from the top left corner to the bottom right corner of the matrix. This path defines the best alignment between sequences.

[0049] The training process of the ensemble learning framework uses historical data batches, with each batch containing all relevant data within a time window. The training objective is to minimize the loss function between the model's predicted values ​​and the true values; the loss function is typically chosen as mean squared error or mean absolute error. The gated recurrent unit (GRU) prediction component receives input and updates the hidden state at each time step. The hidden state carries historical information about the sequence. The device degradation model updates degradation parameters based on new operating data after each training cycle, enabling predictions to track actual performance changes of the device. The spatial network of the geographic attention mechanism uses a dot-product attention model to calculate attention weights. The query vector is the feature vector of the current trajectory point, the key vector is the feature vector of the geographic reference point, and the value vector is the correction vector corresponding to the reference point. The attention score is calculated by the dot product of the query vector and the key vector and normalized using a softmax function. The waveform adapter is typically implemented as a multilayer perceptron, with the input being a vector of trajectory features and the output being a vector of adjusted prediction coefficients. These coefficients directly affect the output value of the GRU prediction component.

[0050] The BeiDou time synchronization acquisition module relies on receiving time signals broadcast by the BeiDou Navigation Satellite System. Internally, the module uses a highly stable temperature-compensated crystal oscillator (TCC) and maintains synchronization with the BeiDou system through a phase-locked loop (PLL) technology. The magnitude of the time synchronization residual directly reflects the synchronization accuracy. The signal-to-noise ratio (SNR) of sensor measurements is calculated by multiplying the logarithm of the signal power to the noise power by ten. The received signal strength (RSS) index is a quantitative representation of the signal amplitude. Latitude and longitude accuracy information is derived from the user distance accuracy factor in the BeiDou navigation message; a smaller accuracy factor indicates higher positioning accuracy. Clock drift is estimated by analyzing the interval changes of the timestamp sequence over a period of time using linear regression. Transmission delay is calculated using the timestamps of network time protocol messages. The adaptive smoothing process in the theoretical inspection value calculation involves a Kalman filter with two main steps: prediction and update. The prediction step predicts the current state and error covariance based on the system state equation, while the update step uses new observations to correct the predicted value and obtain the optimal estimate. The spatiotemporal encoder network structure may include one-dimensional convolutional layers for extracting time-series features, two-dimensional convolutional layers for processing spatially distributed trajectory points, and fully connected layers for fusing features from different sources. The normal fluctuation range is calculated using quantile statistics of historical data, for example, taking the 5th and 95th percentiles as the upper and lower bounds of the fluctuation range. The adaptive update mechanism for equipment health status is achieved by monitoring equipment operating parameters and comparing them with the predicted values ​​of the degradation model. When the deviation continues to exceed the threshold, the model parameters are retrained.

[0051] Example 2: See Figure 3 The spatiotemporal multidimensional comparative analysis comprises three parallel analytical dimensions: time cumulative error calculation, spatial deviation detection, and data stream morphology consistency assessment. Each dimension generates a feature vector, which are ultimately merged to form a difference feature tensor. The time cumulative error calculation uses a preset fixed time period as the analysis window, such as a complete inspection shift or a 24-hour cycle. A rolling window comparison is performed between theoretical and actual inspection values, with the rolling window sliding along the time axis at a fixed step size, covering a new data segment each time. The dynamic time warping algorithm aligns the data streams of theoretical and actual values. Since the data stream may experience asynchronous sampling due to acquisition device response delays or transmission jitter, the dynamic time warping algorithm constructs a cost matrix and finds the path with the minimum cumulative cost to achieve non-linear alignment of the two sequences. The cumulative error within each time period is calculated by summing the absolute differences between corresponding points in the aligned sequences, forming a time error vector where each element represents the total error within an analytical window. The vector length is determined by the number of analytical windows.

[0052] Spatial deviation detection focuses on the spatial difference between theoretical and actual inspection values. The detection process involves transforming the two sets of moving trajectory points using spatial coordinates. This transformation projects the latitude and longitude coordinates from the geodetic coordinate system to a Cartesian coordinate system for distance calculation. The Euclidean distance ratio is used as the basic unit of measurement to calculate the straight-line distance between theoretical and actual trajectory points, and this distance is divided by the average inter-point distance of the entire trajectory point set to obtain the standardized ratio. The deviation metric for each location point is calculated by statistically analyzing the distribution characteristics of all Euclidean distance ratios in the point's neighborhood, such as calculating the mean and standard deviation of the ratio values. The constructed spatial deviation vector contains the deviation metric value for each sampled point on the trajectory, reflecting the degree of spatial alignment between the actual and expected movement paths of the equipment. The implementation of spatial deviation detection includes a detailed sub-process. In the spatial comparison stage, it is necessary to accurately extract the corresponding moving trajectory coordinate point pairs from the theoretical and actual inspection values. A density clustering algorithm is used to divide each set of trajectory points into multi-scale grids. This algorithm can identify noise points in the trajectory points and separate the main trajectory clusters. Multi-scale gridding employs a hierarchical strategy, gradually refining from a coarse grid to an internal grid, analyzing point distribution at each level. Pre-defined key geographic areas are selected based on the range of frequent equipment activity in historical inspection data, ensuring coverage of common inspection area movement ranges. After location density feature extraction, the distribution intensity of theoretical and actual data within the key geographic areas is quantitatively evaluated. Distribution intensity can be calculated using kernel density estimation to determine the probability density value of each location point. Deviation indices for each path segment are extracted based on the relative difference between the two indices. The relative difference is represented by the ratio or difference between the actual and theoretical distribution intensity. The deviation index results for all path segments are aggregated to construct a spatial deviation vector.

[0053] Data stream morphological consistency assessment analyzes the overall shape similarity of data curves, using a shape context algorithm to match the contour distribution of theoretical and actual data streams. The shape context algorithm first detects local extrema in the data streams as feature points, constructing a log-polar histogram centered on each feature point. This histogram describes the relative positional distribution of other feature points around that point, i.e., the shape context descriptor. The matching process calculates the chi-square distance between each pair of feature point descriptors in the theoretical and actual streams and uses the Hungarian algorithm to find the optimal matching pair, achieving contour distribution matching. The time delay difference of data peak points is directly calculated from the time coordinate difference of successfully matched feature point pairs. The Jason divergence of the stream interval distribution quantifies the difference in statistical distribution between the two data streams. It first calculates the probability distribution of the time interval between data points in the theoretical and actual streams, providing a measure of distribution consistency. The generated morphological consistency vector contains multiple indicators, such as the average shape context distance, the root mean square of the peak time delay difference, and the Jason divergence value, collectively characterizing the similarity of the data stream morphology. The generation of the difference feature tensor is the final output step of spatiotemporal multidimensional comparative analysis. It involves tensor superposition of the temporal error vector, spatial deviation vector, and morphological consistency vector. Tensor superposition occurs within a three-dimensional data structure: the temporal error vector is expanded along the temporal dimension, the spatial deviation vector along the spatial dimension, and the morphological consistency vector along the feature dimension, forming a three-dimensional tensor. Feature weight normalization is a necessary step because the numerical dimensions and ranges of different vectors may vary significantly. The normalization process uses a min-max scaling method to linearly transform the values ​​of each feature dimension to the zero-to-one range, eliminating the impact of scale differences on subsequent analysis. The final output difference feature tensor is a standardized multidimensional data cube that comprehensively encodes the differences between theoretical and actual inspection values ​​in the temporal, spatial, and morphological dimensions.

[0054] In calculating cumulative time error, the size of the rolling window needs to be set according to the specific application scenario. A window that is too small will lead to drastic error fluctuations, while a window that is too large will smooth out important local anomalies. The cumulative cost path search in the dynamic time warping algorithm uses dynamic programming. The curvature of the path can be constrained by adding a penalty term to limit excessively distorted alignment. In addition to the sum of absolute differences, the sum of squared differences can also be used to amplify the contribution of larger errors in the calculation of cumulative error. Coordinate transformation in spatial deviation detection typically uses UTM projection or Gauss-Kruger projection to convert spherical coordinates to planar coordinates to reduce distance calculation errors. Standardization of the Euclidean distance ratio makes the deviation metric comparable and unaffected by the overall density of the trajectory point set. The deviation metric calculation can further incorporate orientation angle differences to comprehensively evaluate the spatial deviation of the location points. The density clustering algorithm with multi-scale grid partitioning quickly eliminates large blank areas in the coarse grid layer and accurately locates detailed trajectory features in the fine grid layer, improving analysis efficiency. The selection of key geographic areas can be dynamically adjusted and updated according to the real-time inspection route. Kernel density estimation of distribution intensity requires the selection of an appropriate bandwidth parameter; too large a bandwidth leads to over-smoothing, while too small a bandwidth introduces excessive noise. The process of aggregating path segment deviation indices needs to preserve the order information of the path segments to ensure that the spatial deviation vector can reflect the deviation trend along the path.

[0055] When constructing the shape context descriptor for data stream morphological consistency assessment, the number of radius and angle partitions in the logarithmic polar coordinate system needs to be balanced between descriptive accuracy and computational complexity. Peak point detection requires setting appropriate thresholds to avoid false detections of noise points, and delay difference calculation is only performed on reliably matching feature point pairs. Jason divergence calculation requires accurate probability distribution estimation, typically necessitating a sufficient number of data points to ensure statistical reliability. The various indices of the morphological consistency vector reflect morphological differences from different perspectives, and the weights of each index need to be determined based on the actual application scenario. In the 3D data structure generated by the difference feature tensor, the length of each dimension may differ, requiring handling of dimension alignment issues when stacking tensors. Feature weight normalization can also employ the Z-score standardization method, transforming the data into a distribution with a mean of zero and a standard deviation of one. Normalized tensor data is beneficial for subsequent machine learning model processing, preventing certain dimensions from dominating model training due to their large numerical range. The output difference feature tensor is the direct input to the subsequent geocoding topology inference network, and its quality directly affects the accuracy of anomaly detection.

[0056] Example 3: The anomaly area association and location process comprises four main stages: geographic topology modeling, anomaly diffusion simulation, probabilistic cloud map generation, and physical extent delineation. These four stages work collaboratively to transform the differential feature tensor into anomaly inspection ranges with clear geographical significance. Geographic topology modeling constructs a map grid connection topology based on the known location information of all inspection equipment. The map grid uses regular square grids to cover the entire inspection area, and each grid is assigned a unique identifier. The connection topology is generated based on the Delaunay triangulation algorithm, which connects adjacent equipment location points to form a triangular network, ensuring no cross connections within the network. Path length parameters between grids are labeled, with the path length parameter being the straight-line geographical distance between the center points of connected grids. Propagation loss conditions caused by terrain undulations are superimposed on the topology map. Terrain undulation data is obtained from a high-precision digital elevation model, and propagation loss conditions are calculated using a radio wave propagation model, which considers line-of-sight occlusion and reflection effects. Generate a topology model that includes a path matrix and a grid reachability matrix. The path matrix is ​​a sparse matrix, and its elements record whether there are direct connections between grids and the length of the connecting paths. The grid reachability matrix calculates the shortest path distance between all grid pairs using the Floyd-Warshall algorithm.

[0057] Anomaly propagation simulation maps the difference feature tensor to the corresponding grid in the topology model. The mapping process is based on the nearest neighbor principle between the grid center and the device location; each feature vector in the difference feature tensor is assigned to the grid closest to its source device. Anomaly propagation inference is performed using a graph convolutional network (GCNN), which uses the topology model as the graph structure and the mapped difference features as node features. The anomaly propagation inference calculation includes three sub-processes: calculating the attenuation coefficient of the anomaly data based on the path length parameter (the attenuation coefficient is inversely proportional to the path length; the longer the path, the greater the signal attenuation); capturing cross-grid anomaly association patterns through a self-attention mechanism, which calculates the correlation weights between all grid nodes in the graph, indicating a higher probability of anomaly propagation between grid pairs with higher weights; and simulating the migration path of anomaly data in the topology network using a random walk method. The random walk starts from the grid node with the highest difference feature value and selects the next access node based on the weight probability of the connecting edges, repeating the simulation multiple times to generate a large number of possible propagation paths. In the data stream morphology consistency assessment, the contour matching algorithm uses the shape context distance algorithm to identify the time delay difference by matching the extreme point sets in two data streams. The shape context distance algorithm constructs a feature histogram describing the relative position distribution of its surrounding points for each extreme point. The point set matching is achieved by comparing the differences between the histograms. The time delay difference is the difference between the timestamps of the successfully matched point pairs.

[0058] The probability cloud map generation process statistically analyzes the frequency of anomalous data occurrences along each path in the simulated diffusion. Frequency refers to the number of times an edge between grid cells is traversed in a random walk simulation. The path length parameter is used to calculate the dwell probability of anomalous data. Path length affects the dwell time of anomalous data on a grid cell; shorter paths mean faster migration and a lower dwell probability. A full-coverage anomaly probability cloud map is generated. This probability cloud map is a two-dimensional field where each grid cell is assigned a probability value, obtained by normalizing the weighted product of the frequencies and dwell probabilities of all grid cells. Suspicious path groups with probability values ​​exceeding a set threshold are marked. The threshold is determined based on statistical experience from historical anomalous events. Suspicious path groups are high-probability grid sequences in the probability cloud map that are connected and exceed the threshold. Physical delineation involves performing density clustering analysis on the anomaly probability cloud map. Density clustering algorithms such as DBSCAN can discover dense regions of probability values ​​in space without pre-specifying the number of clusters. Dense regions of anomaly probability are identified. These dense regions are connected regions obtained after density clustering, and the grid cells within these regions have high anomaly probability values. The physical boundaries of abnormal inspections are defined based on the location of the inspection equipment and its connection to the geographic topology. The physical boundaries are obtained by calculating the convex hull or polygonal boundaries of the dense area grid. The boundary lines must ensure that all high-probability equipment location points are included.

[0059] The attenuation coefficient calculation in anomalous diffusion simulation can be expressed using the following mathematical formula:

[0060]

[0061] Where: symbol Indicates abnormal data from the grid propagation to the grid The attenuation coefficient is a dimensionless value between zero and one. (Symbol) Represents a grid With grid The path length parameter between the two points, in meters. (Symbol) It is an attenuation factor constant related to the environmental medium, and the attenuation factor constant is expressed in reciprocals of square meters. Both sides of the equation are dimensionless.

[0062] In the geographic topology modeling stage, the choice of map grid size needs to balance computational accuracy and complexity. Smaller grids provide higher spatial resolution but increase the burden of matrix operations. Delaunay triangulation ensures that there are no excessively long connecting edges in the network, forming a relatively uniform triangular grid. Path length parameters are stored as symmetric matrices because the distance between grids is undirected. Terrain relief data is processed into elevation values ​​for each grid, and the propagation loss model can use the Okumura-Hata model or the COST231-Hata model for urban environments. The construction of the path matrix requires traversing all grid pairs to determine whether there are Delaunay edges connecting them. The calculation of the grid reachability matrix uses a dynamic programming algorithm to record the shortest path between all node pairs. The graph convolutional network used in anomaly diffusion simulation typically has multiple layers to expand the receptive field, and each layer of graph convolutional operations aggregates information from neighboring nodes. In the self-attention mechanism, the query, key, and value vectors are obtained from node features through linear transformation, and the attention weights are normalized using the softmax function. The random walk method can set a restart probability, so that the walk process has a certain chance of jumping back to the starting node, avoiding over-diffusion. The shape context distance algorithm uses a log-polar coordinate grid for histogram partitioning. The logarithmic scale is sensitive to changes in neighboring points, while the polar coordinates maintain rotation invariance. Delay difference calculation requires ensuring that the data stream timestamps have been synchronized with the BeiDou system to avoid errors introduced by system time deviations.

[0063] The frequency statistics generated by the probability cloud map require a sufficient number of random walk simulations to ensure the stability of the statistical results. The calculation of the dwell probability value can consider the grid's inherent properties, such as equipment density or historical failure rate. The visualization of the anomaly probability cloud map uses color gradients to represent probability levels, with warm colors indicating high-probability areas. Threshold settings can employ adaptive methods, such as taking the percentile of the probability value distribution. The labeling of suspicious path groups requires merging spatially adjacent grids whose probability values ​​all exceed the threshold. The density clustering algorithm for physical boundary delineation requires setting neighborhood radius and minimum number of points; these parameters affect the size and number of dense areas. Convex hull calculation uses the Graham scan algorithm or Jarvis step method to find the smallest convex polygon enclosing all grid points in the dense area. After the physical boundary is delineated, it needs to be overlaid with the actual geographic map to confirm the inspection equipment and paths within the boundary.

[0064] Example 4: The anomaly area association and location step generates two important derived information types after generating the anomaly probability cloud map: suspicious device identifiers and anomaly propagation main paths. This information provides direct target guidance for subsequent dynamic configuration of inspection parameters. Suspicious device identifiers are generated based on the map grids connected to the suspicious path groups marked in the anomaly probability cloud map. Map grids are the basic spatial units divided during the geographic topology modeling stage, and each map grid has a unique number. Associating map grids with actual inspection devices requires establishing a grid-device mapping table, which records the physical inspection device numbers contained within each map grid. The process of forming the suspicious device identifier list involves querying all map grids covered by suspicious path groups, extracting the device numbers corresponding to these grids from the grid-device mapping table, and removing duplicate numbers to form the list. The suspicious device identifier list indicates potentially abnormal data sources or infected devices; the devices in the list are the objects that require priority attention and inspection. The anomaly propagation main path is obtained by recording the grid paths and their order traversed during each migration in the anomaly migration process of the random walk simulation. The random walk simulation is a core component of the anomaly propagation simulation. The occurrence count of each path is counted across all simulated paths. The occurrence count refers to the total number of times the same grid node sequence is traversed in multiple random walk simulations. The path sequence with the highest occurrence count is selected as the main path for anomaly propagation. This path sequence is an ordered list of grid numbers, reflecting the main propagation route of the anomalous data in geospatial space. The output main path sequence for anomaly propagation is an ordered list of grids, where the list order indicates the directionality of anomaly propagation.

[0065] The dynamic configuration of inspection parameters is a closed-loop control step in the method, adjusting the data acquisition process in real time based on the output of the anomaly probability cloud map. When the anomaly probability value of a certain area in the anomaly probability cloud map exceeds a set threshold, the system sends an acquisition strategy update command to the acquisition terminal belonging to that area. The threshold is determined based on historical operational data statistics. Executing the acquisition strategy update includes three specific operations: increasing the data sampling rate to an initial multiplier, which is the basic sampling frequency when the equipment is working normally; increasing the multiplier allows for the acquisition of more detailed time-series data. Simultaneously, a high-frequency trajectory monitoring mode is activated, which increases the acquisition frequency of BeiDou positioning data and captures coordinate jump events in the movement trajectory. Coordinate jump events refer to sudden changes in the device's position between consecutive sampling points. An anomaly detector is installed on the terminal side. The anomaly detector is a lightweight rule-based judgment module that records abnormal data segments and marks them with timestamps. Data reliability verification is performed on the adjacent grid devices with the most drastic changes in anomaly probability in the anomaly probability cloud map. The area with the most drastic changes in anomaly probability is determined by calculating the probability gradient field, and multiple types of data stimuli are applied to verify the reliability of the device's response. Refer to Table 1 for the mapping relationship of suspicious device identifiers.

[0066] Table 1: Suspicious Device Identification Mapping Table

[0067]

[0068] The map grid number corresponds to the unique identifier of the grid in the geographic topology modeling. The associated device ID represents the physical device number located within that grid. The anomaly probability value comes from the grid's assignment in the anomaly probability cloud map. The status marker distinguishes between suspected sources and infected devices based on their position in the main anomaly propagation path. The generation of the main anomaly propagation path relies on the statistical results of numerous random walk simulations. Each random walk starts from a grid node with a high anomaly probability and selects the movement direction based on the weighted probability of the connecting edges. The statistical analysis of path occurrences requires setting a minimum support threshold to prevent low-frequency, accidental paths from being mistaken for the main path. The orderliness of the main anomaly propagation path reflects the spatiotemporal sequence of anomaly propagation; the first grid in the list may be the anomaly origin, and subsequent grids represent the propagation direction. The threshold settings in the dynamically configured inspection parameters need to consider the differences in importance of different areas; critical areas can use more sensitive thresholds. Data collection strategy update commands are sent to the terminal via a secure communication protocol to ensure the integrity and timeliness of the commands. The data sampling rate increase multiplier is dynamically adjusted based on device processing capabilities and network bandwidth to avoid excessive data volume causing system overload. The coordinate jump event detection algorithm in high-frequency trajectory monitoring mode needs to set a reasonable displacement threshold to filter out the small-amplitude jitter inherent in GPS positioning. The terminal-side anomaly detector uses a sliding window mechanism to analyze the data stream in real time, with the window size and sampling rate adjusted in tandem. Multiple types of data stimuli for data reliability verification are executed in parallel for adjacent grid devices, and the design of the stimuli signals considers the operating characteristics and safety constraints of the devices.

[0069] The update frequency of the suspicious device identification list is consistent with the generation cycle of the anomaly probability cloud map, ensuring the real-time nature of the list. The visualization of the main path of anomaly propagation maps an ordered grid list onto a geographic background map, forming an intuitive propagation path diagram. The thresholds in the dynamic configuration of inspection parameters can adopt an adaptive mechanism, dynamically adjusting the threshold size according to the system's operating status. The acquisition terminal needs to return confirmation information after receiving a policy update instruction to ensure the instruction is executed correctly. The massive amounts of data generated after the data sampling rate increase require effective compression and storage strategies to reduce transmission and storage pressure. The data preprocessing algorithm under high-frequency trajectory monitoring mode needs optimization to eliminate the impact of positioning errors such as multipath effects. The rule base of the terminal-side anomaly detector needs to be updated regularly to adapt to the emergence of new anomaly patterns. The results of data credibility verification are fed back to the anomaly probability cloud map generation module, forming a closed-loop optimization. The output information of the anomaly area association and positioning step provides the inspection work with a clear target area and device list, and the main path of anomaly propagation reveals the regular characteristics of anomaly propagation. The dynamic configuration of inspection parameters, by adjusting the data acquisition strategy in real time, achieves key monitoring of anomaly areas and cross-validation of data quality. The entire implementation of Example 4 constructs a complete closed loop from anomaly identification to proactive response, improving the intelligence level and fault early warning capability of the inspection system. The combined use of suspicious device identification and the main path of anomaly propagation enables maintenance personnel to quickly locate the source of the problem and take targeted measures, while the dynamic parameter adjustment mechanism ensures optimal allocation of system resources.

[0070] Example 5: Multi-type data stimulation is an active diagnostic mechanism that verifies the reliability of the equipment's data acquisition function and accurately locates abnormally associated equipment by applying controllable test signals to the monitored inspection equipment and analyzing its response. Multi-type data stimulation includes four consecutive technical actions: injecting a sweep frequency test waveform through a programmable signal generator; calculating the theoretical response curve based on the topology model and path matrix; acquiring and recording the measured response data of each grid after the test waveform injection to obtain the measured response curve; calculating the difference between the theoretical and measured response curves based on the abnormal deviation; comparing the difference between the measured and theoretical response curves, calculating the abnormal grid offset rate, and marking the abnormal grid offset rate as a data-abnormal associated device when it exceeds the limit. The programmable signal generator is integrated into the data acquisition terminal of the inspection system or used as a portable test device. The injection of the sweep frequency test waveform is the starting point of the stimulation process. A sweep frequency test waveform is a signal whose frequency changes according to a specific law over time, such as a linear sweep frequency signal that increases linearly from low to high frequency, or a logarithmic sweep frequency signal whose frequency changes according to a logarithmic law. The waveform injection method depends on the sensor type. For electrical parameter sensors, direct electrical coupling injection is used, i.e., the test signal is superimposed on the sensor signal line. For physical quantity sensors, indirect excitation methods may be used, such as applying temperature field changes to temperature sensors or mechanical vibration to vibration sensors. The characteristic parameters of the injected signal need to be carefully designed. The amplitude should be set within the normal measurement range of the equipment to avoid saturation, the frequency range should cover the normal operating frequency band of the sensor, and the frequency sweep speed should be matched with the sampling rate of the data acquisition system to ensure that the complete frequency response can be captured.

[0071] The calculation of the theoretical response curve based on the topology model and path matrix is ​​a crucial step in establishing the expected benchmark for excitation testing. The topology model, a graph structure constructed during the geographic topology modeling phase, describes the inspection equipment and its spatial connections. The path matrix quantifies the transmission characteristics between nodes in the model. The calculation of the theoretical response curve treats the entire system as a linear time-invariant network, utilizing the concept of a network transfer function. The calculation process begins at the injection point, and based on the connection weights and propagation attenuation coefficients provided by the path matrix, the theoretical amplitude and phase of the test signal reaching each node in the network are solved using signal flow graph methods or nodal analysis. The theoretical response curve is a continuous curve describing the frequency response characteristics, including amplitude and phase response, that the swept-frequency test signal should have when propagating from the injection point to the target grid device under ideal conditions. The measured response acquisition is performed simultaneously with the theoretical calculation. While the swept-frequency test waveform is injected, the system records the actual output data of the sensors in each grid device. The measured response acquisition requires a high-precision data acquisition card, which is synchronized with the BeiDou time synchronization system to ensure timestamp consistency. The acquired raw data undergoes preprocessing, including filtering to remove background noise, amplifying the signal amplitude, and analog-to-digital conversion. Response components corresponding to the test signal frequency are extracted from the preprocessed data, and a measured response curve is synthesized. The measured response curve reflects the final result of the test signal propagating through the actual path in the real physical system, being sensed and converted by the equipment, and encompasses the combined effects of equipment characteristics, transmission path effects, and environmental interference.

[0072] Anomaly deviation calculation is the core of diagnosis, achieved by quantitatively comparing the difference between the theoretical response curve and the measured response curve. The difference calculation employs multiple mathematical metrics for comprehensive evaluation, such as calculating the root mean square error between the two curves to assess the overall deviation level, calculating the correlation coefficient to assess the similarity of curve shapes, and calculating the amplitude and phase differences at specific characteristic frequency points to assess local consistency. The result of the difference calculation is a multi-dimensional vector, comprehensively quantifying the degree of inconsistency between theoretical expectations and actual observations. The calculation of the anomaly grid offset rate is a further processing of the difference vector. The offset rate is a scalar value used to comprehensively judge the degree of anomaly of a specific grid device. Calculating the anomaly grid offset rate requires considering the device's historical response data, calculating the deviation of the current difference from the historical average difference, and performing normalization. The offset rate calculation may also incorporate the difference information of neighboring grids, reducing the impact of random fluctuations through spatial smoothing techniques. When the anomaly grid offset rate exceeds a preset limit, it indicates an anomaly in the device or its data link. The limit is set based on the device's accuracy level, the statistical distribution of historical normal operation data, and the acceptable false alarm rate, typically set as several times the standard deviation of the historical offset rate mean. Marking devices associated with data anomalies is the final output of the stimulus test. The marking information is updated to the suspicious device identifier list, triggering corresponding alarms and maintenance work orders. The marking process requires recording detailed diagnostic evidence, including the anomaly's offset rate, the frequency of the greatest discrepancy, test timestamps, and other auxiliary information. Marked devices may be subject to stricter monitoring strategies during subsequent inspection data collection, such as higher sampling rates and more frequent data reliability checks, until maintenance personnel confirm and resolve the anomaly through on-site inspection.

[0073] Taking a specific vibration sensor network inspection as an example, the implementation process of multi-type data excitation is as follows: The system identifies anomaly correlations in the data of several vibration sensors in a certain area through anomaly probability cloud maps. A programmable signal generator injects a linear sweep frequency vibration signal from 5 Hz to 1000 Hz into one of the suspected sensors. This signal is generated by a portable exciter. Based on the topology model, the system calculates the theoretical acceleration response curves that the sweep frequency signal should have when propagating through the network to the surrounding vibration sensors. At the same time, the system synchronously collects and records the actual acceleration readings of all vibration sensors in the area under excitation, forming the measured response curve of each sensor. The amplitude difference and coherence function value between the theoretical curve and the measured curve of each sensor at the center frequency point of 1 / 3 octave band are calculated to obtain the difference vector. Based on the historical baseline data of the sensor over the past 30 days, the offset rate of the current difference is calculated. The system found that the offset rate of three of the sensors exceeded the limit of 2.5 times the standard deviation. The system marked these three vibration sensors as data abnormality associated devices and highlighted them in the management interface, prompting maintenance personnel to focus on checking whether the mounting base of these sensors is loose, whether the connecting cables are damaged, or whether the internal components are aging.

[0074] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A method for automatic acquisition and spatiotemporal correlation of inspection data under the BeiDou spatiotemporal reference, characterized in that, The method performs the following steps: Based on the nanosecond-level time synchronization and centimeter-level spatial positioning information provided by the BeiDou Navigation Satellite System, historical operating condition data of inspection equipment are collected to construct a spatiotemporal dynamic prediction model. The sensor measurement values, movement trajectory point set and timestamp sequence of the inspection equipment are acquired in real time, and the theoretical inspection value is output through the spatiotemporal dynamic prediction model. The theoretical inspection values ​​and actual inspection values ​​are compared and analyzed in a spatiotemporal multidimensional manner. The analysis includes the cumulative error on the time axis, the deviation on the spatial coordinates, and the morphological consistency on the data stream, generating a difference feature tensor for each inspection device. The difference feature tensor is input into the geocoded topology inference network. Combined with the topographic map node attributes and device distribution relationship, the abnormal data diffusion process is simulated to generate an anomaly probability cloud map and delineate the anomaly inspection range. Based on the aforementioned anomaly probability cloud map, inspection parameters are dynamically configured, including enabling enhanced acquisition mode for high-probability anomaly areas and performing data reliability checks on surrounding equipment. The construction of the spatiotemporal dynamic prediction model includes: performing multi-resolution analysis on historical operating data, using ensemble empirical mode decomposition to extract the energy ratio of long-term trends and short-term fluctuation components of sensor measurements, establishing a mapping relationship between the set of mobile trajectory points and the inspection road segments through trajectory segment clustering, and using dynamic time warping algorithm to calibrate the timestamp sequence pattern under different weather conditions. The spatiotemporal dynamic prediction model construction includes inputting preprocessed historical operating condition data into an integrated learning framework. The integrated learning framework includes a gated recurrent unit prediction component based on an equipment degradation model to generate baseline inspection prediction values, a spatial network integrating a geographic attention mechanism to correct prediction deviations caused by positioning errors, and dynamically adjusting prediction coefficients based on a waveform adapter according to the real-time collected mobile trajectory point set. The model also synchronously acquires the signal-to-noise ratio and reception strength indicators of sensor measurements, the latitude and longitude accuracy and time synchronization residual of the mobile trajectory, and the clock drift and transmission delay parameters of the timestamp sequence through a BeiDou time synchronization acquisition module. The theoretical inspection value calculation involves inputting real-time collected data into the spatiotemporal dynamic prediction model to obtain the theoretical inspection value.

2. The method for automatic acquisition and spatiotemporal correlation of inspection data under the BeiDou spatiotemporal reference as described in claim 1, characterized in that, The calculation of the theoretical inspection value includes performing adaptive smoothing processing based on meteorological conditions to eliminate observation noise caused by wind speed and sunshine; fusing the correlation attributes of sensor measurement values, movement trajectory point sets, and timestamp sequences through a spatiotemporal encoder; and outputting a theoretical inspection value that includes a normal fluctuation range, which is adaptively updated according to the health status of the equipment.

3. The method for automatic acquisition and spatiotemporal correlation of inspection data under the BeiDou spatiotemporal reference as described in claim 1, characterized in that, The spatiotemporal multidimensional comparative analysis specifically includes: Time cumulative error calculation: The theoretical inspection value and the actual inspection value are compared in a rolling window at a set time period. The asynchronously sampled data stream is aligned using a dynamic time warping algorithm. The cumulative error in each period is calculated to form a time error vector. Spatial deviation detection: Perform spatial transformation on the set of moving trajectory points of theoretical and actual values, calculate the Euclidean distance ratio of coordinate points, extract the deviation metric of each location point, and construct a spatial deviation vector; Data stream morphology consistency assessment: Based on the shape context algorithm, the contour distribution of theoretical values ​​and actual data streams is matched, the delay difference of data peak points is calculated, the Jason divergence of the stream interval distribution is quantified, and a morphology consistency vector is generated. Difference feature tensor generation: The time error vector, spatial deviation vector, and morphological consistency vector are superimposed as tensors, and the scale difference is eliminated by feature weight normalization to output the difference feature tensor.

4. The method for automatic acquisition and spatiotemporal correlation of inspection data under the BeiDou spatiotemporal reference as described in claim 3, characterized in that, The spatial deviation detection specifically includes: in the spatial comparison stage, extracting the corresponding movement trajectory coordinates from the theoretical inspection values ​​and the actual inspection values, using a density clustering algorithm to divide each group of trajectory points into multi-scale grids, extracting the location density features within a preset key geographical area, where the key geographical area is selected to cover the movement range of common inspection areas, and after extraction, quantitatively evaluating the distribution intensity of theoretical and actual data within the key geographical area, and extracting the deviation index of each path segment based on the relative difference between the two, and collecting the deviation results of all segments to construct a spatial deviation vector.

5. The method for automatic acquisition and spatiotemporal correlation of inspection data under the BeiDou spatiotemporal reference as described in claim 3, characterized in that, In the data stream morphology consistency assessment, the contour matching algorithm uses the shape context distance algorithm to perform shape matching on the set of extreme points in the two data streams and identify the time delay difference.

6. The method for automatic acquisition and spatiotemporal correlation of inspection data under the BeiDou spatiotemporal reference as described in claim 1, characterized in that, The specific steps for locating and associating abnormal regions include: Geographic topology modeling: Construct a map grid connection topology based on the location information of the inspection equipment, label the path length parameters between each grid, and overlay the propagation loss conditions caused by terrain undulations on the topology map to generate a topology model containing a path matrix and a grid reachability matrix. Anomaly diffusion simulation: The difference feature tensor is mapped to the corresponding grid of the topology model; anomaly diffusion inference is performed based on graph convolutional network. The calculation of anomaly diffusion inference includes calculating the decay coefficient of the anomalous data according to the path length parameter, capturing cross-grid anomaly association patterns through self-attention mechanism, and simulating the migration path of anomalous data in the topology network using random walk method. Probability cloud map generation: Statistically analyze the frequency of abnormal data occurrences for each path in the simulation diffusion, calculate the probability value of abnormal data dwelling based on the path length parameter, generate an abnormal probability cloud map covering the entire map, and mark suspicious path groups whose probability values ​​exceed the set threshold; Physical scope delineation: Perform density clustering analysis on the anomaly probability cloud map to identify areas with high anomaly probability; delineate the physical boundaries of anomaly inspections based on the location of the inspection equipment and its geographical topological connectivity.

7. The method for automatic acquisition and spatiotemporal correlation of inspection data under the BeiDou spatiotemporal reference as described in claim 6, characterized in that, The abnormal area association and location step also includes outputting suspicious device identifiers and abnormal spread main paths. The suspicious device identifiers are based on the map grids connected to the suspicious path groups, associating the map grids with real inspection equipment to form a list of suspicious device identifiers, indicating possible abnormal data sources or infected equipment. The abnormal spread main path is obtained by recording the grid paths and their order traversed during each migration in the abnormal migration process simulated by random walk. Among all simulated paths, the occurrence frequency of each path is counted, and the path sequence with the most occurrences is selected as the abnormal spread main path. The output abnormal spread main path sequence is an ordered grid list, reflecting the main propagation route of abnormal data in geospatial space.

8. The method for automatic acquisition and spatiotemporal correlation of inspection data under the BeiDou spatiotemporal reference as described in claim 1, characterized in that, The specific steps for dynamically configuring inspection parameters include: when the anomaly probability value of a certain area in the anomaly probability cloud map exceeds a set threshold, sending an acquisition strategy update command to the acquisition terminal to which the area belongs, increasing the data sampling rate to the initial multiple, simultaneously activating the high-frequency mode of trajectory monitoring, capturing coordinate jump events of the moving trajectory, installing an anomaly detector on the terminal side, and recording abnormal data segments; data credibility verification is performed by applying multiple types of data stimuli to the adjacent grid devices with the most drastic changes in anomaly probability in the anomaly probability cloud map.

9. The method for automatic acquisition and spatiotemporal correlation of inspection data under the BeiDou spatiotemporal reference as described in claim 8, characterized in that, The multi-type data incentives include: Inject a sweep frequency test waveform using a programmable signal generator; Based on the topology model and path matrix, the theoretical response curve is calculated; the measured response is collected and recorded after the test waveform is injected into each grid to obtain the measured response curve; the difference between the theoretical response curve and the measured response curve is calculated by calculating the abnormal deviation; the difference between the measured response curve and the theoretical response curve is compared to calculate the abnormal grid offset rate; when the abnormal grid offset rate exceeds the limit, it is marked as a data abnormality associated device.

Images