A real-time monitoring and early warning method and system for an internet of things device

Abstract

The application provides a kind of real-time monitoring and early warning method and system of internet of things equipment, it is related to data processing technical field, the method includes: collecting the pressure difference monitoring data of micro pressure difference sensor array output arranged on the sidewall of high-pressure natural gas pipeline flange plate upstream and downstream pipe section, and the strain monitoring data of fiber optic strain sensing node output attached to the surface of flange connecting bolt group;The pressure difference monitoring data and strain monitoring data are fused according to time sequence, to construct original time series monitoring data;Cross-node clock drift compensation and multi-channel sampling rate remapping processing are performed on original time series monitoring data, to obtain space-time alignment time series dataset.The application improves the accuracy and early warning effectiveness of real-time monitoring of high-pressure natural gas pipeline flange part.

Description

Technical Field

[0001] This invention relates to the field of data processing technology, and in particular to a method and system for real-time monitoring and early warning of Internet of Things (IoT) devices. Background Technology

[0002] Flange connections, as the most widely used detachable connection method in pipeline systems, directly determine the safety and reliability of pipeline operation through their sealing performance. Long-term exposure to the combined effects of high-pressure media impact, environmental temperature cycling, structural vibration fatigue, and aging of sealing materials can easily lead to microscopic gaps at the flange sealing interface, gradually developing into leakage accidents. Industry statistics show that leakage accidents caused by flange seal failure account for more than 40% of all accidents in long-distance natural gas pipelines, resulting in significant energy losses and environmental pollution. They can also potentially trigger major safety accidents such as explosions and fires, affecting the stability of regional energy supply.

[0003] Currently, the industry generally adopts a single sensor acquisition mode for flange condition monitoring. A certain long-distance natural gas pipeline in operation only deploys a single type of fiber optic strain sensor on the surface of the flange bolts to collect bolt deformation data. It relies on a globally unified fixed threshold to assess operational risks, and the monitoring data is transmitted to a remote platform via a public wireless network for uploading and analysis.

[0004] This monitoring method has technical drawbacks. The monitoring signal is easily covered by low-frequency disturbances from environmental thermal stress and high-frequency noise from micro-vibrations of pipeline fluid, and the weak characteristic signals caused by early micro-leakage cannot be effectively extracted. During wireless transmission, data link breaks often occur in areas with weak signals, such as mountainous areas and valleys, and there is a lack of adaptive interpolation compensation mechanism, resulting in interruption of monitoring time sequence. At the same time, it is difficult to build a dynamic degradation model to track the gradual change trend of sealing performance, and the static early warning judgment form is also difficult to adapt to the dynamic changes of the baseline under different operating conditions, which easily leads to a large number of missed and false early warnings, making it difficult to accurately identify early sealing hazards and achieve stable and reliable real-time monitoring and early warning of pipeline flanges around the clock. Summary of the Invention

[0005] This invention provides a method and system for real-time monitoring and early warning of Internet of Things (IoT) devices, which improves the accuracy and effectiveness of real-time monitoring of flange locations in high-pressure natural gas pipelines.

[0006] To solve the above-mentioned technical problems, the technical solution of the present invention is as follows:

[0007] A first aspect includes a method for real-time monitoring and early warning of Internet of Things (IoT) devices, the method comprising:

[0008] Differential pressure monitoring data is collected from the micro differential pressure sensor array deployed on the sidewalls of the upstream and downstream pipe sections of the high-pressure natural gas transmission pipeline flange, as well as strain monitoring data is collected from the fiber optic strain sensing node attached to the surface of the flange connection bolt group.

[0009] Differential pressure monitoring data and strain monitoring data are fused in time series to construct the original time series monitoring data; cross-node clock drift compensation and multi-channel sampling rate remapping processing are performed on the original time series monitoring data to obtain a spatiotemporally aligned time series dataset;

[0010] Frequency band orthogonal decomposition is performed on the spatiotemporally aligned time series dataset to obtain the low-frequency drift component of environmental thermal stress disturbance and the high-frequency pulsation component of structural micro-vibration transmission; an interference suppression mask is constructed based on the low-frequency drift component and the high-frequency pulsation component.

[0011] An interference suppression mask is applied to a spatiotemporally aligned time series dataset to perform background noise filtering, thereby obtaining a leakage-sensitive component. The frequency band energy distribution spectrum and transient amplitude trajectory of the leakage-sensitive component are then calculated to obtain a feature-decoupled monitoring sequence.

[0012] The monitoring features decouple the monitoring sequence from the data packet integrity identifier on the wireless convergence link. When a data link break segment is identified, adaptive trajectory interpolation compensation is performed on the link break segment through the phase evolution law of the frequency band energy distribution spectrum to obtain a time-continuous full-dimensional state feature stream.

[0013] The full-dimensional state feature flow is mapped to a multi-dimensional degenerate phase space to construct a dynamic baseline drift tracking model to calculate the flange seal failure evolution index; the flange seal failure evolution index is then compared with a preset graded early warning threshold matrix to determine the boundary penetration and obtain real-time monitoring and early warning results.

[0014] Secondly, a real-time monitoring and early warning system for Internet of Things (IoT) devices includes:

[0015] The acquisition module is used to acquire differential pressure monitoring data output by the micro differential pressure sensor array deployed on the sidewalls of the upstream and downstream pipe sections of the high-pressure natural gas transmission pipeline flange, as well as strain monitoring data output by the fiber optic strain sensing node attached to the surface of the flange connection bolt group.

[0016] The fusion module is used to fuse differential pressure monitoring data and strain monitoring data in a time series manner to construct the original time series monitoring data; cross-node clock drift compensation and multi-channel sampling rate remapping processing are performed on the original time series monitoring data to obtain a spatiotemporally aligned time series dataset;

[0017] A module is built to perform frequency band orthogonal decomposition on a spatiotemporally aligned time series dataset to obtain the low-frequency drift component of environmental thermal stress disturbance and the high-frequency pulsation component of structural micro-vibration transmission; an interference suppression mask is built based on the low-frequency drift component and the high-frequency pulsation component.

[0018] The calculation module is used to apply the interference suppression mask to the spatiotemporally aligned time series dataset to perform background noise filtering, obtain the leakage sensitive component, and calculate the frequency band energy distribution spectrum and transient amplitude trajectory of the leakage sensitive component to obtain the feature decoupling monitoring sequence.

[0019] The identification module is used to monitor the integrity identifier of the data packets of the decoupled monitoring sequence on the wireless convergence link. When a data break segment is identified, adaptive trajectory interpolation compensation is performed on the break segment through the phase evolution law of the frequency band energy distribution spectrum to obtain a time-continuous full-dimensional state feature stream.

[0020] The judgment module is used to map the full-dimensional state feature flow to the multi-dimensional degenerate phase space, construct a dynamic baseline drift tracking model to calculate the flange seal failure evolution index; and perform boundary penetration judgment between the flange seal failure evolution index and the preset graded early warning threshold matrix to obtain real-time monitoring and early warning results.

[0021] Thirdly, a computing device, comprising:

[0022] One or more processors;

[0023] A storage device for storing one or more programs that, when executed by one or more processors, cause the one or more processors to implement the method.

[0024] Fourthly, a computer-readable storage medium storing a program that, when executed by a processor, implements the method.

[0025] The above-described solution of the present invention has at least the following beneficial effects:

[0026] By employing techniques such as synchronous acquisition and fusion of multiple physical quantities including micro-pressure difference and fiber optic strain, spatiotemporal alignment correction of time-series data, frequency band orthogonal decomposition and interference masking for noise reduction, phase-constrained chain break interpolation compensation, multi-dimensional phase space degradation modeling, and boundary penetration classification, this approach overcomes the technical problems of traditional single monitoring signals being easily masked by environmental disturbances, discontinuous data time-series breaks, inability to quantify the degree of gradual seal deterioration, and high false alarm rates in static threshold warnings. This allows for the effective extraction of sensitive features of early micro-leakage, ensuring complete and consistent monitoring data, accurately assessing the evolution of flange seal failure, scientifically classifying leakage risk levels, and improving the accuracy and effectiveness of real-time monitoring of flange locations in high-pressure natural gas pipelines. Attached Figure Description

[0027] Figure 1 This is a flowchart illustrating a real-time monitoring and early warning method for IoT devices provided by an embodiment of the present invention.

[0028] Figure 2This is a schematic diagram of a real-time monitoring and early warning system for Internet of Things (IoT) devices provided in an embodiment of the present invention. Detailed Implementation

[0029] Exemplary embodiments of the present disclosure will now be described in more detail with reference to the accompanying drawings. While exemplary embodiments of the present disclosure are shown in the drawings, it should be understood that the present disclosure may be implemented in various forms and should not be limited to the embodiments set forth herein. Rather, these embodiments are provided so that this disclosure will be thorough and complete, and will fully convey the scope of the disclosure to those skilled in the art.

[0030] like Figure 1 As shown, an embodiment of the present invention proposes a real-time monitoring and early warning method for Internet of Things (IoT) devices, the method comprising the following steps:

[0031] Step 1: Collect differential pressure monitoring data output by the micro differential pressure sensor array deployed on the sidewalls of the upstream and downstream pipe sections of the high-pressure natural gas transmission pipeline flange, and strain monitoring data output by the fiber optic strain sensing node attached to the surface of the flange connection bolt group.

[0032] Step 2: Merge differential pressure monitoring data and strain monitoring data according to time series to construct original time series monitoring data; perform cross-node clock drift compensation and multi-channel sampling rate remapping processing on the original time series monitoring data to obtain a spatiotemporally aligned time series dataset;

[0033] Step 3: Perform frequency band orthogonal decomposition on the spatiotemporally aligned time series dataset to obtain the low-frequency drift component of environmental thermal stress disturbance and the high-frequency pulsation component of structural micro-vibration transmission; construct an interference suppression mask based on the low-frequency drift component and the high-frequency pulsation component.

[0034] Step 4: Apply the interference suppression mask to the spatiotemporally aligned time series dataset to perform background noise filtering, obtain the leakage sensitive component, and calculate the frequency band energy distribution spectrum and transient amplitude trajectory of the leakage sensitive component to obtain the feature decoupling monitoring sequence.

[0035] Step 5: Monitor the integrity identifier of the data packets on the wireless convergence link by decoupling the monitoring sequence. When a data break segment is identified, adaptive trajectory interpolation compensation is performed on the break segment by the phase evolution law of the frequency band energy distribution spectrum to obtain a time-continuous full-dimensional state feature flow.

[0036] Step 6: Map the full-dimensional state feature flow to the multi-dimensional degenerate phase space, construct a dynamic baseline drift tracking model to calculate the flange seal failure evolution index; perform boundary penetration judgment between the flange seal failure evolution index and the preset graded early warning threshold matrix to obtain real-time monitoring and early warning results.

[0037] In this embodiment of the invention, an integrated monitoring technology is employed, which combines the collaborative acquisition of dual physical quantities of micro-pressure difference and fiber optic strain, time-series fusion and spatiotemporal alignment correction of multi-source monitoring data, frequency band orthogonal decomposition combined with interference suppression masking for noise reduction, adaptive link break compensation for spectral phase evolution, dynamic baseline tracking modeling of multi-dimensional degraded phase space, and graded threshold boundary penetration judgment. This overcomes the technical problems of traditional flange monitoring methods, such as single physical quantity, misaligned data time sequence, difficulty in filtering environmental and structural interference noise, missing wireless transmission data links, inability to quantify the gradual failure process of the seal, and poor accuracy of static threshold early warning. This effectively removes redundant interference signals, accurately extracts sensitive features of early micro-leakage, ensures complete and continuous monitoring time sequence, dynamically quantifies the degree of flange seal failure evolution, and achieves refined graded early warning of pipeline flange leakage risk. This significantly improves the anti-interference capability, data integrity, and early warning accuracy of high-pressure natural gas pipeline flange equipment monitoring.

[0038] In a preferred embodiment of the present invention, step 1 above may include:

[0039] Step 11: Collect differential pressure monitoring data output from the micro differential pressure sensor array deployed on the sidewalls of the upstream and downstream pipe sections of the high-pressure natural gas transmission pipeline flange, and strain monitoring data output from the fiber optic strain sensing nodes attached to the surface of the flange connection bolt group. Specifically, this includes: deploying monitoring sensing equipment in the high-pressure natural gas transmission pipeline; arranging the micro differential pressure sensor array on the sidewalls of the upstream and downstream pipe sections of the pipeline flange; uniformly arranging the sensing elements within the array along the circumference of the pipe wall with consistent spacing; attaching and fixing fiber optic strain sensing nodes to the surface of the bolt group at the flange connection location; installing one sensing node on the outer wall of each bolt; ensuring stable signal acquisition by ensuring close contact between the sensing elements and the surface of the measured structure; continuously acquiring the pipe wall differential pressure signal formed by the flow of the medium inside the pipeline; and outputting corresponding differential pressure monitoring data according to a fixed sampling frequency; and sensing the deformation of the bolts under load and synchronously outputting corresponding strain monitoring data. The total amount of data acquired by the sensing equipment in a single sampling is set to satisfy the formula: In the formula, This represents the number of data samples acquired within a single data collection period. This represents the number of data points generated per second by the sensing device. This represents the duration of a single continuous data collection. After the two types of sensing devices complete data collection, the raw monitoring signals are uniformly transmitted to the edge processing terminal to summarize and collect the two types of monitoring data.

[0040] In this embodiment of the invention, multiple types of sensing devices respectively pick up pipeline pressure difference and bolt deformation status information, enriching the monitoring dimensions. The layout conforms to the actual structural form of the pipeline flange, and the sensing points cover the key areas of sealing stress, so as to stably obtain basic monitoring data that can reflect the flange sealing condition.

[0041] In a preferred embodiment of the present invention, step 2 above may include:

[0042] Step 21: Extract the original acquisition timestamps carried by the differential pressure monitoring data and strain monitoring data respectively. Using a preset reference clock as a reference, calculate the initial time offset of each sensing node relative to the reference clock. Specifically, this includes: reading the data frames output by the micro differential pressure sensor array and the fiber optic strain sensing node respectively; locating the offset address of the timestamp field after the frame header identifier according to their respective communication protocols, such as Modbus-RTU, CANopen, or the frame format definition of a custom serial frame; extracting the original byte sequence of timestamps recorded in each data frame; and decoding it into a computable numerical time stamp according to the encoding format specified by the protocol, such as Unix timestamp, BCD code, or UTC string. ,in The sensor nodes are numbered, and their timestamps are arranged in the order of acquisition to form an independent original timestamp sequence for each node. The second pulse signal output from the GPS time receiver or a local high-stability crystal oscillator clock synchronized via the NTP network time protocol is used as the preset reference clock source. During system initialization, the reference clock and the clocks of each sensor node are initially aligned and recorded. During operation, the monotonically increasing time scale of the reference clock is continuously maintained for each sensor node. The first valid sampling time of the original timestamp sequence is taken, and the reference time value corresponding to the same physical time is retrieved from the reference clock scale. The difference between the two is the initial time offset. All sensor nodes are traversed and summarized into an initial offset vector. Outlier detection is performed on the initial offset vector: the absolute deviation of each offset relative to the median offset is calculated, and abnormal nodes with deviations exceeding three times the absolute deviation threshold of the median are removed. After re-performing clock calibration or marking the abnormal nodes as nodes to be calibrated, the verified initial offset vector is output.

[0043] Step 22: Construct time axis alignment parameters based on the initial time offset. Apply these parameters to the differential pressure monitoring data and strain monitoring data, and perform sliding interpolation alignment along a unified time grid. Then, stitch and fuse the aligned data in chronological order to construct the original time-series monitoring data. This specifically includes: setting the initial time offset of each sensor node... Inverting the values ​​yields the time axis translation calibration for each node, i.e., the alignment parameters. This parameter means: if the original timestamp of a node lags behind the reference clock. Then its time axis needs to be shifted in the negative direction. To eliminate hysteresis and align the time stamp of the node's data with the reference clock scale, the translation calibration values ​​of all nodes are organized into an alignment parameter set according to the node number, with the reference clock scale as the horizontal axis and the system's preset target sampling interval as the reference. The grid step size is determined based on the Nyquist sampling theorem and the highest frequency of leakage characteristics, starting from the earliest start time of the time axis after all nodes are aligned. Until the latest termination time Divide the time grid into equal intervals , To unify the total number of grid points in the time grid, The target sampling interval preset for the system determines the formula for calculating the time dimension length of the final output time series data:

[0044] ;

[0045] For each grid point on the unified time grid Sliding window interpolation is performed on the original time series of each sensor node: Set the interpolated window width to the center of the window. Covering one standard sampling interval on each side, the sampled area within the window falls within... The original sampling points within the interval are used to calculate the value of the sampled points within the window using a cubic spline interpolation algorithm, while maintaining the continuity of the second derivative. For physical quantities such as pressure difference or strain, if the number of effective sampling points within the window is less than the minimum number of points required for a cubic spline, it degenerates into linear interpolation or nearest neighbor preservation. The pressure difference data interpolation sequence and the strain data interpolation sequence are then stitched together frame by frame according to a unified time grid. Within each time frame, the interpolation data from each pressure difference sensing node channel and the interpolation data from each strain sensing node channel are arranged sequentially, distinguished by channel identifiers, and stacked frame by frame along the time axis to construct a two-dimensional original time-series monitoring data matrix with time frames as the behavior and monitoring channels as the columns. ,in, To unify the total number of grid points in the time grid, This represents the total number of channels.

[0046] Step 23: Using the original time-series monitoring data as the processing object, extract the actual time interval sequence of each data channel within the continuous sampling period, calculate the cumulative clock drift error between the actual time interval sequence and the standard sampling period, and generate a clock drift compensation sequence based on the cumulative clock drift error. Specifically, this includes: the small fluctuation in the actual time interval between two consecutive physical samples of the underlying crystal oscillator of each sensor node; tracing back the original data frames of each sensor node; extracting the actual timestamp difference between two adjacent samples; and constructing the time interval for each channel. Actual sampling time interval sequence ,in, This represents the total number of original sampling points for this channel. For channel No. There are 10 actual sampling intervals, of which 10 are 10 ... For the first Channel No. The original timestamps of each sampling point For the first Channel No. The original timestamps of each sampling point, the actual sampling time interval sequence of each channel, and the system's preset standard sampling period are compared. By subtracting point by point, the instantaneous drift deviation at each sampling interval is obtained. This deviation is then summed along the time direction to obtain the channel value. In the The cumulative clock drift error at each sampling interval is denoted as . The physical meaning of this accumulated error is that, due to a small deviation of the crystal oscillator frequency of the sensing node from the nominal frequency, after experiencing... After one sampling period, the cumulative time offset of the node's local clock relative to the reference clock, if the crystal oscillator frequency is too fast. For positive and follow The error increments; if it's too slow, it becomes negative and accumulates in the negative direction. The accumulated clock drift error sequence for each channel is then inverted to obtain the clock drift compensation sequence. ,in The total number of original sampling points for this channel; the meaning of the compensation sequence is that at each sampling position of the original data, an additive should be applied to the time axis. The offset is used to compensate for accumulated clock drift, and the compensation sequence is stored by channel number.

[0047] Step 24: Apply the clock drift compensation sequence to the original timing monitoring data to perform time axis translation and proportional stretching correction, outputting clock synchronization data; perform equal-interval resampling and physical dimension normalization processing on the clock synchronization data to obtain a spatiotemporally aligned timing dataset, specifically including: processing the original timing monitoring data matrix... Each channel The timeline is used to apply the clock drift compensation value of this channel to each sampling point, thus changing the original time stamp. Updated to corrected time stamp ,in For the first The sensor channel in the first The compensation time offset at each sampling point is used to eliminate the overall time axis offset caused by crystal oscillator drift, so that the time mark of each channel returns to the reference clock scale. The frequency deviation of the clock crystal oscillator causes the overall translation of the time axis and also causes non-uniform scaling of the local sampling interval, that is, the spacing between adjacent sampling points on the time axis is not uniform. According to the first difference of each channel compensation sequence, that is, the instantaneous drift rate per sampling point, the scaling factor of the local time axis is calculated: if the drift rate is positive, it means that the sampling interval is stretched and compression correction needs to be applied; otherwise, stretching is required. The corrected non-uniform time series is mapped to the uniform time grid through local interpolation to complete the proportional stretching correction. After joint correction, the clock synchronization data is output. For the clock synchronization data, in order to eliminate the inconsistency of residual sampling rate and the slight non-uniformity introduced by interpolation, equal-interval resampling is performed at a uniform target sampling rate. The resampling process includes two steps.

[0048] Anti-aliasing low-pass filtering is applied to the clock synchronization data signals of each channel, with the cutoff frequency set at half the target sampling rate to filter out components higher than the Nyquist frequency. Then, decimation or interpolation is performed on the filtered signals according to the target sampling interval to ensure that each channel has a completely consistent number of sampling points and time coordinates. For equally spaced resampled data, physical dimension normalization is performed: the differential pressure values ​​of each channel of the micro-differential pressure sensor are uniformly mapped to the kilopascal dimension. The original data may be millivolt voltage signals, which need to be converted to differential pressure values ​​according to the sensor calibration curve before dimension normalization. The optical fiber should... The strain values ​​of each channel of the variable sensing node are uniformly mapped to the micro-strain dimension. The original data may be wavelength drift, which needs to be converted into strain according to the fiber grating strain sensitivity coefficient before the dimension is unified. The data of each channel after dimension normalization are numerically comparable and computable. The normalized data of each channel are organized into a spatiotemporally aligned time series dataset according to a unified time axis and a unified sampling rate. The core attributes of the dataset include: all channels share the same reference clock scale, the sampling interval is strictly consistent, the physical dimensions are unified and comparable, and the sensor spatial location information is bound to the channel identifier.

[0049] In this embodiment of the invention, a five-step progressive processing chain—initial time offset calculation, sliding interpolation alignment, cumulative drift point-by-point compensation, equal-interval resampling, and physical dimension normalization—elevates the spatiotemporal inconsistency of multi-source heterogeneous sensor data layer by layer from four dimensions: time synchronization, error correction, sampling unification, and dimension normalization. This chain uses GPS / NTP reference clocks as a unified reference to eliminate initial time deviations between nodes, uses the cumulative difference between the actual sampling interval and the standard period to drive time axis translation and proportional stretching dual correction to eliminate long-term crystal oscillator drift, uses equal-interval resampling to unify the multi-channel sampling density, and uses physical dimension normalization to establish numerical comparability across physical quantities. The four-step processing collaboratively ensures that the final spatiotemporally aligned time series dataset has three major characteristics: accurate time base, uniform sampling, and unified dimensions, avoiding feature misassociation and early warning misjudgment caused by spatiotemporal misalignment of data.

[0050] In a preferred embodiment of the present invention, step 3 above may include:

[0051] Step 31: Perform orthogonal frequency band decomposition on the spatiotemporally aligned time series dataset. Perform multi-level orthogonal decomposition operations along the frequency axis to divide the spatiotemporally aligned time series dataset into several independent frequency band sub-signal sequences. Specifically, based on the prior frequency band characteristics of the flange seal micro-leakage signal, the signal energy caused by the micro-deformation of the sealing interface is mainly distributed in the mid-low frequency band from several hertz to tens of hertz. Select orthogonal wavelet basis functions with tight support and high-order vanishing moment characteristics, or select empirical mode decomposition class adaptive basis functions, and preset the number of decomposition levels. This ensures that the center frequency of the lowest frequency general component after decomposition is lower than the lower bound of the thermal stress gradual change range, and the number of decomposition layers... The determining principle is ,in To establish a uniform sampling frequency for spatiotemporally aligned time-series datasets, To define the lower frequency bound of the thermal stress gradual variation range, orthogonal decomposition is performed layer by layer on the signal sequence of each channel in the spatiotemporally aligned time series dataset. Taking orthogonal wavelet decomposition as an example, the first... Layer decomposition will be the first The low-frequency profile components of the layer are passed through a low-pass filter. and high-pass filter Convolution and downsampling are performed separately, and the result is decomposed into the first... Low-frequency overview components and high-frequency detail components of the layer:

[0052] ;

[0053] in, For the first Low-frequency profile components of the layer at discrete sampling points The coefficient value at that point represents the frequency of the signal below a certain value. The smoothing trend component; Discrete sampling points; For the first High-frequency detail components at discrete sampling points The coefficient value at that point represents the frequency in the signal located at... Detailed fluctuation components within the range; and These are the low-pass and high-pass filter coefficient sequences, respectively. For filter coefficient index; For the first Low-frequency profile components of the layer at sampling points The value at that location, when hour, This refers to the original input signal. Each layer of decomposition will decompose the current low-frequency profile components. Downsampling, bandwidth halved, recursive execution After layering, a lowest frequency profile component is obtained. and High-frequency detail components ,total A sequence of orthogonal frequency band sub-signals, for The sub-signal sequences are used to verify their mutual orthogonality: the inner product of any two different sub-signals, the sum of the products at each sampling point should approach zero; the sum of the energies of all sub-signals should be equal to the total energy of the original signal. After verification, the sub-signal sequences corresponding to each channel are organized and stored in a two-dimensional index structure according to the channel number decomposition level.

[0054] Step 32 involves selecting several independent frequency band sub-signal sequences as the filtering objects, extracting sub-signal segments with frequency ranges within a preset thermal stress gradual change range as low-frequency drift components, and extracting sub-signal segments with frequency ranges within a preset mechanical vibration pulse range as high-frequency pulsation components. Specifically, this includes reading the preset thermal stress gradual change range from the system configuration parameter library. and mechanical vibration pulse range These two sets of frequency band boundary parameters were determined by a combination of thermo-mechanical coupled finite element simulation analysis and field hammer impact modal tests of the flange connection structure: the thermal stress gradual change interval corresponds to the frequency range of the flange structure's slow thermal expansion and contraction response caused by the diurnal and seasonal changes in ambient temperature; the mechanical vibration pulse interval corresponds to the frequency range of mechanical waves transmitted from pipeline fluid pulsation, compressor vibration, and valve opening and closing impacts through the pipe wall to the flange connection structure. The two sets of intervals do not overlap on the frequency axis, thus satisfying the condition... traverse all of each channel For each sub-signal sequence, extract its center frequency. Center frequency Defined as the geometric mean or Fourier peak frequency of the corresponding frequency band of the sub-signal, it will satisfy... All sub-signal segments of the condition are marked as low-frequency candidate sub-signals. The candidate sub-signals are then superimposed and synthesized in the time domain on a sampling point-by-sampling basis to obtain the low-frequency drift component of the channel.

[0055] ;

[0056] in, For channel The low-frequency drift component; For the sub-signal layer index set, For channel In the The sub-signal sequence of the layer mainly reflects the slow macroscopic deformation of the flange sealing interface caused by environmental thermal stress. Similarly, traversing all sub-signal sequences of each channel, all sub-signal segments are marked as high-frequency candidate sub-signals. These high-frequency candidate sub-signals are then superimposed and synthesized point by point in the time domain to obtain the high-frequency pulsation component of the channel. This component mainly reflects the small-amplitude high-frequency response of mechanical vibration transmitted to the flange connection through the pipeline structure. The intermediate frequency band sub-signal sequence not included has a frequency range of... and Between these frequencies, the frequency bands that are sensitive to leakage characteristics are reserved and do not participate in the synthesis of interference components. The low-frequency drift components and high-frequency pulsation components of each channel are organized according to the channel index.

[0057] Step 33: Perform instantaneous energy density calculations on the low-frequency drift component and the high-frequency pulsation component respectively to obtain the low-frequency energy distribution sequence and the high-frequency energy distribution sequence. Specifically, this includes: setting the sliding window parameters for the instantaneous energy density calculation, selecting the window type (rectangular window or Hamming window), as rectangular windows are simpler to calculate but have spectral leakage, while Hamming windows offer better sidelobe suppression, and setting the window width. and sliding step size Window width The selection of the time interval needs to balance time resolution and energy estimation stability. If it is too short, the variance of energy estimation will be large; if it is too long, the time resolution will decrease. Generally, a time interval of 10 ... The sliding step size is 3 to 5 times the period corresponding to the lowest frequency in the thermal stress gradual change range. Pick to To ensure appropriate overlap between adjacent windows, the mean square energy value of the signal within the window is calculated by sliding the window position along the time axis for low-frequency drift components. Let the first window be... The center time of each window is The instantaneous energy density of the low-frequency component at that moment is calculated using a window function. Cut exist For a segment within an interval, calculate the weighted average of the squared amplitudes of each sampling point within that segment:

[0058] ;

[0059] In the formula, For the first The low-frequency drift component of each sensing channel in the first... The center moment of the sliding window The instantaneous energy density at a given point has the same dimension as the square of the signal amplitude; if the signal has been normalized, it is dimensionless. The sliding window width used for low-frequency energy calculation, i.e., the total number of sampling points contained within the window, is dimensionless. The time of the sampling point within the window relative to the center of the window The offset index is a dimensionless integer; The normalized weighting coefficient of the window function at offset position n is dimensionless. For the first The center time of the sliding window is seconds. For rectangular windows, a simple arithmetic mean is taken, and after traversing all window positions, the output channel is determined. Low-frequency energy distribution sequence For high-frequency pulsation components, window parameters and calculation methods are used. In order to eliminate the local fluctuations in the energy sequence that may be caused by the sliding window truncation effect, median filtering or exponential weighted moving average smoothing is performed on the energy distribution sequence to obtain the smoothed low-frequency energy distribution sequence and high-frequency energy distribution sequence. The smoothed energy sequences of each channel are summarized according to the channel index.

[0060] Step 34: Using the low-frequency energy distribution sequence and the high-frequency energy distribution sequence as mapping benchmarks, assign corresponding signal attenuation weights according to the degree to which the energy proportion of each frequency band exceeds the preset interference threshold, thus obtaining a frequency domain filtering weight set containing dynamic attenuation coefficients for multiple frequency bands. Specifically, this includes: calculating the proportion of low-frequency interference energy and the proportion of high-frequency interference energy at the center time of each window; for the low-frequency band, calculating the ratio of low-frequency energy to the total signal energy; and for the high-frequency band, calculating similarly. The total signal energy is the sum of the energies of all frequency band sub-signals at that moment, which can be obtained by directly accumulating the energies of each sub-signal during decomposition. Preset interference threshold parameters are read: low-frequency interference threshold and high-frequency interference threshold, and the upper limit quantile of the energy proportion determined by statistical analysis of historical data under normal operating conditions (e.g., taking the 95th quantile). For each frequency band at each moment, an attenuation coefficient is allocated according to rules: For the low-frequency band at moment: if the low-frequency energy proportion at that moment does not exceed the low-frequency interference threshold, then the attenuation coefficient is... That is, all low-frequency components are retained without attenuation; if the threshold is exceeded, the attenuation coefficient is calculated according to the linear mapping of the degree of excess.

[0061] ;

[0062] in This is a low-frequency attenuation adjustment factor, which controls the rate at which the attenuation coefficient decreases with increasing energy percentage after exceeding a threshold. The closer it is to 1, the more rapid the decay. This is the attenuation coefficient; the smaller the value, the stronger the suppression of low-frequency interference components at that moment. Low-frequency interference threshold; The same rules are used to calculate the attenuation coefficient in the high-frequency band, replacing the parameters with the corresponding interference threshold and attenuation adjustment factor for the high-frequency band. The low-frequency attenuation coefficient is then extended and mapped to all sub-signal layers belonging to the low-frequency candidate sub-signal set. These layers all use the same attenuation coefficient at any given time. The high-frequency attenuation coefficient is then extended and mapped to the high-frequency candidate sub-signal set. For all sub-signal layers within the band, the attenuation coefficient is always set to 1 for the intermediate frequency band and the leakage-sensitive frequency band. For the lowest frequency profile component, its attenuation coefficient is calculated and applied after being proportional to the low-frequency attenuation coefficient. By summing the attenuation coefficients of all frequency bands and all times, a system is constructed. The set of multi-band dynamic attenuation coefficients for each channel, i.e., the set of frequency domain filter weights. ,in For time indexing, For frequency bands, sub-signal layer indexes, Each element in the table represents the attenuation weight value for that frequency band at that moment.

[0063] Step 35 involves binding and mapping the frequency domain filter weight set with the spatial distribution characteristics of the sensors to construct an interference suppression mask. Specifically, this includes: reading the three-dimensional spatial installation coordinates of each sensor node from the system configuration file or database, unifying all coordinates to a global coordinate system with the flange geometric center as the origin; the pressure tap position coordinates of each sensor in the micro-differential pressure sensor array can be directly calculated from the geometric parameters of the pipe axis and the flange end face; the bonding position coordinates of each grating in the fiber optic strain sensing node need to be determined based on the circumferential installation angle, nominal diameter, and length of the bolt, transforming the bonding position from the local coordinate system of the bolt surface to the global coordinate system to ensure that the coordinates of the two types of sensor nodes are in the same spatial reference system, thus constructing a spatial position matrix. ,in, The total number of channels is the sum of the number of channels in the micro differential pressure sensor array and the number of fiber optic strain sensing nodes. OK For the first The three-dimensional coordinates of the sensing nodes corresponding to each channel in the global coordinate system. Based on Matrix, calculate any two sensor nodes and Spatial Euclidean distance between Organize the distance values ​​of all node pairs into a spatial distance matrix. Calculating spatial correlation coefficients based on spatial distance: An exponentially decaying spatial correlation function is used to calculate the spatial correlation coefficient based on spatial distance. Mapped to spatial correlation weights ,in The spatially relevant length parameter, determined by finite element simulation analysis of heat conduction and stress transfer in the flange structure, characterizes the attenuation scale of the interference signal's influence in space. The range of values ​​is The closer the distance, the stronger the correlation and the closer the weight is to 1; the farther the distance, the weaker the correlation and the closer the weight is to 0.

[0064] Using the spatial correlation matrix, the frequency domain filter weight set is spatially weighted and smoothed for each channel. At any moment ,frequency band The attenuation weight at a given location is determined by the channel's own energy proportion and also takes into account the attenuation weights of its spatially neighboring channels. This weighted fusion is performed based on spatial correlation coefficients. Through this operation, the interference suppression mask achieves a smooth and continuous spatial distribution. Even if a sensor is not directly installed at a certain location, reasonable interference suppression parameters can be obtained by interpolating the attenuation weights of neighboring sensors through spatial correlation. The spatially smoothed multi-channel, multi-time, and multi-band attenuation weight set is then arranged along the time axis (…). ), frequency axis ( ) and spatial coordinates ( The three dimensions are organized into a three-dimensional tensor structure. Each element value of this tensor represents the time at time t. ,frequency band Spatial location The signal attenuation coefficient that should be applied at the point, This is an interference suppression mask.

[0065] In this embodiment of the invention, a five-step progressive processing link—orthogonal frequency band decomposition, interference component screening and separation, instantaneous energy density calculation, dynamic attenuation coefficient allocation, and spatial distribution feature binding—is used to achieve a complete closed loop from frequency domain interference identification to time-frequency-space three-dimensional joint interference suppression. This link utilizes orthogonal wavelet decomposition along the frequency axis to decompose the monitoring signal into multi-level independent frequency band sub-signals. Using the a priori calibrated thermal stress gradual change range and mechanical vibration pulse range as screening criteria, it accurately separates two types of background interference—environmental thermal-induced low-frequency drift and mechanical vibration high-frequency pulsation—from the leakage-sensitive mid-frequency band. Then, a sliding window instantaneous... Using energy density as the mapping benchmark, the attenuation coefficient is dynamically allocated according to the degree to which the energy proportion of the interference frequency band exceeds the threshold of normal operating conditions. The stronger the interference, the deeper the attenuation; the weaker the interference, the shallower the attenuation. The frequency domain attenuation weight is smoothly extended along the spatial distribution of the sensor based on the spatial correlation length, so that the interference suppression effect is continuously distributed in three-dimensional space, avoiding the suppression blind zone caused by the discrete spatial layout of the sensor. In the time domain, the mask can adaptively track the dynamic changes of the interference intensity. In the frequency domain, it can accurately act on the disassembled and separated interference frequency band without damaging the leakage sensitive frequency band. In the spatial domain, it can achieve continuous coverage of the entire monitoring area.

[0066] In a preferred embodiment of the present invention, step 4 above may include:

[0067] Step 41: Extract the multi-band dynamic attenuation coefficients from the interference suppression mask. Perform a frequency-point weighted multiplication operation on the multi-band dynamic attenuation coefficients and the frequency domain transformation results of the spatiotemporally aligned time series dataset. Perform background interference signal band suppression processing to filter out the spectral characteristics of low-frequency drift components and high-frequency pulsation components. Output the residual time series signal after background purification. Specifically, this includes: by channel index. Iterate through the signal sequences of each channel in the spatiotemporally aligned timing dataset one by one. For each channel signal, a window width consistent with the mask time grid is applied along the time axis. and step length Perform frame segmentation to obtain Each time-domain frame signal segment , For each frame of time-domain signal segment, a discrete Fourier transform is performed to map it from the time domain to the frequency domain, thus obtaining the complex spectrum sequence of that frame.

[0068] ;

[0069] In the formula, For channel In the Frame, First The complex spectral values ​​at each frequency point contain amplitude and phase information; For discrete frequency indexing, For time-domain sampling point index, ; The number of sampling points per frame is equal to the window width. With uniform sampling rate The product; For channel No. Intra-frame The signal amplitude at each sampling point; The imaginary unit, ; Let be a natural constant, Physical frequency corresponding to the index Frequency band index of frequency domain filter weight set Perform the corresponding matching. For the first... Each frequency band determines the range of discrete frequency points it covers. ,in and These are the lower and upper bounds of the frequency band, respectively, from the perspective of the interference suppression mask. Index by Channel Time Index and frequency band index Extract the attenuation weight value at the corresponding position. Assign this weight value to the frequency band All frequency points covered Perform a weighted multiplication operation on the spectrum point by point:

[0070] ;

[0071] In the formula, This is the weighted spectrum value, i.e., the spectrum after background interference has been suppressed; The attenuation weights extracted from the mask have a value range of [value range missing]. When the energy proportion of this frequency band is lower than the interference threshold When the energy percentage significantly exceeds the threshold ; The original spectrum values ​​before weighting; For the first A set of discrete frequency point indices covering all frequency bands. Repeat this operation to complete the frequency-by-frequency weighting of the entire frame's spectrum, and then apply the weighted spectrum... Performing an inverse discrete Fourier transform maps the frequency domain weighted result back to the time domain, reconstructing the time domain signal segment of the frame after background suppression:

[0072] ;

[0073] In the formula, The reconstructed time-domain discrete signal sequence; This refers to the number of sampling points per frame. These are the weighted spectral values; For discrete frequency indexing; For time-domain sampling point index; The imaginary unit; It is a natural constant, taking into account the existence between adjacent frames. In the overlapping region, a weighted smooth stitching process is performed on the overlapping regions of adjacent frames. The overlapping regions are either cross-fade-in / fade-out using linearly varying weights, or the overlapping and summing methods are used to directly superimpose and take the average value to eliminate frame boundary discontinuities. Finally, the residual time-series signals of each channel after background purification are obtained by stitching. .

[0074] Step 42: Using the residual time-series signal after background purification as the processing object, perform time-frequency domain inverse mapping reconstruction to remove residual background baseline noise and extract the leakage-sensitive component characterizing the micro-deformation of the flange sealing interface. Specifically, this includes: processing the residual time-series signal... Perform again with orthogonal wavelets Layer decomposition, to obtain From the sub-signal sequences, extract the sub-signal layer set belonging to the intermediate frequency band and the leakage characteristic sensitive frequency band. , Defined as: all that satisfy The set of sub-signal layer indices, where For the first The center frequency of the layer, for In each sub-signal layer, further detrending processing is performed to remove residual low-frequency baseline noise. The detrending process employs a local weighted regression smoothing method: a local low-order polynomial trend line is fitted to the sub-signal sequence window by window; the fitted trend line is subtracted from the original sub-signal to obtain the detrended residual sequence. This residual sequence represents the pure leakage-sensitive sub-signal after removing residual baseline drift. All detrended sub-signals are then processed... The inner sub-signal layers are superimposed and synthesized at each sampling point in the time domain to obtain the channel leakage sensitive component;

[0075] ;

[0076] In the formula, For channel The leakage-sensitive component at the sampling point The value at; For the first The sub-signal after detrending at the sampling point The value at; For the set of sub-signal layer indices corresponding to the leakage-sensitive frequency band, for all Each channel performs the above operation, outputting the leakage-sensitive component sequence of each channel, which is then passed to two parallel processing branches.

[0077] Step 43: Input the leakage-sensitive component into the sliding frequency domain analysis window, perform frequency band energy integration along the time axis, calculate the cumulative energy value within each target frequency band interval, and plot the frequency band energy distribution spectrum of the leakage-sensitive component. Specifically, this includes setting the frequency domain analysis window parameters, such as window width... and The selection principle is consistent, taking into account both time resolution and frequency resolution; the sliding step size is... Target frequency band segmentation: Based on the prior frequency band characteristics of flange seal micro-leakage, the leakage-sensitive frequency band is segmented. Further subdivided into Sub-band The frequency boundaries of each sub-band are The partitioning method can be either equal bandwidth partitioning or partitioning by octave band, depending on the leakage-sensitive components. Along the time axis by step Slide window, for the first The frequency spectrum amplitude is obtained by performing a discrete Fourier transform on the signal segments within the window. For each sub-band Calculate the energy integral value within this frequency band, which is to sum the squares of the spectral amplitudes at discrete frequency points covered by this frequency band:

[0078] ;

[0079] In the formula, For channel At any moment Sub-band The cumulative energy value within; For the first The signal spectrum within the window is at the 1st The amplitude at each frequency point is obtained by taking the modulo operation on the complex number spectrum; Sub-band The set of discrete frequency point indices covered. ,in For the first For each frequency point, the physical frequency is used to normalize the energy value of each sub-band according to the total energy of each sub-band at the same time. The percentage of energy of each sub-band to the total energy at that time is calculated. The normalized energy sequence is organized into a two-dimensional matrix according to the time axis and the frequency band axis. With time as the horizontal axis, frequency band as the vertical axis, and energy value as color depth or height value, the frequency band energy distribution spectrum of the leakage sensitive component is obtained.

[0080] Step 44: Based on the time-domain waveform characteristics of the leakage-sensitive component, extract the time interval and amplitude change gradient between adjacent extreme points, construct the signal envelope evolution path, trace the peak fluctuation pattern of the signal envelope evolution path along the time axis, and extract the transient amplitude trajectory. Specifically, this includes: for the leakage-sensitive component... Extreme point detection is performed on the time-domain waveform by traversing the sampling point sequence and detecting local maxima and local minima. The criterion for determining local maxima is as follows: and The criteria for determining local minima are similar; we take the value less than or equal to and record the sampling index of each extreme point. and the corresponding signal amplitude ,right Pairs of adjacent maxima in The time interval and amplitude change are calculated. The time interval is the ratio of the difference in sampling indices between adjacent extreme points to the sampling rate. The amplitude change is the difference between the amplitude of the subsequent extreme point and the amplitude of the previous extreme point. The amplitude change gradient between two adjacent maxima is the ratio of the amplitude change within that interval to the time interval. Similarly, for... The amplitude change gradients of adjacent minimum points are calculated, and the calculation results of the maximum and minimum points are merged to construct the signal envelope evolution path of the channel, which is the upper and lower envelope curves formed by connecting the coordinate pairs of each local extreme point in time sequence. Interpolation operations are performed on the upper and lower envelope curves respectively, using cubic spline interpolation or piecewise cubic Hermite interpolation to extend the discrete envelope values ​​defined at the extreme points to all sampling point positions, resulting in continuous upper and lower envelope lines. Median filtering or Savitzky-Golay filtering is performed on the extended envelope lines to eliminate local oscillations introduced by interpolation. Based on the peak fluctuation law of the upper envelope line, the instantaneous change of its envelope peak value is tracked along the time axis, and the amplitude sequence of the upper envelope at each local maximum point is extracted as the transient amplitude trajectory. This trajectory reflects the time-domain energy fluctuation law of the leakage sensitive component: when micro-leakage occurs intermittently or the leakage intensity changes, this trajectory will show the corresponding amplitude jump or gradual change characteristics.

[0081] Step 45: Synchronously align and concatenate the frequency domain feature vector of the band energy distribution spectrum and the time domain feature vector of the transient amplitude trajectory according to the time index to obtain the feature decoupling monitoring sequence. Specifically, this includes: checking whether the time grid of the band energy distribution spectrum and the time grid of the transient amplitude trajectory are consistent. Since the same sliding step size is used, the two time grids are theoretically naturally aligned. If the amplitude trajectory deviates at individual moments due to the discrete effect of extreme point detection, then the transient amplitude trajectory is remapped using the time grid of the band energy distribution spectrum as the reference to ensure that the two have the same sampling time on a unified time axis. For each moment... , frequency band energy distribution spectrum middle The energy values ​​of each sub-band are arranged in a fixed order to form the frequency domain feature vector at that moment;

[0082] ;

[0083] In the formula, For channel At any moment of 3D frequency domain feature row vector; For channel At any moment No. Normalized energy value of each sub-band For each moment Take the transient amplitude trajectory value at that moment. This is concatenated with the amplitude difference values ​​at adjacent time points to form the time-domain feature vector at that time point;

[0084] ;

[0085] In the formula, For channel At any moment 2D time-domain feature row vectors; This represents the transient envelope amplitude at that moment; The first-order backward difference of the amplitudes at adjacent time points. This reflects the rate of change of the envelope, and indexes the frequency domain feature vector and the time domain feature vector at the same time. The following horizontal stitching is performed to obtain the channel at that moment. 3D fusion feature vector;

[0086] ;

[0087] In the formula, For channel At any moment The fused feature vector has a dimension of ; Indicate vector concatenation operation, for all The splicing process is executed at specific times to form a channel. Feature decoupling monitoring sequence For all Perform the same operation on each channel, decoupling the features of each channel and monitoring the sequences by channel index. Organization as a three-dimensional tensor The first dimension is the channel dimension, and there are a total of There are one sensing channel, and the second dimension is the time dimension, totaling [number missing]. The time window is defined as follows: the third dimension is the feature dimension. 2D frequency domain energy characteristics plus 2D time domain amplitude characteristics, This is the final output, which is transmitted to the leak location and graded early warning system based on the joint determination of pressure difference-strain dual-mode characteristics.

[0088] In this embodiment of the invention, a five-step progressive processing link is used to achieve a complete signal refinement and feature structuring process from interference filtering to feature decoupling. This link uses an interference suppression mask as a spectrum screening tool. The time-domain signal is mapped to the frequency domain through frame-by-frame Fourier transform. The multi-band dynamic attenuation coefficient in the mask is used to perform frequency-point weighted multiplication of the spectrum, so that the low-frequency drift component caused by environmental thermal stress and the high-frequency pulsation component caused by mechanical vibration are accurately attenuated in the frequency domain. The suppressed spectrum is reconstructed back to the time domain through inverse Fourier transform. While filtering out background interference, the signal integrity of the leakage-sensitive mid-frequency band is preserved to the maximum extent. Through wavelet domain sub-signal layer screening and detrending processing, the baseline noise remaining after frequency domain suppression is removed, and the pure leakage-sensitive component characterizing the micro-deformation of the flange sealing interface is extracted.

[0089] The frequency domain branch performs sub-band energy integration on the leakage-sensitive component using a sliding frequency domain analysis window, expanding the energy of the leakage signal along the frequency axis to obtain a frequency band energy distribution spectrum reflecting the energy distribution pattern of each frequency band. The time domain branch uses extreme point detection and envelope interpolation techniques to track the peak fluctuation pattern of the signal envelope, extracting the transient amplitude trajectory reflecting the transient changes in leakage intensity. The two features are synchronously spliced ​​according to a unified time reference to obtain a time-frequency dual-dimensional decoupled structured monitoring sequence. This sequence corresponds the frequency domain energy distribution characteristics and the time domain transient amplitude change characteristics of the leakage signal at the same time, preserving the frequency band resolution capability of the spectrum analysis and the transient tracking capability of the time domain waveform. It realizes the complementary expression of frequency domain features and time domain features, effectively solving the technical problems of insufficient signal-to-noise ratio and poor feature distinguishability in the early detection of micro-leakage by traditional single-dimensional feature analysis.

[0090] In a preferred embodiment of the present invention, step 5 above may include:

[0091] Step 51: Read the data packet sequence number and verification identifier carried by the feature decoupling monitoring sequence during the wireless aggregation link transmission to obtain the link transmission status mark stream; based on the link transmission status mark stream, locate the gaps where the verification identifier is missing and the sequence number undergoes discontinuous jumps, and extract the gaps as data link break segments. Specifically, this includes: when the feature decoupling monitoring sequence is transmitted from each sensor node to the edge aggregation node via the wireless aggregation link, the fused feature vector of each frame... They are encapsulated into individual data packets, each with a data packet sequence number. and checksum Two metadata fields: the sequence number is assigned monotonically increasing according to the transmission time sequence, and the check identifier is a fixed-length checksum calculated from the payload of the data packet using a cyclic redundancy check algorithm. At the edge aggregation node receiver, the sequence number and check identifier are extracted packet by packet according to the reception timestamp, and then sorted by channel index. and time index ,in The original arrangement order is organized as a two-dimensional link transmission state mark stream. , for link transmission state marking stream For each data packet, the receiving end recalculates the checksum for the packet's payload and compares the calculated checksum with the checksum identifier carried within the packet. A bit-by-bit comparison is performed. If the two do not match, it is determined that a bit error occurred in the data packet during transmission and is marked as a check failure node. All check failure nodes are recorded by channel and time index to obtain a check failure node set. The link transmission status mark stream is then traversed. The sequence number field of each data packet, along the time axis The sequence number continuity is checked packet by packet. Sequence number continuity is determined by comparing the difference between adjacent packet sequence numbers with a preset expected step size. Under normal transmission conditions, the comparison rule is that the difference between adjacent packet sequence numbers is 1: If... Then the serial number is in and If a discontinuous transition occurs between intervals, indicating data packet loss, the node with the failed verification is merged with the interval of discontinuous sequence number transitions. The node with the failed verification is considered to have unusable data, and the interval of sequence number transitions is considered to have missing data for the corresponding time period. The time index intervals corresponding to both types of anomalies are uniformly marked as empty intervals. Connectivity analysis is performed on the empty intervals along the time axis, merging temporally adjacent empty intervals into the same data break segment. The start time of each data break segment is recorded. Termination time and the corresponding channel index Output a list of data break segments for each channel.

[0092] Step 52: Extract the frequency band energy distribution spectrum within a preset time window before and after the data link breakage segment, extract the slope and curvature characteristics of the phase angle of each frequency band changing with time, and construct a phase evolution trend benchmark. Specifically, this includes: for each data link breakage segment, using its starting time... and termination time Using the boundary as the boundary, cut off the preset lengths forward and backward respectively. The context time window is used to obtain the previous neighbor window. and the next neighbor window Data packets within both the front and back neighbor windows pass verification and have consecutive sequence numbers, indicating complete and usable data. Complex spectral information for each frequency band is extracted from the energy distribution spectrum within these two windows, and then processed in the front neighbor window. and the next neighbor window At various times within the interior The complex spectrum, in sub-band Phase angle extraction per frequency band, for channel At any moment Sub-band The corresponding complex spectrum values, with their real and imaginary parts denoted as , respectively and Then the phase angle of that frequency band at that moment is calculated by the following formula;

[0093] ;

[0094] In the formula, For channel At any moment , No. The phase angle of each sub-band, in radians; This represents the imaginary part of the complex spectrum within this frequency band; This represents the real part of the complex spectrum within this frequency band; Let be the arctangent function, which is applied to all time points and all time points within the preceding and following neighborhood windows. Perform the above extraction on each sub-band to obtain a discrete sequence of phase angle changes over time for each band. The phase angle sequence is used to calculate the first and second rates of change along the time axis to characterize the phase evolution trend. The first rate of change is calculated using the central difference scheme.

[0095] ;

[0096] In the formula, For channel , No. Sub-band at time The phase angle slope, in radians per second; and These are the phase angle values ​​for the preceding and following moments, respectively. For the time interval between adjacent moments, the second rate of change is further centrally differencing based on the slope:

[0097] ;

[0098] In the formula, For channel , No. Sub-band at time The phase angle curvature, in radians per second²; and These are the phase angle slope values ​​at adjacent time points; The time interval between adjacent moments is the time interval for the previous neighbor window. Take the end of the window The forward trend feature is obtained by weighting the phase angle slope and curvature at each time step, with data closer to the chain break segment given higher weight; the subsequent neighborhood window... Take the starting point of the window. The backward trend feature is obtained by weighted averaging of the phase angle slope and curvature at each time step. The forward trend feature and the backward trend feature are then concatenated in chronological order for each sub-band. Form a line from arrive The continuous phase evolution trend baseline curve characterizes the pattern under which the phase angle should smoothly transition within the link break segment of this frequency band, in full... The phase evolution trend benchmarks of each sub-band are combined to form the phase evolution trend benchmark of the broken link segment.

[0099] Step 53: Using the phase evolution trend as a topological constraint, perform manifold trajectory deduction and adaptive interpolation fitting on the missing data nodes within the data break segment to obtain the break-chain compensation time-series data segment. Specifically, this includes: taking the complete fused feature vectors within the front and back neighborhood windows of the break segment as reference sample points. Each fused feature vector is Q+2 dimensional: the first Q dimensions are frequency domain energy features, and the last 2 dimensions are time domain amplitude features. Since there is coupling correlation between the dimensions within the feature vector, the frequency band energy distribution and transient amplitude trajectory reflect the frequency and time domain characteristics of the same leakage physical process. The reference sample points are distributed in the high-dimensional feature space near a low-dimensional manifold. The manifold trajectory deduction includes: performing principal component analysis to reduce the dimensionality of the front and back neighborhood reference sample point sets, taking the first d principal components with a cumulative variance contribution rate ≥95%, and obtaining the d-dimensional manifold embedding space. Let the reference sample point set matrix be... Let N be the total number of samples. After decentering X, the covariance matrix is ​​calculated as follows: In the formula, Given the sample mean matrix, perform eigenvalue decomposition on the covariance matrix. , It is an eigenvalue diagonal matrix. Given an eigenvector matrix, sort the eigenvalues ​​from largest to smallest and select the cumulative variance contribution rate. The first d principal components are used to construct a d-dimensional manifold embedding space. In the d-dimensional manifold space, the local weighted centroids of the last N embedding points in the previous neighborhood are taken as the starting positions, and the local weighted centroids of the first N embedding points in the next neighborhood are taken as the target positions, where N and N are both 5. The weights increase linearly as the distance between the sample point and the boundary of the broken chain segment decreases. The formula for calculating the weighted centroids is as follows:

[0100] ;

[0101] ;

[0102] In the formula, The weights are linear distances; the smaller the boundary distance, the larger the weight. For the neighboring embedded sample points, the phase evolution trend benchmark is mapped to the inference direction vector in the manifold space:

[0103] ;

[0104] In the formula, The weight of the b-th sub-band is the average energy percentage of that band within the window before and after the link break. For the b-th sub-band at time... The slope of the phase angle is obtained from the phase angle difference between adjacent times. ; Let be the unit projection direction vector of the eigenvector corresponding to subband b in the manifold space, from Start, by step length Progressing gradually along the manifold, among which for arrive The geodesic distance is calculated using the following formula: M represents the number of missing data nodes within the broken link segment. The direction of each step is taken from the local tangent space direction of the current point. The weighted synthesis, with weight coefficients of 0.6 and 0.4 respectively, is shown in the following formula: The local tangent space basis vectors of the reference sample point set are calculated using the local linear embedding algorithm. Specifically, for the k nearest neighbor reference sample points of the current inference point, a local reconstruction weight matrix W is constructed, and the local reconstruction error is minimized.

[0105] ;

[0106] Solving for the optimal reconstruction weight matrix Then, the orthogonal basis vectors of the local tangent space are calculated by eigenvalue decomposition of the weight matrix. Using the local coordinate system of the manifold, a local projection correction is performed after each step: the newly generated coordinate points are projected onto the local tangent space of the reference sample point set to ensure that the coordinates after advancement do not deviate from the local principal direction of the manifold. This process is repeated iteratively until the target is reached. The neighborhood is mapped back to the original Q+2 dimensional feature space by the inverse principal component analysis of the coordinates derived from the manifold space. The inverse transformation formula is:

[0107] ;

[0108] In the formula, Let the coordinates be derived from the d-dimensional manifold. Let be the eigenvector matrix of the first d principal components. The original sample mean matrix is ​​used to map the coordinate sequence derived from the manifold space back to the original Q+2 dimensional feature space through the inverse transformation of principal component analysis, thus obtaining preliminary estimates of the fused feature vectors at each missing time point. At this point, the preliminary estimates already satisfy the manifold distribution constraints and phase evolution trend constraints. To eliminate the junction step, adaptive interpolation fitting correction is performed on the preliminary estimates. Using the N intact feature vectors at the end of the previous neighborhood window and the N intact feature vectors at the beginning of the subsequent neighborhood window as boundary conditions, weighted smoothing correction is performed on the preliminary estimates at each time point within the broken chain segment along the time axis, feature dimension by feature. The correction weights are smoothly transitioned, with larger values ​​taken at both ends of the broken chain segment near the intact data and smaller values ​​taken in the middle of the broken chain segment. After completing the correction dimension by dimension, the final fused feature vectors at each missing time point are obtained. The final fused feature vectors at all missing time points are arranged in the original time index order, constituting the broken chain compensation time series data segment for this channel and this broken chain segment.

[0109] Step 54: Seamlessly embed the broken link compensation time series data segment into the corresponding position on the original time axis of the feature decoupling monitoring sequence, replace the broken link data segment, perform full-dimensional feature vector reconstruction, and output a time-continuous full-dimensional state feature stream. Specifically, this includes: for each channel... For each data break segment, using the start and end times of the break segment as the positioning reference, the original feature decoupling monitoring sequence is used. Determine the index range corresponding to this segment on the time axis. Link break compensation time sequence data segment Time index from arrive Strictly aligned with the original timeline, the fused feature vector at each compensation time point has a dimension of [missing value]. Consistent with the dimension of the original feature vector, in the original feature decoupling monitoring sequence In the middle, the index range Existing missing or failed verification nodes within the time series data segment are compensated for with broken links. The compensation feature vector at the corresponding time point is replaced node by node, while maintaining the continuity of the time index and the dimension of the feature vector during the replacement process. For the connection at the boundary of the broken chain segment, i.e. and between, and Between these points, since the junction step has been eliminated through weighted smoothing correction, the eigenvectors at the boundary transition smoothly along all dimensions, resulting in a smooth transition for the replaced channels. Full-time domain feature sequence, time-by-time traversal verification: confirming time index from 1 to... Continuous and complete, the fused feature vector at each time step has a dimension of 1. And there are no non-numerical anomalies, for all Perform the same link break compensation embedding and verification operation on each channel, and then sort the complete feature sequences of each channel by channel index. Reorganized into a three-dimensional tensor The first dimension is the channel dimension. The first sensing channel has one dimension, and the second dimension is the time dimension. A continuous time window, with the third dimension being the feature dimension. Two-dimensional frequency domain energy characteristics plus two-dimensional time domain amplitude characteristics. That is, a temporally continuous, full-dimensional state feature flow.

[0110] In this embodiment of the invention, a systematic compensation and full-dimensional feature restoration for data loss caused by wireless transmission link breakage is achieved through a four-stage progressive processing of the link: link state marking stream construction, frequency band phase evolution trend benchmark extraction, manifold trajectory deduction and adaptive interpolation, and seamless embedding of link breakage compensation. The link uses data packet sequence numbers and check identifiers to construct a link transmission state marking stream. Through a dual mechanism of sequence number continuity check and check anomaly detection, the data link breakage segment is located, ensuring that no data loss of any type caused by bit errors or whole packet loss is missed. By extracting the first-order slope and second-order curvature features of the phase angle of each frequency band as a function of time from the intact data before and after the link breakage segment, a phase evolution trend benchmark that characterizes the phase evolution law in the frequency domain is constructed. The phase evolution of the flange micro-leakage signal carries physical information about the location of the leakage source and the propagation path. The continuous evolution law of the phase is an inviolable physical constraint for loss compensation.

[0111] Using the manifold distribution of adjacent intact data in a high-dimensional feature space as geometric constraints and the phase evolution trend benchmark as tangent vector constraints, the missing node trajectory is deduced along the low-dimensional manifold in the manifold space. Then, the link break compensation data segment is generated through inverse mapping and adaptive interpolation fitting. This method considers the manifold structure correlation within the feature vector and the physical evolution law of the frequency domain phase. Compared with simple linear interpolation or spline interpolation, it can more accurately restore the real feature change pattern of the leakage signal during the link break. Especially when the leakage state changes dynamically, the compensation data can faithfully reflect the transient transition characteristics of the feature parameters. The compensation data is seamlessly embedded into the original sequence and a full-dimensional reconstruction verification is performed. The output is a full-dimensional state feature stream with continuous time index, complete feature dimensions, and consistent physical constraints. This effectively avoids the problem of missed or false alarms caused by transmission link breakage and improves the robustness and monitoring reliability of the system in harsh industrial wireless communication environments.

[0112] In a preferred embodiment of the present invention, step 6 above may include:

[0113] Step 61: Perform a weighted average of the full-dimensional state feature flow along the channel dimension to obtain a one-dimensional feature sequence; perform phase space coordinate transformation on the one-dimensional feature sequence according to the preset embedding dimension and time delay parameters to reconstruct a multi-dimensional trajectory point cloud representing the dynamic evolution of the flange sealing state, and construct a multi-dimensional degenerate phase space, specifically including: taking the full-dimensional state feature flow, the tensor containing all Each sensor channel in all A time window The feature vectors are fused in a dimensional manner by weighting the feature vectors of each channel at each time step along the channel dimension. The weights are taken from the normalized signal-to-noise ratio estimates of each channel, resulting in a one-dimensional feature sequence. , The phase space reconstruction is performed on the one-dimensional feature sequence. According to the preset embedding dimension and time delay, the scalar time series is mapped to a set of trajectory points in the 3D phase space in units of time window steps. The reconstructed 1st... The spatial trajectory points are represented as follows:

[0114] ;

[0115] In the formula, For a moment The corresponding phase space trajectory points; This represents the value of the one-dimensional feature sequence after channel-weighted averaging at the corresponding time index. The preset time delay steps are determined based on the first minimum value criterion of the mutual information method. The embedding dimension is predetermined and determined using the pseudo-nearest neighbor method, which theoretically satisfies... ,in For the attractor fractal dimension of the sealed dynamic system, for all effective time indices Perform the above reconstruction to obtain the point set The point cloud of trajectory in the multidimensional degenerate phase space is formed. The geometric distribution of this point cloud in the phase space directly encodes the complete dynamic evolution information of the flange sealing state from health, micro-leakage initiation, leakage development to critical failure. The migration direction and step size between adjacent trajectory points reflect the rate and magnitude of the change in sealing state.

[0116] Step 62: Using the multidimensional trajectory point cloud as the fitting benchmark, set the state transition rules and covariance update threshold, perform model parameter initialization and baseline convergence calculation, construct a dynamic baseline drift tracking model, and output the benchmark evolution path representing the sealing health state. Specifically, this includes: processing the multidimensional trajectory point cloud... Using observational data, a dynamic baseline drift tracking model is constructed. The model consists of two parts: a state equation and an observation equation. A recursive Bayesian filtering framework is employed to achieve online tracking of baseline drift in the sealed state. The state vector is defined as... The baseline of the sealed health status in the corresponding phase space is in the 1st phase. The position coordinates at each recursive time step; the observation vector takes the newly arrived trajectory point at the current time step. The dynamic baseline drift tracking model is implemented using the Extended Kalman Filter (EKF) framework, and its state equation and observation equation are as follows:

[0117] ;

[0118] ;

[0119] in, For the first The state vector at time t corresponds to the position coordinates of the sealed healthy state baseline in phase space; This is the state transition matrix, which characterizes the slowly varying properties of baseline drift; This is the process noise vector; For the first The observation vector at time t is the newly arrived phase space trajectory point at the current time; The observation matrix is ​​taken as the identity matrix. ; To observe the noise vector, when a new observation arrives, a prediction-update iteration is performed: the prediction step derives a prior estimate of the current baseline position based on the state transition equation. with prior covariance ; Update step to calculate new information The covariance of the new information is used to determine whether the current observation contains a significant anomalous component deviating from the baseline. If the trace of the new information covariance exceeds a preset threshold, the current observation is considered to contain a significantly anomalous component deviating from the baseline. In this case, the observation noise covariance is increased to 10 times its original value, and the confidence weight of the observation noise covariance is reduced, making the model more inclined to maintain the original baseline trend rather than follow anomalous jumps. This threshold is determined based on the statistical distribution of phase space trajectory points during the healthy operation phase, taking the mean of the information covariance trace during the healthy segment plus... Double the standard deviation.

[0120] During the model initialization phase, take the previous value. Using a set of continuous trajectory points as the initial training segment, the mean vector of this segment is calculated as the initial state value, and the covariance matrix of this segment is calculated as the initial state covariance. The observation noise covariance matrix is ​​estimated by the squared average deviation of the segment from its mean vector. After initialization, the parameters are calculated along the time axis for all... The trajectory points are traversed one by one to perform recursive tracing, where The preset time delay steps are determined based on the first minimum value criterion of the mutual information method. The embedding dimension is predetermined and determined using the pseudo-nearest neighbor method, which theoretically satisfies... ,in Let be the attractor fractal dimension of the sealed dynamic system. After a sufficient number of iterations, the state estimation sequence converges into a smooth phase space trajectory, which is the baseline evolution path. ,in For a moment The corresponding phase space trajectory points describe the long-term, slow drift of the flange seal state under conditions of no leakage or only permissible sealing aging.

[0121] Step 63: Using the dynamic baseline drift tracking model and the baseline evolution path as tracking references, calculate in real time the radial deviation distance and tangential slip rate of the multidimensional trajectory point cloud relative to the baseline evolution path to obtain the dynamic baseline drift offset sequence. Specifically, this includes: for the trajectory point cloud... Each trajectory point in In the baseline evolution path The corresponding reference point is determined by searching along the reference evolution path. The benchmark point with the nearest time index and whose state estimation has converged. , track points With reference point Vector difference between Orthogonal decomposition is performed along the local tangent and normal directions of the reference path, and the reference path is in The local tangent vector at a given point is obtained by differentiating the centers of the preceding and following reference points; all directions orthogonal to the tangent vector span a normal subspace. Projecting onto the normal subspace and taking the magnitude of the projection vector yields the radial offset distance. This value quantifies the instantaneous deviation of the current sealing state from the healthy baseline. The first-order forward difference of the tangential component with respect to time is obtained to obtain the tangential slip rate, which describes the speed of the sealing state's advancement along the degradation direction. By traversing all trajectory points, a dynamic baseline drift offset sequence composed of the radial deviation distance sequence and the tangential slip rate sequence is obtained.

[0122] Step 64: Perform state degradation accumulation calculation on the dynamic baseline drift offset sequence, and perform weighted integration processing in conjunction with the creep attenuation coefficient of the sealing interface material to obtain the flange seal failure evolution index characterizing the degree of sealing performance degradation. Specifically, this includes: taking the radial deviation distance sequence and the tangential slip rate sequence as input, performing state degradation accumulation calculation on the time axis, and weighting the radial deviation distance and tangential slip rate on a time-by-time basis to synthesize the instantaneous degradation increment. During the synthesis, the creep model parameters of the sealing interface material are introduced. These parameters are material constants related to the gasket material, service temperature, and preload stress level, and are determined by fitting laboratory accelerated aging tests or historical maintenance data from the field. The instantaneous degradation increment is calculated using the following formula:

[0123] ;

[0124] In the formula, For the first The instantaneous degradation increment within a time window; This represents the radial deviation distance at that moment; Let be the tangential slip rate at that moment; The step size for adjacent time windows; The creep proportionality coefficient is determined by both the gasket material properties and the service temperature, and its unit is matched with the stress index. The equivalent contact stress at the sealing interface is calculated and determined by the bolt preload and the contact area of ​​the gasket, and the unit is megapascals. The stress index reflects the sensitivity of creep rate to stress; the typical value range for non-metallic gaskets is 2. 5； It is a time exponent that reflects the evolution of creep rate over service time. Its typical value range is 0.2 to 0.8, which conforms to the gradual decay characteristics of the Norton-Bailey creep model. The cumulative time, expressed in hours, is calculated from the start of the seal's service life. The first term reflects the cumulative degradation caused by radial deviation during creep, as the material deteriorates. The creep rate varies slowly with service time using a power law and does not increase unbounded. The second term reflects the direct accumulation caused by degradation advancing along the reference direction. By integrating and summing the instantaneous degradation increments over all time windows along the time axis, the flange seal failure evolution index is obtained.

[0125] ;

[0126] In the formula, The flange seal failure evolution index is a dimensionless scalar. The starting index of the valid trajectory points; Index for the latest time window. The value increases monotonically with increasing seal degradation, and the rate of increase gradually flattens with service time, consistent with the physical law of gradual saturation of sealing materials due to creep. The preset threshold range can divide the flange sealing status into four levels: healthy, micro-leakage warning, leakage development, and critical failure, which can directly drive on-site audible and visual alarms or link with the DCS / SCADA system to execute preventive maintenance decisions.

[0127] In this embodiment of the invention, traditional threshold alarms cannot detect progressive degradation, and the seal is often close to failure when an alarm is triggered. By reconstructing the phase space, the temporal features are expanded to a high-dimensional phase space, and the degradation law is represented by the geometric structure of the attractor trajectory. Weak drift is manifested as radial displacement, and the early sensitivity is better than the time domain amplitude detection. The dynamic baseline drift tracking model tracks the slowly changing baseline with recursive Bayesian filtering, and suppresses abnormal jump contamination by the innovation covariance threshold, solving the problem of false alarms due to the migration of fixed baselines due to operating conditions. The radial deviation and tangential slip rate decouple the degradation amplitude dimension and velocity dimension information. The creep decay coefficient of the sealing material is introduced to perform time-weighted integration of the deviation, so that the same deviation contributes more damage in the later stage of service, which is consistent with the reality of creep-accelerated degradation. Overall, without adding sensors, the full-dimensional feature flow is transformed into a quantifiable, traceable, and predictable scalar failure evolution index.

[0128] In a preferred embodiment of the present invention, step 6 above may include:

[0129] Step 65: Extract the temporal evolution trajectory of the flange seal failure evolution index in the multidimensional degradation phase space. Map the static critical value boundary and dynamic rate of change tolerance interval in the hierarchical early warning threshold matrix to the multidimensional degradation phase space to construct a multi-level geometric early warning boundary surface. Specifically, this includes: taking the flange seal failure evolution index sequence and its corresponding time index, and tracing back the phase space trajectory points corresponding to each time index. , for each trajectory point 3D coordinates and the current Scalar value correlation yields temporal evolution trajectories labeled with the degree of degradation. The trajectory is in Traversing in phase space in chronological order, the geometric trajectory encodes the entire process of the sealing state from healthy to deteriorating. A graded early warning threshold matrix is ​​introduced. To construct this matrix, historical failure evolution index data of flange seals under four conditions—normal operation, micro-leakage, leakage development, and critical failure—are collected. The 95th percentile of the failure evolution index for each category is calculated as the static critical value boundary. The ±3σ interval of the rate of change of each index is calculated as the dynamic rate of change tolerance interval. Simultaneously, through accelerated aging tests in the laboratory and on-site leakage simulation tests, the upper limits of normal penetration depth, tangential slip rate, and minimum residence confirmation time corresponding to each level of early warning are determined. These parameters together constitute the graded early warning threshold matrix, with each row corresponding to a preset threshold. There are several warning levels, corresponding to two dimensions: static critical value boundaries and dynamic rate of change tolerance intervals. Among them, the first... The static critical value boundary record corresponding to the Level 1 warning indicates that the failure evolution index is triggered when it reaches this threshold. Level 1 warning; The dynamic change rate tolerance interval corresponding to the Level 1 warning is denoted as This indicates that when the rate of change of the failure evolution index within a unit of time falls into this range, the early warning criterion needs to be adjusted accordingly. The boundaries of static critical values ​​at each level are mapped to a multidimensional degradation phase space to construct... There are nested isotropic hypersurfaces, each defined as the envelope of the set of trajectory points in phase space that satisfy the failure evolution exponent being equal to the corresponding static critical value. The boundary surface of the level-one geometric early warning system is defined as:

[0130] ;

[0131] In the formula, For the first The boundary surface corresponding to the level-one early warning; Let be any point in phase space; This is a mapping function from phase space coordinates to the failure evolution index. This mapping is implicitly determined by the complete computational link, and in actual calculations, it is obtained by applying existing trajectory points. Approximation can be achieved through multidimensional interpolation or locally weighted regression. For the first Level static critical value, the boundary surfaces of each level are nested from the inside to the outside in phase space, the trajectory points of the healthy state are gathered in the innermost surface, when the trajectory points cross outwards Upon entering the Level 1 warning zone, proceed through the following areas in sequence. Corresponding to the progressively escalating warning levels, the dynamic change rate tolerance range The allowable velocity bands in the normal direction of the surface are superimposed on the boundary surfaces of each level. When the trajectory crosses the surface along the normal direction at a speed exceeding this allowable range, even if... Even if the next level of static threshold is not reached, a corresponding level of warning is triggered, resulting in a multi-level geometric warning boundary surface system that integrates static boundaries and dynamic rate constraints.

[0132] Step 66: Discretize the multi-level geometric early warning boundary surface into a spatial grid node array. Calculate the principal curvature tensor and local Gaussian curvature at each node based on the three-dimensional coordinate difference relationship between adjacent grid nodes, and construct the surface curvature gradient distribution field. Project the spatial coordinate sequence of the temporal evolution trajectory onto the surface curvature gradient distribution field, identify the discrete intersection points where the temporal evolution trajectory crosses the curvature abrupt change critical region, and output the initial penetration node set. Specifically, this includes: for each level of the geometric early warning boundary surface... Perform spatial discretization, in upper edge Each dimension of the phase space is divided into uniform grids with a preset resolution, transforming the continuous surface into a shape composed of... A discrete node array consisting of grid nodes, where each node stores its position in phase space. 3D coordinates, due to Usually, 3 or 4 is chosen, when... The time-varying surface degenerates into a two-dimensional manifold embedded in three-dimensional space. For the discretized mesh nodes, the local differential geometric quantities at each node are calculated based on the coordinate difference relationship between adjacent nodes. (The last part, "the node...", appears to be incomplete and requires further context.) and For nearest-neighbor mesh nodes, the first and second fundamental form coefficients at that node are obtained through local quadratic surface fitting or by directly utilizing the finite difference between adjacent nodes. This allows for the solution of the principal curvature tensor, whose two eigenvalues ​​constitute the first principal curvature. Second principal curvature Calculate local Gaussian curvature from principal curvature and mean curvature The above are all well-known fundamental relations in the field of differential geometry, applicable to all... After calculating each node one by one, the surface curvature scalar field is obtained. The gradient of the curvature scalar field is calculated along the two orthogonal directions of the tangent plane of the surface to obtain the surface curvature gradient distribution field. This field describes the degree of curvature change at each position on the warning boundary surface. Regions with large curvature gradient magnitudes constitute the curvature abrupt change critical region. These regions usually correspond to folds, saddle points or curvature sign reversals on the boundary surface and are geometric indicators of accelerated deterioration of the sealed state in the multidimensional degenerative phase space.

[0133] Connecting adjacent trajectory points at each level in the temporal evolution trajectory into spatial polyline segments, projecting them onto the surface curvature gradient distribution field, and then detecting their relationship with each segment along the polyline segment. Geometric intersection relationship: If the two endpoints of a broken line segment are located at... The inner and outer sides are then used to determine whether the broken line segment is related to the inside or outside of the line. Intersections exist; the precise spatial coordinates of these intersections are determined through linear interpolation. Among the resulting intersections, discrete intersections located within the critical region of curvature abrupt change are selected. These intersections have higher warning sensitivity than those crossing regions with gentle curvature, because crossing a high-curvature region signifies that the seal degradation has entered a dynamically unstable and accelerating phase space region. The spatial coordinates and warning levels of all selected discrete intersections are listed below. The initial set of penetration nodes is obtained by taking the corresponding timestamp and the curvature gradient magnitude at that location.

[0134] Step 67: Using the initial penetration node set as the calculation benchmark, perform spatial line segment fitting on adjacent nodes along the time axis, calculate the normal penetration depth and tangential slip rate of each fitted line segment beyond the multi-level geometric warning boundary surface, and combine the continuous residence time of discrete intersection points within the time window to obtain the boundary penetration feature set. Specifically, this includes: arranging the discrete intersection points in the initial penetration node set in ascending order of timestamps, and for those belonging to the same warning level... Furthermore, for two temporally adjacent penetration nodes, a spatial line segment is constructed in phase space connecting them. This line segment reflects the local direction and distance of the sealed trajectory during its crossing of the warning boundary. For the line segment extending beyond the boundary surface... For the outer portion, calculate the normal penetration depth. Take the distance on the line segment. Find the farthest point, and calculate its distance to the surface from that point. The shortest Euclidean distance is the maximum normal penetration depth of the penetration event, which is determined by the following formula:

[0135] ;

[0136] In the formula, The normal penetration depth; The spatial fitting line segment connecting two adjacent penetrating nodes in time, i.e., the set of all points on the line segment; To fit points on the line segment; No. Level geometric early warning boundary surface any point on; For the first Level geometric early warning boundary surface; The distance is Euclidean. In practice, this maximum value is approximated by discretizing the line segment into finely sampled points and calculating the nearest distance to the surface mesh node point by point. The tangential slip rate is taken as the projection length of the line segment on the local tangential plane of the surface divided by the time difference between the two endpoints of the line segment, reflecting how fast the trajectory escapes tangentially along the boundary surface. For each penetration node, the trajectory is statistically analyzed during the crossing... The continuous dwell time outside the boundary surface is recorded. If the trajectory enters and exits the same level boundary surface multiple times, the independent dwell time of each penetration is recorded. The longer the dwell time, the longer the sealing state stays in the danger zone corresponding to the warning level, and the more sufficient the degradation accumulation. The normal penetration depth, tangential slip rate, continuous dwell time and the warning level of each penetration event are combined into a boundary penetration feature vector. The feature vectors of all penetration events constitute the boundary penetration feature set.

[0137] Step 68: Based on the normal penetration depth, tangential slip rate, and continuous residence time in the boundary penetration feature set, match the preset warning level mapping rules in the graded warning threshold matrix to determine the warning trigger level of the current flange sealing status; encapsulate the warning trigger level and leakage risk level identifier to obtain the real-time monitoring and warning results. Specifically, this includes: acquiring the boundary penetration feature set, performing multi-condition joint judgment on the feature vector of each penetration event, and in addition to the static critical value boundary and dynamic change rate tolerance range, the graded warning threshold matrix also presets three judgment parameters corresponding to each level of warning: upper limit of normal penetration depth, upper limit of tangential slip rate, and minimum residence confirmation time. The determination of the warning trigger level follows a priority progression rule, checking the current penetration event. If the normal penetration depth exceeds the upper limit of the normal penetration depth preset for a certain level of warning, and if it does, then that level is designated as a pending trigger level. If it exceeds the upper limits of multiple levels simultaneously, then the highest level is selected. If the tangential slip rate of the current penetration event exceeds the upper limit of the tangential slip rate corresponding to the pending level, and if it does, then the pending level is maintained. If it exceeds by a large margin, then it can be upgraded by one level. If the duration of the current penetration event reaches the minimum duration of the pending level, then the warning is confirmed and the pending level is output as the final warning trigger level. If the duration is insufficient, then it is considered as an instantaneous disturbance where the sealed trajectory briefly pierces the boundary surface and then immediately retreats, and is filtered out. No warning is triggered, and the final warning trigger level is recorded as zero.

[0138] The above three conditions are executed according to a hierarchical logic, prioritizing normal penetration depth, followed by tangential slip rate, and finally confirming the dwell time. This ensures that the early warning judgment captures both the spatial amplitude of the penetration boundary surface of the seal degradation and the dynamic speed of the penetration behavior. False penetrations are eliminated through dwell time. All three are indispensable. The final early warning trigger level is mapped to a leakage risk level identifier: Level 0 corresponds to normal; Level 1 corresponds to Level 1 warning, suspected micro-leakage; Level 2 corresponds to Level 2 warning, leakage is developing; Level 3 corresponds to Level 3 warning, serious leakage risk; and so on up to the preset highest early warning level. The early warning trigger level, leakage risk level identifier, trigger time, current failure evolution index value, and corresponding normal penetration depth, tangential slip rate, and continuous dwell time are encapsulated in a structured data format to generate a standard format real-time monitoring and early warning result message. This message is pushed to the host computer in the central control room and the mobile terminal of maintenance personnel via industrial fieldbus or wireless backhaul link, driving audible and visual alarms, linking emergency shut-off valves, or triggering preventive maintenance work orders.

[0139] In this embodiment of the invention, traditional early warning methods rely solely on single threshold comparisons, making it difficult to perceive the dynamic meaning of degradation trajectories traversing different curvature regions in phase space. By embedding the failure evolution index into a multi-dimensional degradation phase space and constructing multi-level early warning boundaries in the form of nested equivalent hypersurfaces, the early warning thresholds at each level gain a geometric carrier. By calculating the Gaussian curvature and curvature gradient field of the boundary surfaces, the critical regions of curvature abrupt changes are identified. These regions correspond to the phase transition points of sealing degradation dynamics. When the trajectory traverses these regions, the early warning sensitivity is enhanced. The penetration behavior is decoupled into three independent dimensions: normal penetration depth, tangential slip rate, and dwell time. Combined with multi-condition joint judgment, instantaneous disturbances are filtered out, avoiding false alarms caused by operational shocks. Overall, without the need for additional sensors, the scalar threshold is upgraded to a surface geometric penetration criterion system, achieving precise spatiotemporal triggering of graded early warnings.

[0140] like Figure 2 As shown, embodiments of the present invention also provide a real-time monitoring and early warning system for Internet of Things (IoT) devices, comprising:

[0141] The acquisition module is used to acquire differential pressure monitoring data output by the micro differential pressure sensor array deployed on the sidewalls of the upstream and downstream pipe sections of the high-pressure natural gas transmission pipeline flange, as well as strain monitoring data output by the fiber optic strain sensing node attached to the surface of the flange connection bolt group.

[0142] The fusion module is used to fuse differential pressure monitoring data and strain monitoring data in a time series manner to construct the original time series monitoring data; cross-node clock drift compensation and multi-channel sampling rate remapping processing are performed on the original time series monitoring data to obtain a spatiotemporally aligned time series dataset;

[0143] A module is built to perform frequency band orthogonal decomposition on a spatiotemporally aligned time series dataset to obtain the low-frequency drift component of environmental thermal stress disturbance and the high-frequency pulsation component of structural micro-vibration transmission; an interference suppression mask is built based on the low-frequency drift component and the high-frequency pulsation component.

[0144] The calculation module is used to apply the interference suppression mask to the spatiotemporally aligned time series dataset to perform background noise filtering, obtain the leakage sensitive component, and calculate the frequency band energy distribution spectrum and transient amplitude trajectory of the leakage sensitive component to obtain the feature decoupling monitoring sequence.

[0145] The identification module is used to monitor the integrity identifier of the data packets of the decoupled monitoring sequence on the wireless convergence link. When a data break segment is identified, adaptive trajectory interpolation compensation is performed on the break segment through the phase evolution law of the frequency band energy distribution spectrum to obtain a time-continuous full-dimensional state feature stream.

[0146] The judgment module is used to map the full-dimensional state feature flow to the multi-dimensional degenerate phase space, construct a dynamic baseline drift tracking model to calculate the flange seal failure evolution index; and perform boundary penetration judgment between the flange seal failure evolution index and the preset graded early warning threshold matrix to obtain real-time monitoring and early warning results.

[0147] It should be noted that this system is a system corresponding to the above method. All implementation methods in the above method embodiments are applicable to this embodiment and can achieve the same technical effect.

[0148] Embodiments of the present invention also provide a computing device, including: a processor and a memory storing a computer program, wherein the computer program, when executed by the processor, performs the method described above. All implementations in the above method embodiments are applicable to this embodiment and can achieve the same technical effects.

[0149] Embodiments of the present invention also provide a computer-readable storage medium storing instructions that, when executed on a computer, cause the computer to perform the method described above. All implementations in the above method embodiments are applicable to this embodiment and can achieve the same technical effects.

[0150] The above description represents the preferred embodiments of the present invention. It should be noted that those skilled in the art can make various improvements and modifications without departing from the principles of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.

Claims

1. A method for real-time monitoring and early warning of Internet of Things (IoT) devices, characterized in that, The method includes: Differential pressure monitoring data is collected from the micro differential pressure sensor array deployed on the sidewalls of the upstream and downstream pipe sections of the high-pressure natural gas transmission pipeline flange, as well as strain monitoring data is collected from the fiber optic strain sensing node attached to the surface of the flange connection bolt group. Differential pressure monitoring data and strain monitoring data are fused in time series to construct the original time series monitoring data; cross-node clock drift compensation and multi-channel sampling rate remapping processing are performed on the original time series monitoring data to obtain a spatiotemporally aligned time series dataset; Frequency band orthogonal decomposition is performed on the spatiotemporally aligned time series dataset to obtain the low-frequency drift component of environmental thermal stress disturbance and the high-frequency pulsation component of structural micro-vibration transmission; an interference suppression mask is constructed based on the low-frequency drift component and the high-frequency pulsation component. An interference suppression mask is applied to a spatiotemporally aligned time series dataset to perform background noise filtering, thereby obtaining the leakage sensitive component. The frequency band energy distribution spectrum and transient amplitude trajectory of the leakage sensitive component are then calculated to obtain the feature decoupling monitoring sequence. The monitoring features decouple the monitoring sequence from the data packet integrity identifier on the wireless convergence link. When a data break segment is identified, adaptive trajectory interpolation compensation is performed on the break segment through the phase evolution law of the frequency band energy distribution spectrum to obtain a time-continuous full-dimensional state feature stream. The full-dimensional state feature flow is mapped to a multi-dimensional degenerate phase space to construct a dynamic baseline drift tracking model to calculate the flange seal failure evolution index; the flange seal failure evolution index is then compared with a preset graded early warning threshold matrix to determine the boundary penetration and obtain real-time monitoring and early warning results.

2. The method for real-time monitoring and early warning of IoT devices according to claim 1, characterized in that, Differential pressure monitoring data and strain monitoring data are fused together in time series to construct the original time series monitoring data; Cross-node clock drift compensation and multi-channel sampling rate remapping are performed on the raw time-series monitoring data to obtain a spatiotemporally aligned time-series dataset, including: Extract the original acquisition timestamps carried by the differential pressure monitoring data and strain monitoring data respectively, and calculate the initial time offset of each sensing node relative to the reference clock using a preset reference clock as a reference. Based on the initial time offset, a time axis alignment parameter is constructed. The time axis alignment parameter is applied to the differential pressure monitoring data and the strain monitoring data. Sliding interpolation alignment is performed along a unified time grid. The aligned data is then spliced ​​and fused in chronological order to construct the original time-series monitoring data. Using the original time-series monitoring data as the processing object, the actual time interval sequence of each data channel within the continuous sampling period is extracted, the cumulative clock drift error between the actual time interval sequence and the standard sampling period is calculated, and a clock drift compensation sequence is generated based on the cumulative clock drift error. The clock drift compensation sequence is applied to the original timing monitoring data to perform time axis translation and proportional stretching correction, and clock synchronization data is output. The clock synchronization data is then subjected to equal-interval resampling and physical dimension normalization to obtain a spatiotemporally aligned timing dataset.

3. The method for real-time monitoring and early warning of IoT devices according to claim 2, characterized in that, Frequency band orthogonal decomposition was performed on the spatiotemporally aligned time series dataset to obtain the low-frequency drift component of environmental thermal stress disturbance and the high-frequency pulsation component of structural micro-vibration transmission. An interference suppression mask is constructed based on low-frequency drift components and high-frequency pulsation components, including: Orthogonal frequency band decomposition processing is performed on the spatiotemporally aligned time series dataset. Multi-level orthogonal decomposition operations are performed along the frequency axis to divide the spatiotemporally aligned time series dataset into several independent frequency band sub-signal sequences. Using several independent frequency band sub-signal sequences as the screening objects, sub-signal segments with frequency ranges within the preset thermal stress gradual change range are extracted as low-frequency drift components, and sub-signal segments with frequency ranges within the preset mechanical vibration pulse range are extracted as high-frequency pulsation components. Instantaneous energy density calculations were performed on the low-frequency drift component and the high-frequency pulsation component respectively to obtain the low-frequency energy distribution sequence and the high-frequency energy distribution sequence; Using the low-frequency energy distribution sequence and the high-frequency energy distribution sequence as mapping benchmarks, corresponding signal attenuation weights are assigned according to the degree to which the energy proportion of each frequency band exceeds the preset interference threshold, resulting in a frequency domain filtering weight set containing dynamic attenuation coefficients for multiple frequency bands. By binding and mapping the frequency domain filter weight set with the spatial distribution characteristics of the sensor, an interference suppression mask is constructed.

4. The method for real-time monitoring and early warning of IoT devices according to claim 3, characterized in that, An interference suppression mask is applied to a spatiotemporally aligned time-series dataset to perform background noise filtering, obtaining the leakage-sensitive component. The frequency band energy distribution spectrum and transient amplitude trajectory of the leakage-sensitive component are then calculated to obtain a characteristic decoupling monitoring sequence, including: Extract the multi-band dynamic attenuation coefficient from the interference suppression mask, perform a frequency-point weighted multiplication operation on the multi-band dynamic attenuation coefficient and the frequency domain transformation result of the spatiotemporally aligned time series dataset, perform background interference signal frequency band suppression processing, filter out the spectral characteristics of low-frequency drift components and high-frequency pulsation components, and output the residual time series signal after background purification. Using the residual time-series signal after background purification as the processing object, time-frequency domain inverse mapping reconstruction is performed to remove the residual background baseline noise and extract the leakage sensitive component characterizing the micro-deformation of the flange sealing interface. Input the leakage sensitive component into the sliding frequency domain analysis window, perform frequency band energy integration along the time axis, count the cumulative energy value in each target frequency band interval, and plot the frequency band energy distribution spectrum of the leakage sensitive component; Based on the time-domain waveform characteristics of the leakage sensitive component, the time interval and amplitude change gradient of adjacent extreme points are extracted to construct the signal envelope evolution path. The peak fluctuation law of the signal envelope evolution path is tracked along the time axis to extract the transient amplitude trajectory. The frequency domain feature vector of the frequency band energy distribution spectrum and the time domain feature vector of the transient amplitude trajectory are synchronously aligned and spliced ​​according to the time index to obtain the feature decoupling monitoring sequence.

5. The method for real-time monitoring and early warning of IoT devices according to claim 4, characterized in that, The monitoring feature decoupling monitoring sequence identifies the data packet integrity on the wireless convergence link. When a data link break segment is identified, adaptive trajectory interpolation compensation is performed on the break segment based on the phase evolution law of the frequency band energy distribution spectrum, resulting in a time-continuous, full-dimensional state feature stream, including: Read the data packet sequence number and verification identifier carried by the feature decoupling monitoring sequence during the wireless convergence link transmission to obtain the link transmission status mark stream; based on the link transmission status mark stream, locate the missing intervals where the verification identifier is missing and the sequence number undergoes discontinuous jumps, and extract the missing intervals as data link break segments; The frequency band energy distribution spectrum within a preset time window before and after the data link break segment is extracted, and the slope and curvature characteristics of the phase angle of each frequency band changing with time are extracted to construct a phase evolution trend benchmark. Using the phase evolution trend benchmark as the topological constraint, manifold trajectory deduction and adaptive interpolation fitting are performed on the missing data nodes in the data chain breakage segment to obtain the chain breakage compensation time series data segment. The time series data segment with broken link compensation is seamlessly embedded into the corresponding position of the original time axis of the feature decoupling monitoring sequence, the broken link segment is replaced, the full-dimensional feature vector is recombined, and the time series continuous full-dimensional state feature flow is output.

6. The method for real-time monitoring and early warning of IoT devices according to claim 5, characterized in that, By mapping the full-dimensional state feature flow to a multi-dimensional degenerate phase space, a dynamic baseline drift tracking model is constructed to calculate the flange seal failure evolution index, including: A weighted average is applied to the full-dimensional state feature flow along the channel dimension to obtain a one-dimensional feature sequence. The one-dimensional feature sequence is then transformed into phase space coordinates according to the preset embedding dimension and time delay parameters to reconstruct a multi-dimensional trajectory point cloud that characterizes the dynamic evolution of the flange sealing state, and a multi-dimensional degenerate phase space is constructed. Using multidimensional trajectory point cloud as the fitting benchmark, state transition rules and covariance update threshold are set, model parameter initialization and baseline convergence operation are performed, dynamic baseline drift tracking model is constructed, and benchmark evolution path representing the sealed health state is output. Using the dynamic baseline drift tracking model and the baseline evolution path as tracking references, the radial deviation distance and tangential slip rate of the multidimensional trajectory point cloud relative to the baseline evolution path are calculated in real time to obtain the dynamic baseline drift offset sequence. A state degradation accumulation operation is performed on the dynamic baseline drift offset sequence, and a weighted integral is performed in combination with the creep attenuation coefficient of the sealing interface material to obtain the flange seal failure evolution index, which characterizes the degree of sealing performance degradation.

7. The method for real-time monitoring and early warning of IoT devices according to claim 6, characterized in that, The flange seal failure evolution index is compared with a preset graded early warning threshold matrix to determine boundary penetration, resulting in real-time monitoring and early warning results, including: Extract the temporal evolution trajectory of the flange seal failure evolution index in the multidimensional degradation phase space, map the static critical value boundary and dynamic change rate tolerance interval in the graded early warning threshold matrix to the multidimensional degradation phase space, and construct a multi-level geometric early warning boundary surface; The multi-level geometric early warning boundary surface is discretized into a spatial grid node array. Based on the three-dimensional coordinate difference relationship between adjacent grid nodes, the principal curvature tensor and local Gaussian curvature at each node are calculated to construct the surface curvature gradient distribution field. The spatial coordinate sequence of the temporal evolution trajectory is projected onto the surface curvature gradient distribution field to identify the discrete intersection points of the temporal evolution trajectory crossing the curvature change critical region and output the initial penetration node set. Using the initial set of penetration nodes as the calculation benchmark, spatial line segment fitting is performed on adjacent nodes along the time axis. The normal penetration depth and tangential slip rate of each fitted line segment beyond the multi-level geometric warning boundary surface are calculated. Combined with the continuous dwell time of discrete intersection points within the time window, the boundary penetration feature set is obtained. Based on the normal penetration depth, tangential slip rate, and duration of continuous residence in the boundary penetration feature set, the warning level mapping rules in the hierarchical warning threshold matrix are matched to determine the warning trigger level of the current flange sealing status; the warning trigger level and leakage risk level identifier are encapsulated to obtain real-time monitoring and warning results.

8. A real-time monitoring and early warning system for Internet of Things (IoT) devices, wherein the system implements the method as described in any one of claims 1 to 7, characterized in that, include: The acquisition module is used to acquire differential pressure monitoring data output by the micro differential pressure sensor array deployed on the sidewalls of the upstream and downstream pipe sections of the high-pressure natural gas transmission pipeline flange, as well as strain monitoring data output by the fiber optic strain sensing node attached to the surface of the flange connection bolt group. The fusion module is used to fuse differential pressure monitoring data and strain monitoring data in a time series manner to construct the original time series monitoring data; Perform cross-node clock drift compensation and multi-channel sampling rate remapping processing on the original time series monitoring data to obtain a spatiotemporally aligned time series dataset; The module is used to perform frequency band orthogonal decomposition on the spatiotemporally aligned time series dataset to obtain the low-frequency drift component of environmental thermal stress disturbance and the high-frequency pulsation component of structural micro-vibration transmission. An interference suppression mask is constructed based on low-frequency drift components and high-frequency pulsation components; The calculation module is used to apply the interference suppression mask to the spatiotemporally aligned time series dataset to perform background noise filtering, obtain the leakage sensitive component, and calculate the frequency band energy distribution spectrum and transient amplitude trajectory of the leakage sensitive component to obtain the feature decoupling monitoring sequence. The identification module is used to monitor the integrity identifier of the data packets of the decoupled monitoring sequence on the wireless convergence link. When a data break segment is identified, adaptive trajectory interpolation compensation is performed on the break segment through the phase evolution law of the frequency band energy distribution spectrum to obtain a time-continuous full-dimensional state feature stream. The determination module is used to map the full-dimensional state feature flow to the multi-dimensional degenerative phase space and construct a dynamic baseline drift tracking model to calculate the flange seal failure evolution index. The flange seal failure evolution index is compared with the preset graded early warning threshold matrix to determine the boundary penetration and obtain real-time monitoring and early warning results.

9. A computing device, characterized in that, include: One or more processors; A storage device for storing one or more programs, which, when executed by one or more processors, cause the one or more processors to implement the method as described in any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a program that, when executed by a processor, implements the method as described in any one of claims 1 to 7.

Images