Intelligent sensor-based valve leakage intelligent monitoring method

Abstract

This invention discloses an intelligent valve leakage monitoring method based on smart sensors. The method includes: collecting and preprocessing multi-source valve state data to generate a standardized valve state dataset; filtering stable operating windows to generate normal sealing baseline data, and generating valve state residual data by difference between the baseline data and the current monitoring window data; generating periodic state patches and candidate leakage event patches based on the valve state residual data; inputting the two types of patches into an improved TimeXer model to generate internal leakage state representation, external leakage state representation, and initial equivalent leakage orifice diameter; using a dung beetle optimizer combined with the valve topology manifold and leakage severity momentum to optimize the model parameters; re-inputting the two types of patches, and generating valve leakage monitoring results through a continuous orifice diameter readout layer. This invention achieves accurate identification, path location, and orifice diameter estimation of valve leaks, improving monitoring real-time performance and early warning reliability.

Description

Technical Field

[0001] This invention relates to the field of intelligent monitoring technology for industrial equipment, and in particular to an intelligent valve leakage monitoring method based on intelligent sensors. Background Technology

[0002] Valve leakage monitoring involves determining the operational status of the valve inlet pipe section, valve outlet pipe section, valve core sealing area, valve seat annular gap area, valve stem packing area, valve cover connection area, flange connection area, and external environment area. Existing technologies for valve leakage identification primarily employ pressure sensors, flow sensors, acoustic emission sensors, vibration sensors, temperature sensors, and concentration sensors to collect valve status data. This data is then preprocessed, feature extracted, anomaly identified, and leakage type determined using methods such as fixed threshold judgment, rule matching, ordinary time-series prediction models, convolutional neural networks, recurrent neural networks, and conventional classification networks. Alarm information is then output based on the classification results.

[0003] Current technologies still have shortcomings. Multi-source valve status data is often directly input into classification models, lacking processes for filtering stable operating windows, generating normal sealing baseline data, and constructing valve status residual data. This results in insufficient ability to distinguish between operating condition fluctuations, sensor drift, and early minor leaks. The anomaly identification process often segments data according to fixed time windows, lacking processes for extracting candidate anomaly anchor points, dynamic event time warping, and generating candidate leakage event patches. This leads to insufficient representation of the arrival order among pressure surges, flow deviations, acoustic emission peaks, vibration spectrum shifts, and concentration increases. The model diagnostic process lacks processes for valve gap path status tokens, valve gap propagation kernel self-attention calculation, cross-covariance modulation, and valve topology manifold projection optimization, making it difficult to stably generate internal leakage status representations, external leakage status representations, equivalent leakage orifice diameters, and valve leakage monitoring results.

[0004] Therefore, how to provide a smart valve leakage monitoring method based on intelligent sensors is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0005] One objective of this invention is to propose an intelligent valve leakage monitoring method based on smart sensors. This invention utilizes multi-source sensor acquisition, dynamic event time warping, an improved TimeXer model, and a dung beetle optimizer to generate internal leakage state characterization, external leakage state characterization, and equivalent leakage orifice diameter. It has the advantages of accurate leakage identification, clear distinction between internal and external leakage, timely response to minor leakage, and reliable monitoring results.

[0006] According to an embodiment of the present invention, a valve leakage intelligent monitoring method based on a smart sensor includes: Based on the collection of valve multi-source state datasets by intelligent sensors, preprocessing is performed on the valve multi-source state datasets to generate standardized valve state datasets; Stable operating windows are selected based on standardized valve status datasets to generate normal sealing baseline data; the standardized valve status data corresponding to the current monitoring window is differentially processed with the normal sealing baseline data to generate valve status residual data; Based on valve status residual data, periodic status patches are generated. Candidate anomaly anchor points are extracted from the periodic status patches and dynamic event time warping is performed to generate candidate leakage event patches. The periodic state patch and candidate leakage event patch are input into the improved TimeXer model to perform valve gap path state token generation, valve gap propagation kernel self-attention calculation and cross covariance modulation to generate internal leakage state representation and external leakage state representation. Based on the internal leakage state representation and external leakage state representation, the initial equivalent leakage aperture is generated. The dung beetle optimizer is used to optimize the parameters of the improved TimeXer model. The valve topology manifold is constructed and the candidate parameters are rolled over and then projected onto the valve topology manifold. The leakage severity momentum is generated based on the initial equivalent leakage orifice diameter change, the candidate parameter search step size is corrected, and the optimized improved TimeXer model parameters are generated. Based on the optimized parameters of the improved TimeXer model, the periodic state patch and candidate leakage event patch are re-input into the improved TimeXer model to generate the optimized equivalent leakage orifice diameter, and valve leakage monitoring results are generated through the continuous readout layer of the leakage orifice diameter.

[0007] Optionally, the valve multi-source status dataset includes pressure, flow rate, acoustic emission, vibration, temperature, concentration, and operating condition data for the valve inlet pipe section, valve outlet pipe section, valve core sealing area, valve seat annular gap area, valve stem packing area, valve cover connection area, flange connection area, and external environment area.

[0008] Optionally, the preprocessing of the valve multi-source state dataset includes time alignment, outlier removal, and standardization.

[0009] Optionally, generating normal sealing reference data includes: The standardized valve status dataset is divided into continuous windows according to the monitoring time sequence to generate a continuous monitoring window set. Extract the pressure sequence, flow sequence, acoustic emission sequence, vibration sequence, temperature sequence, concentration sequence, and operating condition data sequence corresponding to each continuous monitoring window from the continuous monitoring window set, and calculate the fluctuation amplitude and change trend of each sequence; Based on the fluctuation amplitude and changing trend, continuous monitoring windows that meet the stable operation conditions are selected from the set of continuous monitoring windows to generate stable operation windows; Normal sealing reference data is generated based on the pressure sequence, flow sequence, acoustic emission sequence, vibration sequence, temperature sequence, concentration sequence, and operating condition data sequence within the stable operating window.

[0010] Optionally, the generation of valve state residual data includes: Read the standardized valve status data corresponding to the current monitoring window, and extract the pressure sequence, flow sequence, acoustic emission sequence, vibration sequence, temperature sequence, concentration sequence and operating condition data sequence from the standardized valve status data; Based on the data type matching relationship, subtract the corresponding reference values ​​in the normal sealing reference data from the sequence values ​​in the current monitoring window to generate pressure residual, flow residual, acoustic emission residual, vibration residual, temperature residual, concentration residual and operating condition residual; Valve status residual data are generated by arranging pressure residual, flow residual, acoustic emission residual, vibration residual, temperature residual, concentration residual, and operating condition residual according to the sampling time sequence and sensor installation area.

[0011] Optionally, generating candidate leak event patches includes: The valve status residual data is windowed according to the monitoring cycle, and the periodic status patch is generated according to the sampling time sequence and the sensor installation area marking window segmentation results. Extract pressure residual abrupt change points, flow residual deviation points, acoustic emission residual peak points, vibration residual spectrum shift points, temperature residual drift points, and concentration residual rise points from the periodic state patch to generate candidate anomaly anchor points; For candidate abnormal anchor points, record the anchor point type, anchor point occurrence time, anchor point duration, anchor point residual amplitude, sensor installation area and valve structure area, generate candidate abnormal anchor point records and construct anchor point spatiotemporal pairing relationships, calculate the regularization cost between candidate abnormal anchor points based on anchor point spatiotemporal pairing relationships, and generate regularization cost matrix. The minimum normalized path between candidate anomaly anchor points is searched according to the normalized cost matrix, and candidate anomaly anchor points with continuous arrival relationships within the same monitoring period are aggregated to generate candidate leakage event patches. The candidate leakage event patches include event type, event start time, event duration, residual amplitude set, sensor installation area set, valve structure area set, and event arrival order.

[0012] Optionally, generating the initial equivalent leakage aperture includes: An improved TimeXer model is constructed, which includes a patch embedding layer, a valve path state token layer, a valve propagation kernel self-attention layer, a cross-covariance modulation layer, an inner-drain and outer-drain branch output layer, and an aperture initial readout layer. The patch embedding layer reads the periodic state patch and the candidate leakage event patch, and generates the periodic state patch vector and the candidate leakage event patch vector according to the sampling time label, sensor installation area label, event type and event arrival order; The valve gap path status token layer reads the path start area, path end area, medium propagation direction, valve structure distance, sensor installation area and event arrival sequence corresponding to valve core internal leakage, valve seat annular gap leakage, valve stem packing external leakage, valve cover connection external leakage and flange connection external leakage, and generates valve gap path status tokens. The valve gap propagation kernel self-attention layer generates valve core internal leakage path state representation, valve seat annular gap leakage path state representation, valve stem packing external leakage path state representation, valve cover connection external leakage path state representation, and flange connection external leakage path state representation based on periodic state patch vector, candidate leakage event patch vector, valve gap path state token, time difference between candidate abnormal anchor points, valve structure distance, and medium propagation direction. The cross-covariance modulation layer reads the endogenous variable representations corresponding to pressure residuals, flow residuals, acoustic emission residuals, vibration residuals, temperature residuals, and concentration residuals, and the exogenous variable representations corresponding to operating condition residuals. It calculates the covariance relationship between the endogenous variable representations and the exogenous variable representations to generate the modulated path state representation. The internal leakage and external leakage branch output layers generate internal leakage status representation based on the modulated valve core internal leakage path status representation and the modulated valve seat annular gap leakage path status representation, and generate external leakage status representation based on the modulated valve stem packing external leakage path status representation, the modulated valve cover connection external leakage path status representation and the modulated flange connection external leakage path status representation. The initial readout layer reads the internal and external leakage state characteristics to generate the initial equivalent leakage aperture.

[0013] Optionally, the generation of optimized improved TimeXer model parameters includes: Read the improved TimeXer model parameters and encode them into candidate parameters. The candidate parameters include valve gap path state token parameters, valve gap propagation kernel parameters, cross covariance modulation parameters, internal and external leakage branch output parameters, orifice initial readout parameters, and leakage orifice continuous readout parameters. Construct the valve topology manifold based on the path start region, path end region, medium propagation direction, valve structure distance, and sensor installation area corresponding to the valve core internal leakage, valve seat annular gap leakage, valve stem packing external leakage, valve cover connection external leakage, and flange connection external leakage. Based on the dung beetle optimizer, the candidate parameters are rolled over and updated to generate rolling candidate parameters. The rolling candidate parameters are then projected onto the valve topology manifold according to the parameter segments corresponding to valve core internal leakage, valve seat annular gap leakage, valve stem packing external leakage, valve cover connection external leakage and flange connection external leakage, respectively, to generate topology projection candidate parameters. The candidate parameters of topological projection are written into the improved TimeXer model. Based on the periodic state patch and the candidate leakage event patch, the equivalent leakage aperture after perturbation is generated. The direction and magnitude of change of the equivalent leakage aperture after perturbation and the initial equivalent leakage aperture are compared to generate the leakage severity momentum. The search step size of the topology projection candidate parameters is adjusted based on the leakage severity momentum. The candidate parameters are updated based on the topology projection candidate parameters after the adjustment of the search step size. The candidate parameters are then filtered according to the number of missed reports, the number of leakage path identification errors, the number of internal and external leakage confusions, and the equivalent leakage aperture error, to generate the optimized and improved TimeXer model parameters.

[0014] Optionally, the generation of valve leakage monitoring results includes: Write the optimized improved TimeXer model parameters into the improved TimeXer model, and re-input the periodic state patch and candidate leakage event patch into the improved TimeXer model to generate the optimized internal leakage state representation and the optimized external leakage state representation. The continuous readout layer reads the optimized internal leakage state characterization, optimized external leakage state characterization, event duration, sensor installation area and valve structure area, and generates the optimized equivalent leakage orifice diameter. Based on the optimized equivalent leakage orifice diameter, the event arrival order in the candidate leakage event patch, and the valve structure region, the leakage status, leakage type, leakage path number, leakage start region, and abnormal duration are generated. Based on the leakage status, leakage type, leakage path number, leakage initiation area, abnormal duration, and optimized equivalent leakage orifice diameter, valve leakage monitoring results are generated.

[0015] The beneficial effects of this invention are: This invention proposes an intelligent valve leakage monitoring method based on smart sensors. By constructing a process involving multi-source valve state data acquisition, normal sealing baseline data generation, valve state residual data calculation, periodic state patch generation, and candidate leakage event patch generation, pressure, flow, acoustic emission, vibration, temperature, concentration, and operating condition data are transformed into structured input data for leakage identification. Compared to traditional fixed threshold judgment and single-sensor alarm methods, this invention can reduce the impact of operating condition fluctuations, sensor drift, and environmental disturbances on leakage identification results, and improve the stability of identifying minute leaks, sudden leaks, internal leaks, and external leaks.

[0016] This invention improves the TimeXer model by generating valve gap path state tokens, calculating valve gap propagation kernel self-attention, and modulating cross-covariance to generate internal and external leakage state representations. It further optimizes the TimeXer model parameters by combining valve topology manifold projection and leakage severity momentum updates using a dung beetle optimizer. Compared to traditional time-series prediction models and conventional classification networks, this invention enhances the path differentiation capabilities between valve core internal leakage, valve seat annular gap leakage, valve stem packing external leakage, valve cover connection external leakage, and flange connection external leakage. It also improves the accuracy of equivalent leakage orifice diameter estimation and generates leakage status, leakage type, leakage path number, leakage initiation region, anomaly duration, and valve leakage monitoring results. This invention offers advantages such as accurate leakage location, timely risk assessment, and strong interpretability of monitoring results. Attached Figure Description

[0017] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They explain the invention together with the embodiments of the invention and do not constitute a limitation thereof. In the drawings: Figure 1 This is an overall flowchart of a valve leakage intelligent monitoring method based on intelligent sensors proposed in this invention; Figure 2 This is a schematic diagram of the improved TimeXer model of the intelligent valve leakage monitoring method based on intelligent sensors proposed in this invention. Figure 3 This is a schematic diagram of the optimization process of the dung beetle optimizer for a valve leakage intelligent monitoring method based on intelligent sensors proposed in this invention; Detailed Implementation

[0018] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0019] refer to Figure 1 , Figure 2 and Figure 3 A smart valve leakage monitoring method based on smart sensors includes: Based on the collection of valve multi-source state datasets by intelligent sensors, preprocessing is performed on the valve multi-source state datasets to generate standardized valve state datasets; Stable operating windows are selected based on standardized valve status datasets to generate normal sealing baseline data; the standardized valve status data corresponding to the current monitoring window is differentially processed with the normal sealing baseline data to generate valve status residual data; Based on valve status residual data, periodic status patches are generated. Candidate anomaly anchor points are extracted from the periodic status patches and dynamic event time warping is performed to generate candidate leakage event patches. The periodic state patch and candidate leakage event patch are input into the improved TimeXer model to perform valve gap path state token generation, valve gap propagation kernel self-attention calculation and cross covariance modulation to generate internal leakage state representation and external leakage state representation. Based on the internal leakage state representation and external leakage state representation, the initial equivalent leakage aperture is generated. The dung beetle optimizer is used to optimize the parameters of the improved TimeXer model. The valve topology manifold is constructed and the candidate parameters are rolled over and then projected onto the valve topology manifold. The leakage severity momentum is generated based on the initial equivalent leakage orifice diameter change, the candidate parameter search step size is corrected, and the optimized improved TimeXer model parameters are generated. Based on the optimized parameters of the improved TimeXer model, the periodic state patch and candidate leakage event patch are re-input into the improved TimeXer model to generate the optimized equivalent leakage orifice diameter, and valve leakage monitoring results are generated through the continuous readout layer of the leakage orifice diameter.

[0020] In this embodiment, the valve multi-source status dataset includes pressure, flow rate, acoustic emission, vibration, temperature, concentration, and operating condition data for the valve inlet pipe section, valve outlet pipe section, valve core sealing area, valve seat annular gap area, valve stem packing area, valve cover connection area, flange connection area, and external environment area.

[0021] In this embodiment, the preprocessing of the valve multi-source state dataset includes time alignment, outlier removal, and standardization.

[0022] In this embodiment, generating normal sealing reference data includes: The standardized valve status dataset is divided into continuous windows according to the monitoring time sequence to generate a set of continuous monitoring windows. Specifically, the division of the standardized valve status dataset into continuous windows according to the monitoring time sequence is as follows: Based on the unified sampling timestamp of the standardized valve condition dataset, the window length is set to 30 minutes and the sliding step size is 5 minutes. Starting from the first sampling point, data segments are extracted in an increasing direction over time. Each data segment contains pressure, flow, acoustic emission, vibration, temperature, concentration and operating condition data sequences within the same time range. All data segments are numbered according to their start time to form a continuous monitoring window set. Extract the pressure, flow, acoustic emission, vibration, temperature, concentration, and operating condition data sequences corresponding to each continuous monitoring window from the continuous monitoring window set. Calculate the fluctuation amplitude and trend of each sequence. Specifically, the calculation of the fluctuation amplitude of each sequence is as follows: For the same type of sequence within each continuous monitoring window, the maximum, minimum, and average values ​​are read. The difference between the maximum and minimum values ​​is taken as the absolute fluctuation amplitude, and the ratio of the absolute fluctuation amplitude to the average value is taken as the relative fluctuation amplitude. Corresponding fluctuation amplitudes are generated for pressure sequence, flow sequence, acoustic emission sequence, vibration sequence, temperature sequence, concentration sequence, and operating condition data sequence, respectively. The specific trend is as follows: Linear fitting is performed on the same type of sequence within each continuous monitoring window according to the sampling time order. The slope of the fitted line is used as the trend parameter. A positive slope indicates that the sequence is on an upward trend, a negative slope indicates that the sequence is on a downward trend, and the absolute value of the slope is less than 1% of the mean of the corresponding sequence, indicating that the sequence is in a stable trend. A stable operation window is generated by selecting continuous monitoring windows that meet the stable operation conditions from the set of continuous monitoring windows based on fluctuation amplitude and changing trend. Specifically, the selection of continuous monitoring windows that meet the stable operation conditions based on fluctuation amplitude and changing trend is as follows: Stable operating conditions include relative fluctuations of pressure series not exceeding 3%, relative fluctuations of flow series not exceeding 3%, relative fluctuations of acoustic emission series not exceeding 5%, relative fluctuations of vibration series not exceeding 5%, relative fluctuations of temperature series not exceeding 2%, relative fluctuations of concentration series not exceeding 2%, relative fluctuations of operating condition data series not exceeding 2%, and all series change trend parameters satisfying a stable trend. Based on the pressure sequence, flow sequence, acoustic emission sequence, vibration sequence, temperature sequence, concentration sequence, and operating condition data sequence within the stable operating window, normal sealing reference data is generated. Specifically, the generation of normal sealing reference data involves: For the pressure sequence, flow sequence, acoustic emission sequence, vibration sequence, temperature sequence, concentration sequence, and operating condition data sequence within the stable operation window, calculate the baseline mean, baseline median, baseline upper limit, and baseline lower limit respectively. The baseline upper limit is the baseline mean plus 3 times the standard deviation, and the baseline lower limit is the baseline mean minus 3 times the standard deviation. Combine the baseline mean, baseline median, baseline upper limit, and baseline lower limit corresponding to each type of sequence according to the sensor installation area and sampling time order to generate normal sealing baseline data.

[0023] In this embodiment, generating valve state residual data includes: Read the standardized valve status data corresponding to the current monitoring window, and extract the pressure sequence, flow sequence, acoustic emission sequence, vibration sequence, temperature sequence, concentration sequence, and operating condition data sequence from the standardized valve status data, including: Read the standardized valve status data corresponding to the current monitoring window, specifically: The current sampling time is used as the end time of the window. A 30-minute data segment is extracted from the previous time as the current monitoring window. The current monitoring window includes standardized valve status data corresponding to the valve inlet pipe section, valve outlet pipe section, valve core sealing area, valve seat annular gap area, valve stem packing area, valve cover connection area, flange connection area, and external environment area. Extracting each sequence from the standardized valve condition data, specifically: According to the data type field, pressure data, flow data, acoustic emission data, vibration data, temperature data, concentration data, and operating condition data are read from the current monitoring window respectively. They are arranged in ascending order of sampling timestamp to generate pressure sequence, flow sequence, acoustic emission sequence, vibration sequence, temperature sequence, concentration sequence, and operating condition data sequence. The operating condition data sequence includes valve opening sequence, medium temperature sequence, pipeline operating load sequence, and ambient temperature sequence. Based on data type matching relationships, each sequence value within the current monitoring window is subtracted from the corresponding baseline value in the normal sealing baseline data to generate pressure residuals, flow residuals, acoustic emission residuals, vibration residuals, temperature residuals, concentration residuals, and operating condition residuals, among which: The data type matching relationship is as follows: The pressure sequence in the current monitoring window is matched with the pressure reference value in the normal sealing reference data; the flow rate sequence is matched with the flow rate reference value; the acoustic emission sequence is matched with the acoustic emission reference value; the vibration sequence is matched with the vibration reference value; the temperature sequence is matched with the temperature reference value; the concentration sequence is matched with the concentration reference value; and the operating condition data sequence is matched with the operating condition reference value. The matching process is performed according to the same data type, the same sensor installation area, and the same sampling time and location. The benchmark values ​​are as follows: The benchmark values ​​of the same type are taken as the benchmark mean of the corresponding data type in the normal sealing benchmark data. If the sampling time position of the current monitoring window does not completely coincide with the sampling time position in the normal sealing benchmark data, the benchmark mean corresponding to the nearest benchmark time position is read. If there are two benchmark time positions that are the same distance apart, the arithmetic mean of the two benchmark means is taken. Generate various types of residuals, specifically: The pressure residual is generated by subtracting the pressure reference value from the pressure value at each sampling time location within the current monitoring window; the flow residual is generated by subtracting the flow reference value from the flow rate value; the acoustic emission residual is generated by subtracting the acoustic emission reference value from the acoustic emission value; the vibration residual is generated by subtracting the vibration reference value from the vibration value; the temperature residual is generated by subtracting the temperature reference value from the temperature value; the concentration residual is generated by subtracting the concentration reference value from the concentration value; and the operating condition residual is generated by subtracting the operating condition reference value from the operating condition data value. Valve state residual data is generated by arranging pressure residuals, flow residuals, acoustic emission residuals, vibration residuals, temperature residuals, concentration residuals, and operating condition residuals according to sampling time sequence and sensor installation area. Specifically, the arrangement of pressure residuals, flow residuals, acoustic emission residuals, vibration residuals, temperature residuals, concentration residuals, and operating condition residuals according to sampling time sequence and sensor installation area is as follows: Using the sampling timestamp as the first sorting index and the sensor installation area as the second sorting index, the sensor installation areas are arranged in the following order: valve inlet pipe section, valve outlet pipe section, valve core sealing area, valve seat annular gap area, valve stem packing area, valve cover connection area, flange connection area, and external environment area. Pressure residual, flow residual, acoustic emission residual, vibration residual, temperature residual, concentration residual, and operating condition residual are written sequentially under each sampling timestamp to generate valve status residual data.

[0024] In this embodiment, generating candidate leak event patches includes: The valve status residual data is windowed according to the monitoring cycle, and a periodic status patch is generated based on the windowing results according to the sampling time sequence and the sensor installation area markings. The valve status residual data is windowed according to the monitoring cycle, specifically as follows: The monitoring period is 5 minutes. Valve status residual data for 5 minutes consecutively backward from the current sampling time is taken to form a periodic window. The window sliding step is 1 minute. 4 minutes of overlapping data are retained between adjacent periodic windows. The sampling timestamp is used as the segmentation index. Pressure residual, flow residual, acoustic emission residual, vibration residual, temperature residual, concentration residual and operating condition residual are written into the same periodic window to form residual segments arranged by time. The results are segmented according to the sampling time sequence and the sensor installation area marking window, as follows: Generate area numbers according to the valve inlet pipe section, valve outlet pipe section, valve core sealing area, valve seat annular gap area, valve stem packing area, valve cover connection area, flange connection area and external environment area, write the area numbers into the corresponding residual fragments, and generate periodic status patches. From the periodic state patch, extract pressure residual abrupt change points, flow residual deviation points, acoustic emission residual peak points, vibration residual spectral shift points, temperature residual drift points, and concentration residual rise points to generate candidate anomaly anchor points, where: The pressure residual abrupt change point is specifically: Read the pressure residual difference between two adjacent sampling points within the periodic status patch. When the absolute value of the difference is greater than three times the pressure standard deviation in the normal sealing reference data, mark the next sampling point as the pressure residual abrupt change point. The flow residual deviation point is as follows: Read the flow residuals of the valve inlet pipe section and valve outlet pipe section within the periodic status patch. When the difference between the flow residuals of the two regions is greater than twice the standard deviation of the flow in the normal sealing reference data for three consecutive sampling points, mark the starting sampling point of the continuous interval as the flow residual deviation point. The peak point of the acoustic emission residual is as follows: Read the local peak value of the acoustic emission residual. When the local peak value is greater than the mean of the acoustic emission residual within the same period state patch plus 3 times the standard deviation of the acoustic emission residual, mark the sampling point corresponding to the local peak value as the peak value point of the acoustic emission residual. The vibration residual spectrum shift point is specifically: Perform a fast Fourier transform on the vibration residual, read the dominant frequency position and dominant frequency amplitude. When the dominant frequency position deviates from the reference dominant frequency in the normal sealing reference data by more than 5%, mark the corresponding sampling point as the vibration residual spectrum offset point. The temperature residual drift point is specifically: Read the linear fitting slope of the temperature residual over 10 consecutive sampling points. When the linear fitting slope is greater than twice the temperature change slope in the normal sealed reference data, mark the starting sampling point of the continuous interval as the temperature residual drift point. The concentration residual rise point is specifically: Read the concentration residual of the external environment area. When the concentration residual increases for 5 consecutive sampling points and the final concentration residual is greater than twice the concentration standard deviation in the normal sealed reference data, mark the starting sampling point of the continuous interval as the concentration residual rising point. For candidate abnormal anchor points, record the anchor point type, anchor point occurrence time, anchor point duration, anchor point residual amplitude, sensor installation area, and valve structure area. Generate candidate abnormal anchor point records and construct spatiotemporal pairing relationships for anchor points, where: Generate candidate anomaly anchor point records, specifically as follows: For each candidate abnormal anchor point, write the anchor point type, anchor point occurrence time, anchor point duration, anchor point residual amplitude, sensor installation area and valve structure area. The anchor point duration is the length of time during which the candidate abnormal anchor point continuously meets the corresponding judgment condition, and the anchor point residual amplitude is the maximum absolute value of the residual within the interval where the corresponding judgment condition is continuously met. The spatiotemporal pairing relationship of anchor points is as follows: Based on the conditions that the time difference between anchor points does not exceed 60 seconds, the sensor installation areas are adjacent areas, and the valve structure areas have a direct connection relationship, two candidate abnormal anchor points are paired in a spatiotemporal relationship. The valve inlet pipe section is adjacent to the valve core sealing area, the valve core sealing area is adjacent to the valve seat annular gap area, the valve seat annular gap area is adjacent to the valve outlet pipe section, and the valve stem packing area, valve cover connection area, and flange connection area are adjacent to the external environment area respectively. Based on the spatiotemporal pairing relationship of anchor points, the normalization cost between candidate anomaly anchor points is calculated, generating a normalization cost matrix, where: The adjustment costs are as follows: The normalization cost is obtained by adding the normalized value of the time difference of anchor point occurrence, the normalized value of the duration difference of anchor point, the cost of the direction of change of anchor point residual amplitude, the cost of the distance of sensor installation area, and the cost of the adjacency of valve structure area. The coefficient of each normalization cost is 1. The normalized cost matrix is ​​as follows: Arrange the candidate anomaly anchors in ascending order of their occurrence time. Write the normalization cost between any two candidate anomaly anchors into the corresponding position in the matrix. Write the maximum cost of 10 into the candidate anomaly anchor pairs that do not form a spatiotemporal pairing relationship. This forms a normalization cost matrix. The minimum normalized path between candidate anomaly anchor points is searched according to the normalized cost matrix, and candidate anomaly anchor points with continuous arrival relationships within the same monitoring period are aggregated to generate candidate leakage event patches. Each candidate leakage event patch includes the event type, event start time, event duration, residual amplitude set, sensor installation area set, valve structure area set, and event arrival order. The minimum regularization path is as follows: From the normalization cost matrix, select the candidate outlier anchor point connection path with the smallest cumulative normalization cost. The candidate outlier anchor points within the connection path are arranged in ascending order according to the anchor point occurrence time. When the cumulative normalization costs are the same, select the connection path with more anchor points. When the number of anchor points is the same, select the connection path with the larger residual amplitude of the initial anchor point. Candidate anomaly anchors with continuous arrival relationships are as follows: If the time difference between the occurrence of adjacent candidate abnormal anchor points within the same minimum regular path does not exceed 60 seconds, and there is an adjacent relationship between the sensor installation area and the valve structure area, then the candidate abnormal anchor points that meet the above conditions are determined to have a continuous arrival relationship. The patch for the candidate leak incident is as follows: Candidate abnormal anchor points with continuous arrival relationships within the minimum regularization path are aggregated into one candidate leakage event patch. The event type is determined based on the anchor point type with the most occurrences within the minimum regularization path. The event start time is taken as the earliest anchor point occurrence time, and the event duration is taken as the difference between the latest anchor point end time and the event start time. The residual amplitude set consists of the anchor point residual amplitudes of each candidate abnormal anchor point. The sensor installation area set consists of the sensor installation areas of each candidate abnormal anchor point. The valve structure area set consists of the valve structure areas of each candidate abnormal anchor point. The event arrival order is arranged in ascending order according to the anchor point occurrence time.

[0025] In this embodiment, generating the initial equivalent leakage aperture includes: An improved TimeXer model is constructed, comprising a patch embedding layer, a valve path state token layer, a valve propagation kernel self-attention layer, a cross-covariance modulation layer, an inner-drain and outer-drain branch output layer, and an aperture initial readout layer. Specifically, the improved TimeXer model is constructed as follows: This paper retains the endogenous variable patch embedding structure, exogenous variable token structure, encoder block, patch-level self-attention framework, variable-level cross-attention framework, feedforward network, layer normalization, residual connection path, and output projection framework of the traditional TimeXer model. A patch embedding layer is added at the traditional endogenous time series patch embedding location, transforming the original input structure that receives a single continuous time patch into a dual-patch input structure that receives both periodic state patches and candidate leakage event patches. A valve gap path state token layer is added at the traditional global endogenous token location, transforming the original global token that aggregates the overall endogenous time series state into path state tokens corresponding to valve core internal leakage, valve seat annular gap leakage, valve stem packing external leakage, valve cover connection external leakage, and flange connection leakage. The traditional patch-level self-attention framework... An internal valve gap propagation kernel self-attention layer is set up, transforming the original structure that only performs attention calculation based on patch vector similarity into a path propagation structure that performs attention calculation based on patch vector similarity, time difference between candidate anomaly anchor points, valve structure distance, and medium propagation direction. A cross-covariance modulation layer is set up between the traditional variable-level cross-attention framework and the feedforward network, transforming the original structure where exogenous variables directly participate in cross-attention into a condition disturbance stripping structure where the covariance relationship between endogenous variable representation and exogenous variable representation participates in channel correction. An internal leakage and external leakage branch output layer and an initial aperture readout layer are set up at the position of the traditional output projection framework, transforming the original predicted time series output structure into an output structure of internal leakage state representation, external leakage state representation, and initial equivalent leakage aperture. The patch embedding layer includes: Periodic status patch reading area: Reads the pressure residual, flow residual, acoustic emission residual, vibration residual, temperature residual, concentration residual, operating condition residual, sampling time label and sensor installation area label from the periodic status patch, and arranges them in ascending order of sampling time label to form a periodic status patch sequence; Candidate Leakage Event Patch Reading Area: Reads the event type, event start time, event duration, residual amplitude set, sensor installation area set, valve structure area set, and event arrival order from the candidate leakage event patches, and arranges them in order of event arrival from earliest to latest to form a candidate leakage event patch sequence; Patch Embedding Fusion Area: The residual values, sampling time labels, and sensor installation area labels in the periodic state patch sequence are concatenated. The event type, event start time, event duration, residual amplitude set, sensor installation area set, valve structure area set, and event arrival order in the candidate leakage event patch sequence are concatenated. The periodic state patch vector and the candidate leakage event patch vector are generated by linear projection with an output dimension of 256, respectively. Valve gap path status token layer, including: Path structure reading area: Reads the starting area, ending area, valve structure distance, sensor installation area, and event arrival sequence of the paths corresponding to valve core internal leakage, valve seat annular gap leakage, valve stem packing external leakage, valve cover connection external leakage, and flange connection external leakage. Among them, the valve core internal leakage path consists of the valve inlet pipe section, valve core sealing area, and valve outlet pipe section; the valve seat annular gap leakage path consists of the valve core sealing area, valve seat annular gap area, and valve outlet pipe section; the valve stem packing external leakage path consists of the valve stem packing area and the external environment area; the valve cover connection external leakage path consists of the valve cover connection area and the external environment area; and the flange connection external leakage path consists of the flange connection area and the external environment area. Medium propagation direction coding area: The medium propagation direction is configured according to the direction from the starting area of ​​the path to the ending area of ​​the path. The valve core internal leakage and valve seat annular gap leakage are configured as the propagation direction inside the valve. The valve stem packing external leakage, valve cover connection external leakage and flange connection leakage are configured as the propagation direction from the inside of the valve to the external environment area. Path token generation area: The path start area, path end area, media propagation direction, valve structure distance, sensor installation area and event arrival order are mapped to 128-dimensional embedding vectors respectively. After splicing the embedding vectors, the output dimension of 256 linear projection is used to generate valve core internal leakage path status tokens, valve seat annular gap leakage path status tokens, valve stem packing external leakage path status tokens, valve cover connection external leakage path status tokens and flange connection external leakage path status tokens. The gap propagation core self-attention layer includes: Propagation kernel input area: Read the periodic status patch vector, candidate leakage event patch vector and valve gap path status token, and write the time difference between candidate abnormal anchor points, valve structure distance and medium propagation direction into the propagation kernel input field; Attention score calculation area: Calculate the vector similarity value between the periodic state patch vector and the candidate leakage event patch vector, calculate the time difference normalization value between candidate abnormal anchor points, the valve structure distance normalization value and the media propagation direction consistency value, add the vector similarity value, the reverse value of the time difference normalization value, the reverse value of the valve structure distance normalization value and the media propagation direction consistency value to generate the valve gap propagation kernel attention score, where the media propagation direction consistency value is 1 when the direction is consistent and 0 when the direction is inconsistent; Path state output area: Based on the valve gap propagation kernel attention score, aggregate the periodic state patch vector and the candidate leakage event patch vector to generate the valve core internal leakage path state representation, valve seat annular gap leakage path state representation, valve stem packing external leakage path state representation, valve cover connection external leakage path state representation and flange connection external leakage path state representation. The cross-covariance modulation layer includes: Variable characterization reading area: Read the endogenous variable characterizations corresponding to pressure residual, flow residual, acoustic emission residual, vibration residual, temperature residual and concentration residual, and read the exogenous variable characterizations corresponding to operating condition residuals. Operating condition residuals include valve opening residual, medium temperature residual, pipeline operating load residual and ambient temperature residual. Covariance calculation area: Calculate the number of times the deviation direction of each endogenous variable channel and each exogenous variable channel is consistent at the same sampling time position, and multiply the deviation magnitude by the cumulative value to generate the covariance relationship between the endogenous variable representation and the exogenous variable representation; Channel modulation region: The responses of endogenous variable channels with a covariance relationship higher than 0.70 are adjusted to 0.70 of the original channel response, while the responses of endogenous variable channels with a covariance relationship not higher than 0.70 and matching the candidate leakage event patch are retained as the original channel response, generating a modulated path state representation. Here, 0.70 is the threshold for determining strong correlation between operating condition disturbance and the threshold is set based on the median of the covariance distribution of the operating condition disturbance channel and the leakage response channel in the training samples. The internal and external leakage branch output layer includes: Internal leakage branch region: Read the modulated valve core internal leakage path state characterization and the modulated valve seat annular gap leakage path state characterization, splice them according to the internal propagation order of the valve, and generate the internal leakage state characterization through linear mapping and ReLU activation function; External leakage branch area: Read the modulated valve stem packing external leakage path status representation, the modulated valve cover connection external leakage path status representation, and the modulated flange connection external leakage path status representation, splice them according to the external leakage structure area numbering order, and generate the external leakage status representation through linear mapping and ReLU activation function; The initial readout layer of the aperture includes: Aperture readout vector generation region: The inner leakage state representation, outer leakage state representation, event duration and residual amplitude set are concatenated to generate the initial aperture readout vector; Continuous regression region for orifice diameter: Perform continuous regression mapping on the initial orifice diameter readout vector to output five orifice diameter estimates corresponding to valve core internal leakage, valve seat annular gap leakage, valve stem packing external leakage, valve cover connection external leakage and flange connection external leakage. Take the maximum value among the five orifice diameter estimates as the initial equivalent leakage orifice diameter. The initial equivalent leakage orifice diameter is recorded in millimeters. In the improved TimeXer model, periodic state patches and candidate leakage event patches enter the patch embedding layer to generate periodic state patch vectors and candidate leakage event patch vectors. Path structure data enters the valve gap path state token layer to generate valve gap path state tokens. Periodic state patch vectors, candidate leakage event vectors, and valve gap path state tokens enter the valve gap propagation kernel self-attention layer to generate five types of path state representations. The five types of path state representations enter the cross-covariance modulation layer to generate modulated path state representations. The modulated path state representations enter the inner and outer leakage branch output layer to generate inner and outer leakage state representations. The inner and outer leakage state representations enter the aperture initial readout layer to generate the initial equivalent leakage aperture. The patch embedding layer reads the periodic state patch and candidate leakage event patch, and generates periodic state patch vectors and candidate leakage event patch vectors according to the sampling time label, sensor installation area label, event type, and event arrival order, where: Generate the periodic state patch vector, specifically as follows: The patch embedding layer reads the pressure residual, flow residual, acoustic emission residual, vibration residual, temperature residual, concentration residual, operating condition residual, sampling time label, and sensor installation area label from the periodic state patch. It normalizes each residual value to the interval between 0 and 1, maps the sampling time label to a 128-dimensional time embedding vector, maps the sensor installation area label to a 128-dimensional area embedding vector, and concatenates each residual value, time embedding vector, and area embedding vector and performs linear mapping to generate the periodic state patch vector. Generate candidate leak event patch vectors, specifically as follows: The patch embedding layer reads the event type, event start time, event duration, residual amplitude set, sensor installation area set, valve structure area set, and event arrival order from the candidate leakage event patch. It maps the event type to a 128-dimensional event embedding vector, maps the event arrival order to a 128-dimensional sequence embedding vector, normalizes the event duration and residual amplitude set to the interval between 0 and 1, and concatenates the embedding vector and normalized values ​​and performs a linear mapping to generate the candidate leakage event patch vector. The valve gap path status token layer reads the path start area, path end area, medium propagation direction, valve structure distance, sensor installation area, and event arrival sequence corresponding to valve core internal leakage, valve seat annular gap leakage, valve stem packing external leakage, valve cover connection external leakage, and flange connection external leakage, and generates a valve gap path status token, wherein: The path start region and path end region are as follows: The path of internal leakage of the valve core starts at the valve inlet pipe section and ends at the valve outlet pipe section. The path of leakage in the valve seat annular gap starts at the valve core sealing area and ends at the valve outlet pipe section. The path of external leakage of the valve stem packing starts at the valve stem packing area and ends at the external environment area. The path of external leakage of the valve cover connection starts at the valve cover connection area and ends at the external environment area. The path of external leakage of the flange connection starts at the flange connection area and ends at the external environment area. The direction of medium propagation is as follows: The direction of medium propagation is from the starting area of ​​the path to the ending area of ​​the path. The internal leakage of the valve core and the annular leakage of the valve seat correspond to the propagation direction inside the valve. The external leakage of the valve stem packing, the external leakage of the valve cover connection, and the external leakage of the flange connection correspond to the propagation direction from inside the valve to the external environment. The direction of medium propagation is written into the direction coding field. Valve structure distance, specifically: The distance between valve structures is calculated based on the number of connecting edges between valve structure areas. The distance between adjacent structure areas is recorded as 1, and the distance between two structure areas is recorded as 2. The distance between the valve inlet pipe section, valve core sealing area, valve seat annular gap area, and valve outlet pipe section is calculated according to the internal connection sequence. The distance between the valve stem packing area, valve cover connection area, and flange connection area and the external environment area is calculated according to the external leakage connection sequence. Generate valve gap path status tokens, specifically: The valve gap path status token layer reads the path start area, path end area, medium propagation direction, valve structure distance, sensor installation area and event arrival sequence corresponding to valve core internal leakage, valve seat annular gap leakage, valve stem packing external leakage, valve cover connection external leakage and flange connection external leakage respectively. The path start area, path end area, medium propagation direction and sensor installation area are converted into unique thermal codes according to category numbers. The valve structure distance is normalized according to the maximum structure distance. The event arrival sequence is converted into sequential codes according to the order of anchor point occurrence time. The unique thermal encoding of the path start region, the unique thermal encoding of the path end region, the unique thermal encoding of the medium propagation direction, the normalized valve structure distance, the unique thermal encoding of the sensor installation area, and the event arrival sequence encoding are respectively input into the corresponding embedding table to generate 6 128-dimensional embedding vectors. Following the fixed field order of path start region, path end region, medium propagation direction, valve structure distance, sensor installation area, and event arrival order, six 128-dimensional embedded vectors are concatenated into a 768-dimensional path concatenation vector. The linear mapping layer is set to have an input dimension of 768 and an output dimension of 256. The linear mapping layer includes a 768×256 mapping weight matrix and a bias vector of length 256. The mapping weight matrix is ​​initialized with a uniform distribution, and the initialization range is determined based on the input dimension of 768 and the output dimension of 256. The initial value of the bias vector is set to 0. The linear mapping layer performs a dimension-wise multiplication and addition calculation on the 768-dimensional path concatenation vector, that is, multiplies each of the 768 values ​​in the path concatenation vector with the weights of the corresponding columns in the mapping weight matrix and sums them up, and adds the bias values ​​of the corresponding dimensions in the bias vector to obtain a 256-dimensional path state vector. The same linear mapping process is performed on the valve core internal leakage, valve seat annular gap leakage, valve stem packing external leakage, valve cover connection external leakage and flange connection external leakage to generate valve core internal leakage path status token, valve seat annular gap leakage path status token, valve stem packing external leakage path status token, valve cover connection external leakage path status token and flange connection leakage path status token. The valve gap propagation kernel self-attention layer generates state representations for the following types of leakage paths based on periodic state patch vectors, candidate leakage event patch vectors, valve gap path state tokens, time differences between candidate abnormal anchor points, valve structure distances, and media propagation directions: valve core internal leakage path state representation, valve seat annular gap leakage path state representation, valve stem packing external leakage path state representation, valve cover connection external leakage path state representation, and flange connection external leakage path state representation. The input of the valve gap propagation kernel from the attention layer is specifically as follows: The valve gap propagation kernel reads the periodic state patch vector, the candidate leakage event patch vector, and the valve gap path state token from the attention layer. It uses the periodic state patch vector and the candidate leakage event patch vector as patch input, the valve gap path state token as path query input, and the time difference between candidate abnormal anchor points, valve structure distance, and medium propagation direction as propagation kernel input. The time difference between candidate anomaly anchor points is as follows: Read the anchor occurrence time of two candidate anomaly anchor points, subtract the earlier anchor occurrence time from the later anchor occurrence time to obtain the time difference between the candidate anomaly anchor points. The time difference is recorded in seconds. After the time difference is normalized to the interval of 0 to 1, it is written into the valve gap propagation kernel self-attention layer. The self-attention calculation of the valve gap propagation kernel is as follows: The valve gap propagation kernel first calculates the vector similarity value between patch vectors from the attention layer, and then reads the time difference between candidate anomaly anchor points, valve structure distance and medium propagation direction. The smaller the time difference, the shorter the valve structure distance and the more consistent the medium propagation direction, the larger the propagation kernel similarity value. The vector similarity value and the propagation kernel similarity value are added to obtain the attention score, and the patch vectors are aggregated based on the attention score. Generate state representations for each path, specifically as follows: The valve core internal leakage path status token aggregates the pressure residual, flow residual, and acoustic emission residual corresponding to the valve inlet pipe section, valve core sealing area, and valve outlet pipe section to generate a valve core internal leakage path status characterization. The valve seat annular gap leakage path status token aggregates the acoustic emission residual, vibration residual, and pressure residual corresponding to the valve seat annular gap area to generate a valve seat annular gap leakage path status characterization. The valve stem packing external leakage path status token aggregates the acoustic emission residual, vibration residual, temperature residual, and concentration residual corresponding to the valve stem packing area and the external environment area to generate a valve stem packing external leakage path status characterization. The valve cover connection external leakage path status token aggregates the acoustic emission residual, vibration residual, temperature residual, and concentration residual corresponding to the valve cover connection area and the external environment area to generate a valve cover connection external leakage path status characterization. The flange connection external leakage path status token aggregates the acoustic emission residual, vibration residual, temperature residual, and concentration residual corresponding to the flange connection area and the external environment area to generate a flange connection external leakage path status characterization. The cross-covariance modulation layer reads the endogenous variable representations corresponding to pressure residuals, flow residuals, acoustic emission residuals, vibration residuals, temperature residuals, and concentration residuals, and the exogenous variable representations corresponding to operating condition residuals. It calculates the covariance relationship between the endogenous and exogenous variable representations to generate the modulated path state representation, where: Endogenous variables are characterized as follows: The path state representations corresponding to pressure residual, flow residual, acoustic emission residual, vibration residual, temperature residual, and concentration residual are used as endogenous variable representations. The endogenous variable representations are used to record the sensor residual changes related to valve leakage response. The exogenous variable is characterized as follows: The path state representation corresponding to the operating condition residual is used as the exogenous variable representation. The operating condition residual includes valve opening residual, medium temperature residual, pipeline operating load residual and ambient temperature residual. The exogenous variable representation is used to record the influence of operating condition changes on pressure residual, flow residual, acoustic emission residual, vibration residual, temperature residual and concentration residual. The covariance relationship is as follows: The cross-covariance modulation layer reads the endogenous variable representation and the exogenous variable representation, and calculates the degree of synchronous change between each endogenous variable channel and each exogenous variable channel. The degree of synchronous change is obtained by multiplying the number of times the deviation direction of the two channels is consistent and the deviation amplitude at the same sampling time position by the cumulative value, thus forming the covariance relationship. The modulated path state representation is generated as follows: The cross-covariance modulation layer identifies endogenous variable channels that are significantly affected by the operating condition residuals based on the covariance relationship, reduces the corresponding channel response to 0.70 of the original channel response, and retains the original channel response for endogenous variable channels that are not synchronized with the operating condition residuals and are consistent with the candidate leakage event patch. After channel correction, the modulated path state representation is generated. The internal leakage and external leakage branch output layers generate an internal leakage state characterization based on the modulated valve core internal leakage path state characterization and the modulated valve seat annular gap leakage path state characterization, and generate an external leakage state characterization based on the modulated valve stem packing external leakage path state characterization, the modulated valve cover connection external leakage path state characterization, and the modulated flange connection external leakage path state characterization, wherein: Generate an internal leakage state representation, specifically as follows: The internal and external leakage branch output layer reads the modulated valve core internal leakage path state representation and the modulated valve seat annular gap leakage path state representation, splices the two types of path state representations according to the internal propagation order of the valve, and then generates the internal leakage state representation through linear mapping and nonlinear activation. Generate a representation of the leaked state, specifically as follows: The internal and external leakage branch output layer reads the modulated valve stem packing external leakage path status representation, the modulated valve cover connection external leakage path status representation, and the modulated flange connection external leakage path status representation. The three types of path status representations are spliced ​​together according to the external leakage structure area numbering order, and then linear mapping and nonlinear activation are performed to generate the external leakage status representation. The initial readout layer reads the internal and external leakage state characteristics to generate an initial equivalent leakage aperture, where: The initial readout layer reads the internal leakage state representation, external leakage state representation, event duration and residual amplitude set from the candidate leakage event patch. The internal leakage state representation is used as the valve internal leakage response vector, and the external leakage state representation is used as the valve external leakage response vector. The event duration is normalized according to 30 minutes as the maximum normalized duration. The pressure residual amplitude, flow residual amplitude, acoustic emission residual amplitude, vibration residual amplitude, temperature residual amplitude and concentration residual amplitude in the residual amplitude set are normalized according to the maximum training residual amplitude of the corresponding residual type, respectively, to obtain the event duration normalized value and the residual amplitude normalized vector. The valve internal leakage response vector, valve external leakage response vector, event duration normalized value, and residual amplitude normalized vector are concatenated according to a fixed field order to generate the orifice initial readout vector. The fixed field order is: valve internal leakage response vector, valve external leakage response vector, event duration normalized value, pressure residual amplitude normalized value, flow residual amplitude normalized value, acoustic emission residual amplitude normalized value, vibration residual amplitude normalized value, temperature residual amplitude normalized value, and concentration residual amplitude normalized value. The initial equivalent leakage aperture is generated as follows: The initial readout layer consists of a first linear readout layer, a nonlinear activation layer, a second linear readout layer, and an aperture scale conversion layer. The first linear readout layer reads the initial readout vector and maps it to a 128-dimensional aperture latent vector. The nonlinear activation layer performs a modified linear activation process on the 128-dimensional aperture latent vector to obtain a non-negative aperture latent vector. The second linear readout layer maps the non-negative aperture latent vector to a 5-dimensional aperture normalized vector. The first dimension of the 5-dimensional aperture normalization vector corresponds to the normalized value of the leakage aperture inside the valve core, the second dimension corresponds to the normalized value of the leakage aperture in the valve seat annular gap, the third dimension corresponds to the normalized value of the leakage aperture outside the valve stem packing, the fourth dimension corresponds to the normalized value of the leakage aperture outside the valve cover connection, and the fifth dimension corresponds to the normalized value of the leakage aperture outside the flange connection. The aperture scale conversion layer multiplies the 5 aperture normalization values ​​by 2.00mm to obtain the aperture estimates corresponding to the 5 paths. 2.00mm is the maximum equivalent leakage aperture set in the training samples. The initial orifice reading layer compares the orifice estimates corresponding to the five paths, and determines the orifice estimate with the largest value as the initial equivalent leakage orifice diameter. The path corresponding to the orifice estimate with the largest value is recorded as the initial orifice path. The initial orifice path is determined from the valve core internal leakage, valve seat annular gap leakage, valve stem packing external leakage, valve cover connection external leakage, and flange connection external leakage. The initial equivalent leakage orifice diameter is recorded in millimeters. The training objective of the first and second linear readout layers is to reduce the absolute error between the estimated orifice diameter and the labeled equivalent leakage orifice diameter. The labeled equivalent leakage orifice diameter in the training samples is obtained from the leakage orifice diameter calibration experiment. The range of the labeled equivalent leakage orifice diameter is set from 0 mm to 2.00 mm. When the estimated orifice diameter is less than 0.05 mm, it corresponds to a normal sealing state. When the estimated orifice diameter is not less than 0.05 mm and less than 0.20 mm, it corresponds to a suspected minor leakage state. When the estimated orifice diameter is not less than 0.20 mm, it corresponds to a confirmed leakage state.

[0026] In this embodiment, generating the optimized improved TimeXer model parameters includes: Read the improved TimeXer model parameters and encode them into candidate parameters. These candidate parameters include valve gap path state token parameters, valve gap propagation kernel parameters, cross-covariance modulation parameters, internal and external leakage branch output parameters, orifice initial readout parameters, and leakage orifice continuous readout parameters. The improved TimeXer model parameters are encoded as candidate parameters, specifically as follows: The valve gap path state token parameter, valve gap propagation kernel parameter, cross covariance modulation parameter, internal leakage and external leakage branch output parameter, orifice initial readout parameter, and leakage orifice continuous readout parameter are expanded into numerical vectors respectively. They are divided into 5 parameter segments according to valve core internal leakage, valve seat annular gap leakage, valve stem packing external leakage, valve cover connection external leakage, and flange connection external leakage. Each candidate parameter is obtained by splicing the 5 parameter segments in the order of path number P1 to P5. The dung beetle population size is set to 30 candidate parameters, and the maximum number of iterations is set to 80. The range of values ​​for the candidate parameters is as follows: The valve gap path status token parameter ranges from 0 to 1, the valve gap propagation kernel parameter ranges from 0 to 1, the cross covariance modulation parameter ranges from 0 to 1, the internal and external leakage branch output parameter ranges from 0 to 1, the orifice initial readout parameter ranges from 0 to 1, the leakage orifice continuous readout parameter ranges from 0 to 1, and candidate parameters exceeding the range are truncated according to the boundary values. The valve topology manifold is constructed based on the starting and ending regions of the paths corresponding to internal leakage of the valve core, annular leakage of the valve seat, external leakage of the valve stem packing, external leakage of the valve cover connection, and external leakage of the flange connection, as well as the medium propagation direction, valve structure distance, and sensor installation area. The valve topological manifold is as follows: Number the valve inlet pipe section as 1, the valve outlet pipe section as 2, the valve core sealing area as 3, the valve seat annular gap area as 4, the valve stem packing area as 5, the valve cover connection area as 6, the flange connection area as 7, and the external environment area as 8. The internal leakage path of the valve core is denoted as P1. The path node sequence of P1 is the valve inlet pipe section, the valve core sealing area, and the valve outlet pipe section. The leakage path of the valve seat annular gap is denoted as P2. The path node sequence of P2 is valve core sealing area, valve seat annular gap area and valve outlet pipe section. The leakage path of the valve stem packing is denoted as P3, and the path node sequence of P3 is the valve stem packing area and the external environment area. The external leakage path of the valve cover connection is denoted as P4. The path node sequence of P4 is the valve cover connection area and the external environment area. The external leakage path of the flange connection is denoted as P5, and the path node sequence of P5 is the flange connection area and the external environment area. Establish path connection edges according to the node order of each path, and encode the path number, starting area number, ending area number, medium propagation direction number, valve structure distance, sensor installation area number and path node order into a topological coordinate vector; The media propagation direction numbering includes internal propagation direction numbering and external dissipation direction numbering. The internal propagation direction numbering is used for the valve core internal leakage path and the valve seat annular gap leakage path, while the external dissipation direction numbering is used for the valve stem packing external leakage path, the valve cover connection external leakage path, and the flange connection external leakage path. The topological coordinate vector, path connecting edges, and path node order of each path together form the valve topological manifold of the corresponding path. Five valve topological manifolds are generated from P1 to P5 respectively. Valve structure distance, specifically: The distance between valve structures is calculated based on the number of connecting edges of the valve structure area. The distance between adjacent valve structure areas is recorded as 1, and the distance between 1 valve structure area is recorded as 2. The valve inlet pipe section, valve core sealing area, valve seat annular gap area and valve outlet pipe section are calculated according to the internal connection sequence. The valve stem packing area, valve cover connection area and flange connection area are calculated with the external environment area according to the external leakage connection sequence. The topological coordinate vector is specifically: The topological coordinate vector consists of path number, starting region number, ending region number, medium propagation direction number, valve structure distance, sensor installation area number, and path node sequence number; The path number is used to distinguish P1 to P5, the start area number and end area number are used to limit the leakage path range to which the candidate parameter belongs, the medium propagation direction number is used to distinguish the internal propagation direction and the external dissipation direction, the valve structure distance is used to indicate the structural interval between path nodes, the sensor installation area number is used to indicate the data source area corresponding to the candidate parameter, and the path node sequence number is used to indicate the order of the parameter projection along the leakage path. Generate the valve topology manifold as follows: Generate a set of topological coordinate vectors for P1 to P5 respectively, and connect adjacent topological coordinate vectors within the same path according to the path connection edge to obtain the path topological coordinate chain. Using the path topology coordinate chain as the projection space of candidate parameters, the parameter segments belonging to the same path in the rolling candidate parameters are only projected onto the corresponding path topology coordinate chain, generating valve topology manifolds corresponding to valve core internal leakage, valve seat annular gap leakage, valve stem packing external leakage, valve cover connection external leakage, and flange connection external leakage, respectively.

[0027] Based on the dung beetle optimizer, rolling updates are performed on the candidate parameters to generate rolling candidate parameters. These rolling candidate parameters are then projected onto the valve topology manifold according to the parameter segments corresponding to valve core internal leakage, valve seat annular clearance leakage, valve stem packing external leakage, valve cover connection external leakage, and flange connection external leakage, respectively, to generate topology projection candidate parameters. Among these: Rolling updates, specifically: The dung beetle optimizer reads the current candidate parameters, the current best candidate parameters, and the rolling step size. The initial value of the rolling step size is 10% of the length of the candidate parameter value range. It updates each parameter segment in the direction of the current candidate parameters toward the current best candidate parameters to generate rolling candidate parameters. Every 10 iterations, the rolling step size is reduced to 0.85 times the rolling step size of the previous stage. Projecting onto the valve topology manifold, specifically: Read the corresponding topological coordinates of parameter segments P1 to P5 in the rolling candidate parameters, calculate the angle between the parameter segment update direction and the corresponding valve topological manifold direction, project the parameter segment update direction onto the valve topological manifold direction with the smallest angle, delete the update components that are inconsistent with the corresponding valve topological manifold, and generate topological projection candidate parameters. The candidate parameters of the topological projection are written into the improved TimeXer model. Based on the periodic state patch and the candidate leakage event patch, the equivalent leakage aperture after perturbation is generated. The direction and magnitude of change between the equivalent leakage aperture after perturbation and the initial equivalent leakage aperture are compared to generate the leakage severity momentum, where: The equivalent leakage aperture after perturbation is generated based on periodic state patches and candidate leakage event patches, specifically: The topological projection candidate parameters are written into the improved TimeXer model. The fixed period state patch and candidate leakage event patch remain unchanged. The patch embedding, valve path state token generation, valve propagation kernel self-attention calculation, cross covariance modulation, internal and external leakage branch output and aperture initial readout are re-executed to generate the perturbation equivalent leakage aperture. The direction and magnitude of the change are as follows: When the equivalent leakage orifice diameter after disturbance is greater than the initial equivalent leakage orifice diameter, the direction of change is recorded as increasing; when the equivalent leakage orifice diameter after disturbance is less than the initial equivalent leakage orifice diameter, the direction of change is recorded as decreasing; when the equivalent leakage orifice diameter after disturbance is equal to the initial equivalent leakage orifice diameter, the direction of change is recorded as unchanged; and the absolute value of the difference between the equivalent leakage orifice diameter after disturbance and the initial equivalent leakage orifice diameter is recorded as the magnitude of change. Leakage severity momentum, specifically: The leakage severity momentum is obtained by adding 0.80 times the leakage severity momentum of the previous round and 0.20 times the value corresponding to the current aperture change direction. The value corresponding to the aperture change direction is 1 when it increases, -1 when it decreases, and 0 when it remains unchanged. The initial leakage severity momentum is 0. The search step size of the topology projection candidate parameters is adjusted based on the leakage severity momentum. The candidate parameters are updated based on the adjusted search step size. Candidate parameters are then filtered according to the number of missed detections, the number of incorrect leakage path identifications, the number of internal and external leakage confusions, and the equivalent leakage aperture error, generating optimized and improved TimeXer model parameters. The search step size based on the leakage severity momentum-corrected topological projection candidate parameters is as follows: When the leakage severity momentum is greater than 0 and the number of missed detections decreases, the search step size of the topology projection candidate parameters is adjusted to 1.20 times the current search step size. When the leakage severity momentum is less than 0 and the equivalent leakage aperture error decreases, the search step size of the topology projection candidate parameters is adjusted to 0.80 times the current search step size. When the number of missed detections increases, the topology projection candidate parameters are restored to the candidate parameters before the rolling update. The lower limit of the search step size is 1% of the length of the candidate parameter value range, and the upper limit of the search step size is 20% of the length of the candidate parameter value range. The specific steps for filtering candidate parameters are as follows: Candidate parameters are selected in the following order: number of missed reports, number of leakage path identification errors, number of internal and external leakage confusions, and equivalent leakage aperture error. Candidate parameters with fewer missed reports are retained first. When the number of missed reports is the same, the candidate parameter with fewer leakage path identification errors is selected. When the number of leakage path identification errors is the same, the candidate parameter with fewer internal and external leakage confusions is selected. When the number of internal and external leakage confusions is the same, the candidate parameter with smaller equivalent leakage aperture error is selected. The optimized and improved TimeXer model parameters are generated as follows: The optimization stops when the number of iterations reaches 80 or the candidate parameters retained for 10 consecutive iterations remain unchanged. The final retained candidate parameters are then split into valve gap path state token parameters, valve gap propagation kernel parameters, cross covariance modulation parameters, internal and external leakage branch output parameters, orifice initial readout parameters, and leakage orifice continuous readout parameters, generating the optimized improved TimeXer model parameters.

[0028] In this embodiment, generating valve leakage monitoring results includes: Write the optimized improved TimeXer model parameters into the improved TimeXer model, and re-input the periodic state patch and candidate leakage event patch into the improved TimeXer model to generate the optimized internal leakage state representation and the optimized external leakage state representation. The continuous readout layer reads the optimized internal leakage state characterization, optimized external leakage state characterization, event duration, sensor installation area, and valve structure area to generate the optimized equivalent leakage orifice diameter. Specifically, the generation of the optimized equivalent leakage orifice diameter involves: The continuous readout layer reads the optimized internal leakage state characterization, the optimized external leakage state characterization, the event duration in the candidate leakage event patch, the sensor installation area set, and the valve structure area set, and arranges the read content into an aperture readout vector according to the path number order; The continuous readout layer of leakage orifice performs continuous regression mapping on the orifice readout vector and outputs orifice value in millimeters. The orifice value corresponding to the path with the largest orifice value among valve core internal leakage, valve seat annular gap leakage, valve stem packing external leakage, valve cover connection external leakage and flange connection external leakage is used as the optimized equivalent leakage orifice. Based on the optimized equivalent leakage orifice diameter, the event arrival order in the candidate leakage event patches, and the valve structure region, the leakage status, leakage type, leakage path number, leakage start region, and anomaly duration are generated, where: The leak status is as follows: A normal state is generated when the optimized equivalent leakage orifice diameter is less than 0.05 mm and the abnormal duration is less than 2 min; a suspected leakage state is generated when the optimized equivalent leakage orifice diameter is not less than 0.05 mm and less than 0.20 mm or the abnormal duration is not less than 2 min and less than 10 min; and a confirmed leakage state is generated when the optimized equivalent leakage orifice diameter is not less than 0.20 mm or the abnormal duration is not less than 10 min. The types of leaks are as follows: When the path response intensity corresponding to the optimized internal leakage state characterization is greater than that corresponding to the optimized external leakage state characterization, an internal leakage type is generated. When the path response intensity corresponding to the optimized external leakage state characterization is greater than that corresponding to the optimized internal leakage state characterization, an external leakage type is generated. When the response intensities of the two types of paths are equal, a leakage type is generated according to the path with the larger optimized equivalent leakage aperture. The leakage path number and leakage starting area are as follows: The internal leakage path of the valve core is numbered P1, the leakage path of the valve seat annular gap is numbered P2, the external leakage path of the valve stem packing is numbered P3, the external leakage path of the valve cover connection is numbered P4, and the external leakage path of the flange connection is numbered P5. The path number corresponding to the optimized equivalent leakage orifice diameter is read as the leakage path number, and the valve structure area where the earliest candidate abnormal anchor point appears in the corresponding path is read as the leakage starting area. The duration of the abnormality is as follows: Read the event start time and the end time of the last candidate anomaly anchor point from the candidate leak event patch. Subtract the event start time from the end time of the last candidate anomaly anchor point to obtain the anomaly duration. The anomaly duration is recorded in minutes. Based on the leakage status, leakage type, leakage path number, leakage initiation area, abnormal duration, and optimized equivalent leakage orifice diameter, valve leakage monitoring results are generated.

[0029] Example 1: In a continuous online valve monitoring cycle, the monitoring object is an industrial pipeline control valve. A total of 24 intelligent sensors are arranged in the valve inlet section, valve outlet section, valve core sealing area, valve seat annular gap area, valve stem packing area, valve cover connection area, flange connection area, and external environment area to collect pressure, flow rate, acoustic emission, vibration, temperature, concentration, and operating condition data. After continuous data collection for 180 minutes, a multi-source state dataset of the valve is obtained, with a total of 259,200 raw data entries. The inlet pressure range is 0.62MPa–0.86MPa, the outlet pressure range is 0.41MPa–0.67MPa, the inlet flow rate range is 38.4m³ / h–47.8m³ / h, the outlet flow rate range is 38.1m³ / h–47.6m³ / h, the acoustic emission amplitude range is 0.006V–0.092V, the dominant vibration frequency range is 102Hz–146Hz, and the concentration range of the external environment area is 2ppm–38ppm. After the data enters the processing flow, the system performs time alignment according to a unified 1-second timestamp, removes 31 data points with obvious jumps, normalizes various monitoring values ​​to the 0-1 range, and generates a standardized valve status dataset.

[0030] The system divided the continuous monitoring window set into 31 windows, each with a 30-minute window length and a 5-minute sliding step. In the first to sixth continuous monitoring windows, the maximum relative pressure fluctuation was 1.8%, the maximum relative flow rate fluctuation was 1.5%, the maximum acoustic emission fluctuation was 3.2%, the maximum vibration fluctuation was 3.7%, the maximum concentration fluctuation was 1.1%, and the maximum operating condition data fluctuation was 1.4%, all meeting stable operating conditions. The system selected six stable operating windows to generate normal sealing baseline data: the average pressure baseline was 0.732 MPa, the average flow rate baseline was 42.6 m³ / h, the average acoustic emission baseline was 0.014 V, the dominant vibration frequency was 116 Hz, the average temperature baseline was 39.4 °C, and the average concentration baseline was 3.6 ppm.

[0031] Between 68 and 78 minutes, a minor internal leak occurred in the valve seat annular region. At 68 minutes, the residual pressure at the valve inlet was 0.006 MPa, the residual pressure at the valve outlet was 0.004 MPa, the residual inlet flow was 0.13 m³ / h, the residual outlet flow was 0.09 m³ / h, and the residual acoustic emission in the valve seat annular region was 0.012 V. At 72 minutes, the residual pressure at the valve inlet increased to 0.019 MPa, the residual pressure at the valve outlet increased to 0.016 MPa, the difference between the residual inlet and residual outlet flow reached 0.58 m³ / h, and the peak value of the residual acoustic emission in the valve seat annular region increased to 0.043 V. The system differentiated the current monitoring window with the normal sealing baseline data to form valve status residual data. The acoustic emission response in the valve seat annular region was 3.1 times the average of the normal baseline.

[0032] The system generates periodic status patches according to a 5-minute monitoring cycle and a 1-minute sliding step. In the periodic status patches from the 70th to the 75th minute, the system extracts three types of candidate anomaly anchor points: the acoustic emission residual peak point at the 71st minute, the pressure residual abrupt change point at the 72nd minute, and the flow residual deviation point at the 73rd minute. After establishing the spatiotemporal pairing relationship between the anchor points, the time difference between the acoustic emission residual peak point and the pressure residual abrupt change point is 48 seconds, the sensor installation area distance is one region, and the normalization cost is 0.28; the time difference between the pressure residual abrupt change point and the flow residual deviation point is 51 seconds, and the normalization cost is 0.31. The normalization cost matrix search yields the minimum normalization path, with a cumulative normalization cost of 0.59. The system aggregates the three candidate anomaly anchor points into one candidate leakage event patch. The event starts at the 71st minute, lasts for 3.4 minutes, and arrives in the order of acoustic emission residual peak point, pressure residual abrupt change point, and flow residual deviation point.

[0033] After inputting periodic state patches and candidate leakage event patches into the improved TimeXer model, the patch embedding layer generates 256-dimensional periodic state patch vectors and 256-dimensional candidate leakage event patch vectors. The valve gap path state token layer generates five path state tokens: the average response of the valve seat annular gap leakage path state token is 0.74, the average response of the valve core internal leakage path state token is 0.46, and the average response of the valve stem packing external leakage path state token is 0.18. The valve gap propagation kernel self-attention layer calculates the similarity values ​​of the patch vectors and the propagation kernel; the attention score for the valve seat annular gap leakage path is 0.82, while the highest attention score for other paths is 0.39. After reading the operating condition residuals, the cross-covariance modulation layer finds that the valve opening residual is 0.6%, the pipeline operating load residual is 1.2%, and the number of strongly correlated channels for operating condition disturbances is 1. After modulation, the pressure residual channel response is adjusted from 0.71 to 0.68, while the acoustic emission residual channel response remains at 0.86. The mean value of the internal leakage state characterization generated by the internal leakage and external leakage branch output layers is 0.79, the mean value of the external leakage state characterization is 0.21, and the initial equivalent leakage aperture of the initial readout layer output is 0.24 mm.

[0034] The Dung Beetle optimizer reads the parameters of the improved TimeXer model and encodes 30 candidate parameters. Each candidate parameter includes valve gap path state token parameters, valve gap propagation kernel parameters, cross-covariance modulation parameters, internal and external leakage branch output parameters, orifice initial readout parameters, and leakage orifice continuous readout parameters. The system constructs a valve topology manifold, projecting the valve seat annular gap leakage parameter segment onto the corresponding topological directions of the valve core sealing region, valve seat annular gap region, and valve outlet pipe segment. After the first round of rolling updates, the perturbed equivalent leakage orifice diameter is 0.27 mm, the orifice diameter changes in the direction of increase, the leakage severity momentum changes from 0 to 0.20, and the search step size is corrected from 10% to 12% of the candidate parameter range length. After the 8th iteration, the equivalent leakage aperture after perturbation is 0.31mm, the actual simulated aperture is 0.30mm, the equivalent leakage aperture error is 0.01mm, the number of missed reports is 0, the number of leakage path identification errors is 0, the number of internal and external leakage confusions is 0, and the system outputs the optimized improved TimeXer model parameters.

[0035] Between 118 and 132 minutes, external leakage occurred in the valve stem packing region. The peak value of the acoustic emission residual in the valve stem packing region reached 0.061V, the dominant vibration frequency in the valve stem packing region shifted from 116Hz to 132Hz, the concentration residual in the external environment region increased from 4ppm to 29ppm, and the temperature residual increased by 1.9℃ within 10 minutes. The system extracted the peak value of the acoustic emission residual at 119 minutes, the vibration residual spectrum shift point at 121 minutes, and the concentration residual increase point at 125 minutes. After the candidate leakage event patch was input into the optimized improved TimeXer model, the mean value of the external leakage state characterization was 0.83, the mean value of the internal leakage state characterization was 0.19, the optimized equivalent leakage orifice diameter was 0.48mm, the leakage status was confirmed leakage, the leakage type was external leakage, the leakage path number was P3, the leakage initiation region was the valve stem packing region, and the anomaly duration was 12.6 minutes.

[0036] This embodiment compares a traditional fixed threshold method with a standard CNN classification method. The training sample size is 3200 sets, and the test sample size is 800 sets. The test samples include 300 normal samples, 120 valve core internal leakage samples, 100 valve seat annular gap leakage samples, 100 valve stem packing external leakage samples, 90 valve cover connection external leakage samples, and 90 flange connection external leakage samples. The traditional fixed threshold method sets a pressure residual threshold of 0.025 MPa, a flow residual threshold of 0.80 m³ / h, an acoustic emission residual threshold of 0.050 V, and a concentration residual threshold of 20 ppm. The standard CNN classification method inputs a fixed 30-minute window of data and outputs three categories: normal, internal leakage, and external leakage. The method of this invention inputs a periodic state patch and a candidate leakage event patch, and outputs the leakage status, leakage type, leakage path number, leakage start area, abnormal duration, and optimized equivalent leakage orifice diameter.

[0037] In the same batch of test data, the traditional fixed threshold method achieved an overall recognition accuracy of 82.4%, a micro-leakage recall rate of 68.5%, an internal-to-external leakage differentiation accuracy of 79.2%, a false alarm rate of 9.8%, and an average alarm delay of 7.6 minutes. The ordinary CNN classification method achieved an overall recognition accuracy of 88.7%, a micro-leakage recall rate of 76.3%, an internal-to-external leakage differentiation accuracy of 84.1%, a false alarm rate of 7.1%, an average alarm delay of 5.2 minutes, and an average absolute error of 0.16 mm for the equivalent leakage aperture. The method of this invention achieved an overall recognition accuracy of 95.6%, a micro-leakage recall rate of 91.4%, an internal-to-external leakage differentiation accuracy of 94.2%, a false alarm rate of 3.2%, an average alarm delay of 1.8 minutes, and an average absolute error of 0.05 mm for the equivalent leakage aperture. In the external leakage simulation at 118 minutes, the traditional fixed threshold method triggered an alarm at 126 minutes, the ordinary CNN classification method triggered an alarm at 123 minutes, and the method of this invention outputs a suspected leakage status at 120 minutes and a confirmed leakage status at 121 minutes. As can be seen from Example 1, this invention forms a clear data change process at each step, and can maintain stable identification results under conditions of fluctuating operating conditions, small internal leaks, and lag in external leak concentration, while improving the accuracy of leak path location and aperture estimation.

[0038] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A smart sensor based valve leakage intelligent monitoring method, characterized in that, include: Based on the collection of valve multi-source state datasets by intelligent sensors, preprocessing is performed on the valve multi-source state datasets to generate standardized valve state datasets; Stable operating windows are selected based on standardized valve status datasets to generate normal sealing baseline data; the standardized valve status data corresponding to the current monitoring window is differentially processed with the normal sealing baseline data to generate valve status residual data; Based on valve status residual data, periodic status patches are generated. Candidate anomaly anchor points are extracted from the periodic status patches and dynamic event time warping is performed to generate candidate leakage event patches. The periodic state patch and candidate leakage event patch are input into the improved TimeXer model to perform valve gap path state token generation, valve gap propagation kernel self-attention calculation and cross covariance modulation to generate internal leakage state representation and external leakage state representation. Based on the internal leakage state representation and external leakage state representation, the initial equivalent leakage aperture is generated. The dung beetle optimizer is used to optimize the parameters of the improved TimeXer model. The valve topology manifold is constructed and the candidate parameters are rolled over and then projected onto the valve topology manifold. The leakage severity momentum is generated based on the initial equivalent leakage orifice diameter change, the candidate parameter search step size is corrected, and the optimized improved TimeXer model parameters are generated. Based on the optimized parameters of the improved TimeXer model, the periodic state patch and candidate leakage event patch are re-input into the improved TimeXer model to generate the optimized equivalent leakage orifice diameter, and valve leakage monitoring results are generated through the continuous readout layer of the leakage orifice diameter.

2. A method of intelligent monitoring of valve leakage based on smart sensor as claimed in claim 1, wherein, The valve multi-source status dataset includes pressure, flow rate, acoustic emission, vibration, temperature, concentration, and operating condition data for the valve inlet pipe section, valve outlet pipe section, valve core sealing area, valve seat annular gap area, valve stem packing area, valve cover connection area, flange connection area, and external environment area.

3. A method of intelligent monitoring of valve leakage based on smart sensor as claimed in claim 1, wherein, The preprocessing of the valve multi-source state dataset includes time alignment, outlier removal, and standardization.

4. The intelligent sensor based valve leak intelligent monitoring method as claimed in claim 1, wherein, The generation of normal sealing reference data includes: The standardized valve status dataset is divided into continuous windows according to the monitoring time sequence to generate a continuous monitoring window set. Extract the pressure sequence, flow sequence, acoustic emission sequence, vibration sequence, temperature sequence, concentration sequence, and operating condition data sequence corresponding to each continuous monitoring window from the continuous monitoring window set, and calculate the fluctuation amplitude and change trend of each sequence; Based on the fluctuation amplitude and changing trend, continuous monitoring windows that meet the stable operation conditions are selected from the set of continuous monitoring windows to generate stable operation windows; Normal sealing reference data is generated based on the pressure sequence, flow sequence, acoustic emission sequence, vibration sequence, temperature sequence, concentration sequence, and operating condition data sequence within the stable operating window.

5. A method of intelligent monitoring of valve leakage based on smart sensor as claimed in claim 1, wherein, The generated valve state residual data includes: Read the standardized valve status data corresponding to the current monitoring window, and extract the pressure sequence, flow sequence, acoustic emission sequence, vibration sequence, temperature sequence, concentration sequence and operating condition data sequence from the standardized valve status data; Based on the data type matching relationship, subtract the corresponding reference values ​​in the normal sealing reference data from the sequence values ​​in the current monitoring window to generate pressure residual, flow residual, acoustic emission residual, vibration residual, temperature residual, concentration residual and operating condition residual; Valve status residual data are generated by arranging pressure residual, flow residual, acoustic emission residual, vibration residual, temperature residual, concentration residual, and operating condition residual according to the sampling time sequence and sensor installation area.

6. A smart sensor based valve leak intelligent monitoring method as claimed in claim 1, wherein, The generation of candidate leak event patches includes: The valve status residual data is windowed according to the monitoring cycle, and the periodic status patch is generated according to the sampling time sequence and the sensor installation area marking window segmentation results. Extract pressure residual abrupt change points, flow residual deviation points, acoustic emission residual peak points, vibration residual spectrum shift points, temperature residual drift points, and concentration residual rise points from the periodic state patch to generate candidate anomaly anchor points; For candidate abnormal anchor points, record the anchor point type, anchor point occurrence time, anchor point duration, anchor point residual amplitude, sensor installation area and valve structure area, generate candidate abnormal anchor point records and construct anchor point spatiotemporal pairing relationships, calculate the regularization cost between candidate abnormal anchor points based on anchor point spatiotemporal pairing relationships, and generate regularization cost matrix. The minimum normalized path between candidate anomaly anchor points is searched according to the normalized cost matrix, and candidate anomaly anchor points with continuous arrival relationships within the same monitoring period are aggregated to generate candidate leakage event patches. The candidate leakage event patches include event type, event start time, event duration, residual amplitude set, sensor installation area set, valve structure area set, and event arrival order.

7. A smart sensor based valve leak intelligent monitoring method as claimed in claim 1, wherein, The generation of the initial equivalent leakage aperture includes: An improved TimeXer model is constructed, which includes a patch embedding layer, a valve path state token layer, a valve propagation kernel self-attention layer, a cross-covariance modulation layer, an inner-drain and outer-drain branch output layer, and an aperture initial readout layer. The patch embedding layer reads the periodic state patch and the candidate leakage event patch, and generates the periodic state patch vector and the candidate leakage event patch vector according to the sampling time label, sensor installation area label, event type and event arrival order; The valve gap path status token layer reads the path start area, path end area, medium propagation direction, valve structure distance, sensor installation area and event arrival sequence corresponding to valve core internal leakage, valve seat annular gap leakage, valve stem packing external leakage, valve cover connection external leakage and flange connection external leakage, and generates valve gap path status tokens. The valve gap propagation kernel self-attention layer generates valve core internal leakage path state representation, valve seat annular gap leakage path state representation, valve stem packing external leakage path state representation, valve cover connection external leakage path state representation, and flange connection external leakage path state representation based on periodic state patch vector, candidate leakage event patch vector, valve gap path state token, time difference between candidate abnormal anchor points, valve structure distance, and medium propagation direction. The cross-covariance modulation layer reads the endogenous variable representations corresponding to pressure residuals, flow residuals, acoustic emission residuals, vibration residuals, temperature residuals, and concentration residuals, and the exogenous variable representations corresponding to operating condition residuals. It calculates the covariance relationship between the endogenous variable representations and the exogenous variable representations to generate the modulated path state representation. The internal leakage and external leakage branch output layers generate internal leakage status representation based on the modulated valve core internal leakage path status representation and the modulated valve seat annular gap leakage path status representation, and generate external leakage status representation based on the modulated valve stem packing external leakage path status representation, the modulated valve cover connection external leakage path status representation and the modulated flange connection external leakage path status representation. The initial readout layer reads the internal and external leakage state characteristics to generate the initial equivalent leakage aperture.

8. A smart sensor based valve leak intelligent monitoring method as claimed in claim 1, wherein, The parameters for generating the optimized and improved TimeXer model include: Read the improved TimeXer model parameters and encode them into candidate parameters. The candidate parameters include valve gap path state token parameters, valve gap propagation kernel parameters, cross covariance modulation parameters, internal and external leakage branch output parameters, orifice initial readout parameters, and leakage orifice continuous readout parameters. Construct the valve topology manifold based on the path start region, path end region, medium propagation direction, valve structure distance, and sensor installation area corresponding to the valve core internal leakage, valve seat annular gap leakage, valve stem packing external leakage, valve cover connection external leakage, and flange connection external leakage. Based on the dung beetle optimizer, the candidate parameters are rolled over and updated to generate rolling candidate parameters. The rolling candidate parameters are then projected onto the valve topology manifold according to the parameter segments corresponding to valve core internal leakage, valve seat annular gap leakage, valve stem packing external leakage, valve cover connection external leakage and flange connection external leakage, respectively, to generate topology projection candidate parameters. The candidate parameters of topological projection are written into the improved TimeXer model. Based on the periodic state patch and the candidate leakage event patch, the equivalent leakage aperture after perturbation is generated. The direction and magnitude of change of the equivalent leakage aperture after perturbation and the initial equivalent leakage aperture are compared to generate the leakage severity momentum. The search step size of the topology projection candidate parameters is adjusted based on the leakage severity momentum. The candidate parameters are updated based on the topology projection candidate parameters after the adjustment of the search step size. The candidate parameters are then filtered according to the number of missed reports, the number of leakage path identification errors, the number of internal and external leakage confusions, and the equivalent leakage aperture error, to generate the optimized and improved TimeXer model parameters.

9. The intelligent valve leakage monitoring method based on intelligent sensors according to claim 1, characterized in that, The generated valve leakage monitoring results include: Write the optimized improved TimeXer model parameters into the improved TimeXer model, and re-input the periodic state patch and candidate leakage event patch into the improved TimeXer model to generate the optimized internal leakage state representation and the optimized external leakage state representation. The continuous readout layer reads the optimized internal leakage state characterization, optimized external leakage state characterization, event duration, sensor installation area and valve structure area, and generates the optimized equivalent leakage orifice diameter. Based on the optimized equivalent leakage orifice diameter, the event arrival order in the candidate leakage event patch, and the valve structure region, the leakage status, leakage type, leakage path number, leakage start region, and abnormal duration are generated. Based on the leakage status, leakage type, leakage path number, leakage initiation area, abnormal duration, and optimized equivalent leakage orifice diameter, valve leakage monitoring results are generated.

Images