Method for predicting service life of pipeline valve based on finite element analysis

Abstract

The application relates to the technical field of equipment life prediction, and particularly discloses a pipeline valve life prediction method based on finite element analysis, which comprises collecting valve shell stress distribution time series data, sealing surface wear accumulation data and medium pressure fluctuation time series to form a multi-time series monitoring data set. A valve stress field tensor sequence is generated through finite element grid node mapping, a sealing surface degradation trend curve is obtained through segmented fitting of the wear rate, and a pressure cycle load spectrum is extracted through frequency domain decomposition of the pressure fluctuation sequence. The stress field tensor sequence and the pressure cycle load spectrum time axis are aligned to construct a stress-load coupling feature matrix, joint time series coding is completed by using a cycle gating network, a valve residual life probability distribution is output, an expected failure time window is demarcated according to the probability distribution, and pipeline valve life prediction is completed.

Description

Technical Field

[0001] This invention belongs to the field of equipment life prediction technology, specifically relating to a method for predicting the life of pipeline valves based on finite element analysis. Background Technology

[0002] As critical control components in fluid transport systems, pipeline valves are continuously subjected to the combined effects of shell structural stress, sealing surface wear, and dynamic fluctuations in medium pressure during long-term service. Current pipeline valve life prediction methods generally employ independent analysis of single physical parameters, lacking an integrated processing architecture for multi-source time-series data including shell stress distribution, cumulative sealing surface wear, and medium pressure fluctuations. Stress monitoring data undergoes only discrete numerical analysis without incorporating finite element mesh node mapping, thus failing to generate a structured stress field tensor sequence. Sealing surface wear data is analyzed using conventional linear trends without piecewise fitting of wear rates, and medium pressure time-series data is retained only in the time domain, without frequency domain decomposition to extract the cyclic load spectrum.

[0003] Existing technologies lack a time axis alignment mechanism for stress sequences and pressure load spectra, making it impossible to construct a feature matrix characterizing the coupling effects of multiple physics fields. Conventional lifetime assessments often rely on traditional empirical models for extrapolation, failing to employ temporal deep learning networks to jointly encode multiple degradation features and thus unable to output a probabilistic distribution of remaining lifetime.

[0004] Traditional life assessment methods can only provide a single failure time point and cannot define time windows in the form of interval f. This makes it difficult to adapt to the actual changes in the nonlinear degradation of valves during service, and also fails to meet the application requirements of full life cycle health management, refined life assessment, and standardized service cycle management of pipeline valves. Summary of the Invention

[0005] This invention addresses the shortcomings of existing technologies by providing a method for predicting the lifespan of pipeline valves based on finite element analysis, thereby solving the problem of accurate prediction of pipeline valve lifespan.

[0006] This invention achieves this objective through the following technical solution: Methods for predicting the lifespan of pipeline valves based on finite element analysis include: Step 1: Obtain a multi-time series monitoring data set of the target pipeline valve during its service life. The multi-time series monitoring data set includes time series data of valve body stress distribution, cumulative data of sealing surface wear, and time series of medium pressure fluctuations. Step 2: Perform finite element mesh node mapping processing on the time series data of the stress distribution of the valve body to generate the valve stress field tensor sequence; Step 3: Perform piecewise fitting processing on the cumulative wear data of the sealing surface to obtain the degradation trend curve of the sealing surface; Step 4: Perform frequency domain decomposition on the medium pressure fluctuation time series to extract the pressure cyclic load spectrum; Step 5: Align the valve stress field tensor sequence with the pressure cyclic load spectrum along the time axis to generate a stress-load coupling feature matrix; Step 6: Call the pre-trained recurrent gating network to perform joint temporal encoding on the stress-load coupling feature matrix and the sealing surface degradation trend curve, and output the valve remaining life probability distribution; Step 7: Determine the expected failure time window of the valve based on the probability distribution of the valve's remaining lifespan, and mark the expected failure time window as the lifespan prediction result.

[0007] As a further aspect of the present invention, step 2 specifically includes: Finite element meshing is performed on each frame of the stress cloud map in the time series data of the stress distribution of the valve body to obtain the stress component value corresponding to each mesh node; Extract the three principal stress direction values ​​of each grid node from the stress component values, and combine the three principal stress direction values ​​into a node stress vector; The nodal stress vectors of all mesh nodes are arranged in a matrix according to the mesh topology to form a single-frame stress field tensor; The stress field tensors of each frame at each time step are stacked in order to generate the valve stress field tensor sequence.

[0008] As a further aspect of the present invention, step 3 specifically includes: The cumulative wear data of the sealing surface is sorted according to the acquisition time sequence to obtain the wear time sequence; Perform a sliding window difference operation on the wear amount time series to calculate the instantaneous wear rate value within each time window; The wear amount time series is segmented according to the change range of the instantaneous wear rate value to obtain multiple wear stage sub-sequences; For each wear stage subsequence, least squares linear fitting is performed to obtain the stage wear rate straight line corresponding to each wear stage; The wear rate lines corresponding to each wear stage are spliced ​​at the segmentation points to generate the degradation trend curve of the sealing surface.

[0009] As a further aspect of the present invention, step 4 specifically includes: The medium pressure fluctuation time series is pre-processed to remove the static pressure baseline component, resulting in a dynamic pressure fluctuation series. Apply a discrete Fourier transform to the dynamic pressure fluctuation sequence to generate a pressure fluctuation spectrum. In the pressure fluctuation spectrum, discrete frequency components with amplitudes exceeding the background noise threshold are identified to obtain the set of dominant frequency components. Extract the waveform amplitude and phase angle corresponding to each discrete frequency component from the set of main frequency components, and combine the waveform amplitude and phase angle into a cyclic load parameter pair; The pressure cyclic load spectrum is generated by arranging all cyclic load parameter pairs in ascending order of frequency.

[0010] As a further aspect of the present invention, step 5 specifically includes: Extract the global stress extremum point and its grid node position corresponding to each moment in the valve stress field tensor sequence; Extract the peak pressure and valley pressure corresponding to each pressure cycle in the pressure cycle load spectrum; The global stress extremum points are paired with the peak pressures in the pressure cyclic load spectrum in chronological order to obtain a stress-pressure peak pairing sequence. Spatially correlate the grid node location of the global stress extreme point with the valve load area corresponding to the peak pressure to obtain a spatial correlation marker; The stress-pressure peak pairing sequence is merged and encoded with the spatial correlation marker to generate the stress-load coupling feature matrix.

[0011] As a further aspect of the present invention, step 6 specifically includes: The stress-load coupling feature matrix is ​​expanded into a sequence of input feature vectors over time steps, and the input feature vectors of each time step are fed into the input layer of the recurrent gated network in sequence. The slope value of the degradation trend curve of the sealing surface at the current time step is used as a state adjustment factor and synchronously input into the hidden layer of the cyclic gating network. In each time step of the recurrent gating network, the updated gate activation value and the reset gate activation value are calculated based on the input feature vector and the hidden state of the previous time step. The candidate hidden state is modulated based on the updated gate activation value and the reset gate activation value to generate the encoded hidden state at the current time step; The encoded hidden state of the last time step is input into the fully connected output layer of the recurrent gating network to obtain the predicted mean and predicted variance of the valve's remaining lifespan. The predicted mean and predicted variance are then combined to form the probability distribution of the valve's remaining lifespan.

[0012] As a further aspect of the present invention, in step 6, at each time step of the recurrent gating network, the updated gate activation value and the reset gate activation value are calculated based on the input feature vector and the hidden state of the previous time step, specifically including: The input feature vector is concatenated with the hidden state of the previous time step to generate a concatenated feature vector. Multiply the concatenated feature vector by the update gate weight matrix to obtain the linear combination value of the update gate; A nonlinear activation operation is applied to the linear combination value of the update gate to generate the update gate activation value; Multiply the spliced ​​feature vector by the reset gate weight matrix to obtain the reset gate linear combination value; A nonlinear activation operation is applied to the linear combination value of the reset gate to generate the reset gate activation value.

[0013] As a further aspect of the present invention, step 7 specifically includes: The predicted mean and predicted variance are extracted from the probability distribution of the valve's remaining life, and an initial confidence interval is constructed with the predicted mean as the center. The width of the initial confidence interval is adjusted according to the magnitude of the predicted variance to generate the expected failure time window; Verify whether the lower boundary of the expected failure time window is greater than zero. If the lower boundary is less than or equal to zero, then correct the lower boundary to the preset minimum positive value. The upper and lower boundaries of the revised expected failure time window are marked as the latest and earliest predicted failure times of the valve, respectively. The latest predicted failure time and the earliest predicted failure time are combined to form the lifetime prediction result.

[0014] As a further aspect of the present invention, step 1 further includes data cleaning processing of the multi-time-series monitoring data set to obtain a cleaned monitoring data set, specifically including: Abnormal stress peaks are detected in the time series data of stress distribution in the valve body. Abnormal stress frames that exceed the upper limit of stress range are removed to obtain stress data after cleaning. Missing values ​​are located in the cumulative wear data of the sealing surface, and the wear values ​​at the missing time points are filled in using a linear interpolation method to obtain complete wear data. The pressure fluctuation time series of the medium is subjected to high-frequency noise filtering, and pressure spikes are removed by a sliding median filtering algorithm to obtain smooth pressure fluctuation data. The post-cleaning stress data, complete wear data, and smoothed pressure fluctuation data are combined into the post-cleaning monitoring data set.

[0015] As a further aspect of the present invention, in step 5, before aligning the valve stress field tensor sequence with the pressure cyclic load spectrum along the time axis, the method further includes dimensionality reduction and compression of the valve stress field tensor sequence, specifically including: Perform a matrix flattening operation on each single frame stress field tensor in the valve stress field tensor sequence to obtain the stress field long vector. Calculate the covariance matrix among all stress field long vectors, and perform eigenvalue decomposition on the covariance matrix to obtain the eigenvalue sequence and eigenvector matrix; The eigenvectors corresponding to the eigenvalues ​​whose cumulative contribution rate exceeds a preset retention threshold are selected from the eigenvalue sequence to form the projection basis matrix; Perform matrix multiplication on each stress field long vector and the projection basis matrix to obtain the dimension-reduced stress field eigenvectors. All the dimensionality-reduced stress field eigenvectors are rearranged according to the original temporal order to generate a dimensionality-reduced and compressed valve stress field tensor sequence.

[0016] Compared with the prior art, the beneficial effects of this invention are as follows: Finite element mesh node mapping processing of valve shell stress distribution time-series data can convert discrete monitoring data into a standardized stress field tensor sequence, fully preserving the spatial and temporal variation characteristics of structural stress. Piecewise fitting of wear rate to cumulative sealing surface wear data can generate a continuous degradation trend curve based on data variation patterns, accurately presenting the evolution of sealing surface wear over time. Frequency domain decomposition of medium pressure fluctuation time series can separate the effective load components and extract the pressure cyclic load spectrum, removing redundant information from the time-series data. Stress, wear, and pressure monitoring data can all be converted into standardized feature expressions. The stress field tensor sequence and pressure cyclic load spectrum are aligned along the time axis to form a stress-load coupling feature matrix with correlation characterization capabilities, achieving structured integration of the temporal correspondence of multiple physics parameters.

[0017] A cyclic gating network performs joint temporal encoding on the stress-load coupling feature matrix and the sealing surface degradation trend curve. Multi-dimensional temporal features are deeply abstracted and fused, outputting a probability distribution of the valve's remaining life that reflects the equipment's degradation pattern. Using this probability distribution as a reference, an expected failure time window is defined, replacing the fixed-time-point judgment method and fully presenting the gradual degradation attributes of the valve's service state. State variations and temporal degradation characteristics during the life evolution process are quantitatively expressed, allowing continuous characterization of the valve's full-cycle operating state changes. The life prediction results have interval-based expression capabilities, adapting to the refined assessment logic of long-term service state of pipeline valves. Attached Figure Description

[0018] Figure 1 This is a flowchart of the pipeline valve life prediction method based on finite element analysis of the present invention. Figure 2 This is a flowchart of the finite element mesh node mapping and valve stress field tensor sequence generation of the present invention; Figure 3 This is a flowchart of the wear rate segmentation fitting and sealing surface degradation trend curve generation of the present invention. Detailed Implementation

[0019] Exemplary embodiments of the present invention will now be described in more detail with reference to the accompanying drawings. While exemplary embodiments of the present invention are shown in the drawings, it should be understood that the present invention can be implemented in various forms and should not be limited to the embodiments set forth herein. Rather, these embodiments are provided to enable a more thorough understanding of the present invention and to fully convey the scope of the invention to those skilled in the art. It should be noted that, unless otherwise specified, the embodiments and features described herein can be combined with each other. The present invention will now be described in detail with reference to the accompanying drawings and embodiments.

[0020] See Figure 1 This invention discloses a method for predicting the lifespan of pipeline valves based on finite element analysis, including the following: A multi-time-series monitoring dataset of the target pipeline valve during its service life is acquired. This dataset includes time-series data on valve body stress distribution, cumulative data on sealing surface wear, and time series data on medium pressure fluctuations. Finite element mesh node mapping is performed on the valve body stress distribution time-series data to generate a valve stress field tensor sequence. Wear rate piecewise fitting is performed on the cumulative sealing surface wear data to obtain a sealing surface degradation trend curve. Frequency domain decomposition is performed on the medium pressure fluctuation time series to extract the pressure cyclic load spectrum. The valve stress field tensor sequence and the pressure cyclic load spectrum are aligned along the time axis to generate a stress-load coupling feature matrix. A pre-trained cyclic gating network is used to perform joint temporal encoding on the stress-load coupling feature matrix and the sealing surface degradation trend curve, outputting the valve's remaining life probability distribution. The expected failure time window of the valve is determined based on the valve's remaining life probability distribution, and this expected failure time window is marked as the life prediction result.

[0021] In one embodiment of the present invention, when performing finite element mesh node mapping processing on the time-series data of the valve body stress distribution to generate the valve stress field tensor sequence, the specific operations include: (See below) Figure 2The stress cloud map of each frame in the time series data of the stress distribution of the valve body is divided into finite element meshes to obtain the stress component values ​​corresponding to each mesh node; the three principal stress direction values ​​of each mesh node are extracted from the stress component values, and the three principal stress direction values ​​are combined into a node stress vector; the node stress vectors of all mesh nodes are matrix-arranged according to the mesh topology to form a single-frame stress field tensor; the single-frame stress field tensors at each time step are stacked according to the time step order to generate the valve stress field tensor sequence.

[0022] The acquisition of a multi-time-series monitoring data set of the target pipeline valve during its service life also includes data cleaning of the multi-time-series monitoring data set to obtain a cleaned monitoring data set. Specifically, this involves: detecting abnormal stress peaks in the valve body stress distribution time-series data and removing abnormal stress frames exceeding the upper limit of the stress range to obtain cleaned stress data; locating missing values ​​in the cumulative wear data of the sealing surface and filling in the missing wear values ​​at the missing time points using linear interpolation to obtain complete wear data; performing high-frequency noise filtering on the medium pressure fluctuation time series and using a sliding median filtering algorithm to remove pressure spikes to obtain smooth pressure fluctuation data; and merging the cleaned stress data, complete wear data, and smoothed pressure fluctuation data into the cleaned monitoring data set.

[0023] In the specific implementation, a DN300 gate valve installed on a high-pressure natural gas transmission pipeline was used as the target pipeline valve. This target pipeline valve has been in continuous service for 36 months. During its service, 16 resistance strain gauges and 2 fiber optic strain sensors arranged on the surface of the valve body were used to collect time-series data on the stress distribution of the valve body at a sampling frequency of once per hour. Eddy current displacement sensors installed at the sealing surface were used to collect cumulative data on the wear of the sealing surface at a sampling frequency of once per day. A pressure transmitter set at the valve inlet was used to collect the time series of medium pressure fluctuations at a sampling frequency of once per minute. In this way, a multi-time series monitoring data set containing time-series data on the stress distribution of the valve body, cumulative data on the wear of the sealing surface, and time series of medium pressure fluctuations was obtained.

[0024] In specific implementation, when generating the valve stress field tensor sequence by mapping the time-series data of the valve body stress distribution using finite element mesh nodes, each frame of the stress cloud map in the time-series data of the valve body stress distribution is meshed using finite element meshes. Specifically, hexahedral elements are used to discretize the geometric model of the valve body, generating a total of 12,580 mesh nodes. For each mesh node, three normal stress components and three shear stress components are extracted, resulting in six stress component values ​​corresponding to each mesh node. From the stress component values, the three principal stress direction values ​​of each mesh node are extracted. By solving the eigenvalues ​​of the stress tensor matrix at that mesh node, the first principal stress is obtained. The values ​​of the second and third principal stresses are obtained, and the values ​​of the three principal stress directions are combined into a 3×1 nodal stress vector. The nodal stress vectors of all mesh nodes are matrix-arranged according to the topological order of the finite element mesh. The row index is arranged in ascending order of the node number, and the column index is arranged in the order of the three principal stress components, forming a single-frame stress field tensor matrix of size 12580×3. The single-frame stress field tensors of each sampling time are stacked in the order of the time step. Assuming that stress cloud maps are collected at T time steps, the T 12580×3 matrices are stacked along the third dimension to generate the valve stress field tensor sequence of size T×12580×3. Optionally, when performing abnormal stress peak detection on the time-series data of the stress distribution of the valve body, the upper limit of the stress range is set to 450 MPa. The stress component values ​​of all grid nodes in each frame of the stress cloud map are compared with 450 MPa. If any stress component value of any grid node in a certain frame exceeds 450 MPa, the frame is determined to be an abnormal stress frame and is removed. The remaining effective stress frames after removal are rearranged according to the original time sequence to obtain the cleaned stress data.

[0025] In some embodiments, when locating missing values ​​in the cumulative wear data of the sealing surface, the daily collected wear values ​​are arranged according to a date sequence. If a blank period with a time interval greater than 24 hours is detected in the date sequence, the time corresponding to the blank period is marked as a missing time point. For each missing time point, the wear value at the time before and the wear value at the time after the missing time point are taken, and a fill value is calculated using a linear interpolation method. The formula for calculating the fill value is as follows:

[0026] in: This indicates the wear and tear values ​​to fill in for missing time points. This represents the wear and tear value at the time of data collection preceding the missing time point. This represents the wear value at the next data collection point after the missing time point. This represents the timestamp value of the data collected before the missing time point. This represents the timestamp value of the data collection point following the missing time point. The timestamp value represents the missing time point. The formula is used to calculate the fill value for each missing time point in turn. After filling in the wear value values ​​for all missing time points, the complete wear data is obtained.

[0027] Optionally, when performing high-frequency noise filtering on the medium pressure fluctuation time series, a sliding median filtering algorithm is adopted. The sliding window width is set to 5 consecutive sampling points. Within each window, the 5 pressure values ​​within the window are sorted, and the median value after sorting is taken as the pressure filtering output value at the center of the window. The window is moved successively with a step size of 1 sampling point, traversing the entire pressure fluctuation sampling sequence. After removing pressure spikes, smooth pressure fluctuation data is obtained.

[0028] In some embodiments, when the post-cleaning stress data, the complete wear data, and the smoothed pressure fluctuation data are combined into the post-cleaning monitoring data set, the data is aligned according to a unified timestamp index. For the stress frame corresponding to each sampling time in the post-cleaning stress data, the wear value in the complete wear data and the pressure value in the smoothed pressure fluctuation data at that time are appended to the same data structure to form a data record for each time containing three fields: shell stress distribution matrix, sealing surface wear, and medium pressure value. The data records for all times are arranged in ascending order of time to form the post-cleaning monitoring data set.

[0029] It is understandable that, after the above data cleaning process, the post-cleaning monitoring data set includes valve body stress distribution time series data with outlier frames removed and valid stress frames retained; cumulative sealing surface wear data with all missing points filled to form a continuous time series; and medium pressure fluctuation time series with high-frequency spikes filtered out to form a smooth curve. All three are organized along a unified time axis. It is also understandable that, during the finite element mesh node mapping process for the valve body stress distribution time series data, each stress cloud map comes from a valid stress frame in the cleaned stress data. The mesh topology used for finite element mesh generation remains consistent across all time series frames to ensure the comparability of the generated valve stress field tensor sequence in the time dimension.

[0030] In one embodiment of the present invention, when performing piecewise fitting processing on the cumulative wear data of the sealing surface to obtain the sealing surface degradation trend curve, the specific operation is as follows: (Refer to...) Figure 3The cumulative wear data of the sealing surface is sorted according to the acquisition time sequence to obtain a wear time series sequence. A sliding window difference operation is performed on the wear time series sequence to calculate the instantaneous wear rate value within each time window. The wear time series sequence is segmented according to the variation amplitude of the instantaneous wear rate value to obtain multiple wear stage subsequences. Least square linear fitting is performed on each wear stage subsequence to obtain the stage wear rate line corresponding to each wear stage. The stage wear rate lines corresponding to each wear stage are spliced ​​at the segmentation points to generate the sealing surface degradation trend curve.

[0031] The specific steps for extracting the pressure cyclic load spectrum by performing frequency domain decomposition on the medium pressure fluctuation time series are as follows: First, perform detrending preprocessing on the medium pressure fluctuation time series to eliminate the static pressure baseline component, obtaining a dynamic pressure fluctuation series. Second, apply a discrete Fourier transform to the dynamic pressure fluctuation series to generate a pressure fluctuation spectrum. Third, identify discrete frequency components in the pressure fluctuation spectrum whose amplitude exceeds the background noise threshold to obtain a set of dominant frequency components. Fourth, extract the waveform amplitude and phase angle corresponding to each discrete frequency component from the set of dominant frequency components, and combine the waveform amplitude and phase angle into a cyclic load parameter pair. Fifth, arrange all cyclic load parameter pairs in ascending order of frequency to generate the pressure cyclic load spectrum.

[0032] In the specific implementation, a DN300 gate valve installed on a high-pressure natural gas transmission pipeline was used as the target pipeline valve. This target pipeline valve has been in continuous service for 36 months. 1080 cumulative data points of sealing surface wear were collected by an eddy current displacement sensor at a sampling frequency of once a day, forming a cumulative data set of sealing surface wear. At the same time, about 1.57 million pressure sampling points were collected for the medium pressure fluctuation time series at a sampling frequency of once per minute.

[0033] In specific implementation, when performing piecewise fitting processing on the cumulative wear data of the sealing surface to obtain the degradation trend curve of the sealing surface, the cumulative wear data of the sealing surface is sorted according to the collection time order, and the daily wear values ​​are arranged sequentially from the first day of service to the 1080th day to obtain the wear time series sequence. A sliding window difference operation is performed on the wear time series sequence, with the sliding window width set to 30 consecutive sampling points. Within each window, the wear difference between the end time and the beginning time of the window is calculated, and then divided by the 30-day time interval to obtain the instantaneous wear rate value corresponding to the center time of that window. After traversing the entire wear time series sequence, the time series of instantaneous wear rate values ​​is obtained. The wear time series sequence is segmented according to the variation range of the instantaneous wear rate values, and the stage where the instantaneous wear rate value is stable in the range of 0 to 0.05 mm / day is marked as the running-in wear stage. The wear rate is divided into three subsequences: a stable wear stage where the instantaneous wear rate is stable within the range of 0.05 to 0.12 mm / day; a severe wear stage where the instantaneous wear rate exceeds 0.12 mm / day and continues to rise; and a subsequence of wear amount is divided into three wear stage subsequences according to the boundary times of the three stages. Least squares linear fitting is performed on each wear stage subsequence, with time t as the independent variable and wear amount W as the dependent variable, to obtain the slope and intercept of each straight line. The stage wear rate corresponding to each wear stage is equal to the slope of the fitted straight line for that stage. The stage wear rate straight lines of the running-in wear stage, the stable wear stage, and the severe wear stage are spliced ​​at the segmentation points, i.e., the wear amounts of the two straight lines are equal at the segmentation points, forming a segmented continuous polygonal line, which generates the degradation trend curve of the sealing surface.

[0034] Optionally, when performing a sliding window differencing operation on the wear time series, the sliding window width is selected as 30 days to cover a complete pressure cycle. The formula for calculating the instantaneous wear rate value is:

[0035] in: This represents the instantaneous wear rate value corresponding to the i-th sampling point. This represents the cumulative wear data of the sealing surface at the 15th sampling time after the i-th sampling point. This represents the cumulative wear data of the sealing surface at the 15th sampling time before the i-th sampling point, and 30 represents the 30-day time interval corresponding to the window width.

[0036] In some embodiments, when extracting the pressure cyclic load spectrum by frequency domain decomposition of the medium pressure fluctuation time series, the medium pressure fluctuation time series undergoes detrending preprocessing. The linear trend term in the medium pressure fluctuation time series is fitted using the least squares method to obtain the static pressure baseline component. The static pressure baseline component is then subtracted from the original medium pressure fluctuation time series to obtain the dynamic pressure fluctuation series. A discrete Fourier transform is applied to the dynamic pressure fluctuation series to convert the time-domain signal into a frequency-domain signal, generating a pressure fluctuation spectrum, where the horizontal axis represents frequency and the vertical axis represents amplitude. Discrete frequency components with amplitudes exceeding the background noise threshold are identified in the pressure fluctuation spectrum. The noise threshold is set to three times the average amplitude of all frequency components. Frequency components exceeding this threshold are extracted to obtain a set of dominant frequency components. The waveform amplitude and phase angle corresponding to each discrete frequency component are extracted from this set. The waveform amplitude of each discrete frequency component is equal to the amplitude value of that frequency point in the spectrum, and the phase angle of each discrete frequency component is equal to the phase value of that frequency point in the spectrum. The waveform amplitude and phase angle are combined into a cyclic load parameter pair. All cyclic load parameter pairs are arranged in ascending order of frequency to form a list. Each list element contains three values: frequency value, waveform amplitude, and phase angle, generating the pressure cyclic load spectrum. Optionally, when applying a discrete Fourier transform to the dynamic pressure fluctuation sequence, a fast Fourier transform algorithm is used, with a sampling frequency set to once per minute. According to the Nyquist sampling theorem, the analyzable frequency range is 0 to 0.0083 Hz, and the transform length is 131,072 sampling points. Sequences exceeding this length are truncated, and sequences shorter than this length are zero-padded.

[0037] In some embodiments, when identifying discrete frequency components whose amplitude exceeds the background noise threshold, the background noise threshold is calculated as follows: the arithmetic mean of the amplitudes of all frequency components in the pressure fluctuation spectrum is calculated, and this arithmetic mean is multiplied by 3 to obtain the background noise threshold. All frequency components are iterated through, and frequency components with amplitudes greater than the threshold are identified as the dominant frequency components, while DC components near the fundamental frequency are removed. It can be understood that after the above-described piecewise fitting of the wear rate, the sealing surface degradation trend curve reflects the complete degradation process from the break-in wear stage to the severe wear stage. Each stage corresponds to a fixed stage wear rate, and the splicing at the segmentation points ensures the continuity of the curve. It can also be understood that after the above-described frequency domain decomposition, the pressure cyclic load spectrum contains the frequency, amplitude, and phase information of all significant cyclic load components in the medium pressure fluctuation. This information collectively describes the complete characteristics of the pressure cyclic load and is used for subsequent coupling analysis with the stress field tensor sequence.

[0038] In one embodiment of the present invention, when aligning the valve stress field tensor sequence with the pressure cyclic load spectrum along the time axis to generate a stress-load coupling feature matrix, the specific operations are as follows: extracting the global stress extremum point and its grid node position corresponding to each moment in the valve stress field tensor sequence; extracting the peak pressure and valley pressure corresponding to each pressure cycle in the pressure cyclic load spectrum; pairing the global stress extremum point with the peak pressure in the pressure cyclic load spectrum in time order to obtain a stress-pressure peak pairing sequence; spatially associating the grid node position of the global stress extremum point with the valve load area corresponding to the peak pressure to obtain a spatial association marker; merging and encoding the stress-pressure peak pairing sequence and the spatial association marker to generate the stress-load coupling feature matrix.

[0039] Before aligning the valve stress field tensor sequence with the pressure cyclic load spectrum along the time axis, the process further includes dimensionality reduction and compression of the valve stress field tensor sequence. Specifically, this involves: performing matrix flattening on each single-frame stress field tensor in the valve stress field tensor sequence to obtain a stress field long vector; calculating the covariance matrix among all stress field long vectors and performing eigenvalue decomposition on the covariance matrix to obtain an eigenvalue sequence and an eigenvector matrix; selecting eigenvectors corresponding to eigenvalues ​​whose cumulative contribution rate exceeds a preset retention threshold from the eigenvalue sequence to form a projection basis matrix; performing matrix multiplication on each stress field long vector and the projection basis matrix to obtain a dimensionality-reduced stress field eigenvector; and rearranging all dimensionality-reduced stress field eigenvectors according to their original temporal order to generate a dimensionality-reduced and compressed valve stress field tensor sequence.

[0040] In the specific implementation, a DN300 gate valve installed on a high-pressure natural gas transmission pipeline is taken as the target pipeline valve. The valve stress field tensor sequence and pressure cyclic load spectrum have been obtained for the target pipeline valve. The size of the valve stress field tensor sequence is T×12580×3, where T represents the number of time steps, 12580 represents the number of grid nodes, and 3 represents the number of principal stress components for each node. The pressure cyclic load spectrum contains 47 principal frequency components, and each principal frequency component contains three values: frequency value, waveform amplitude, and phase angle.

[0041] In specific implementation, when performing dimensionality reduction and compression on the valve stress field tensor sequence, a matrix flattening operation is performed on each single-frame stress field tensor in the valve stress field tensor sequence. The single-frame stress field tensor matrix of size 12580×3 is expanded into a stress field long vector of length 37740 in row-major order. For T time steps, a total of T stress field long vectors of length 37740 are obtained. The covariance matrix among all stress field long vectors is calculated. Each stress field long vector is regarded as a sample, with a sample size of T and a feature dimension of 37740. The specific method for calculating the covariance matrix is ​​to subtract the mean vector of all stress field long vectors from each stress field long vector, multiply by its own transpose, and then divide by T-1 to obtain a covariance matrix of size 37740×37740. The covariance matrix is ​​then subjected to eigenvalue decomposition to solve for the eigenvalue sequence and eigenvector matrix. The eigenvalues ​​are arranged in descending order. Each eigenvalue corresponds to an eigenvector. From the eigenvalue sequence, eigenvectors corresponding to eigenvalues ​​whose cumulative contribution rate exceeds a preset retention threshold are selected. The preset retention threshold is set to 95%. The cumulative contribution rate is calculated as the sum of the first k eigenvalues ​​divided by the sum of all eigenvalues. When the cumulative contribution rate first exceeds or equals 95%, the value of k is recorded. The eigenvectors corresponding to the first k eigenvalues ​​are selected to form a projection basis matrix, with a size of 37740×k. Each stress field length vector is multiplied by the projection basis matrix. Each stress field length vector is a 1×37740 row vector, and the projection basis matrix is ​​a 37740×k matrix. After multiplication, a 1×k dimension-reduced stress field eigenvector is obtained. All dimension-reduced stress field eigenvectors are rearranged according to their original temporal order to generate a dimension-reduced and compressed valve stress field tensor sequence. The size of this sequence is T×k, where k is much smaller than 37740. Optionally, the formula for calculating the cumulative contribution rate is:

[0042] in: This represents the cumulative contribution rate of the first k eigenvalues. This represents the value of the i-th eigenvalue in descending order. Let represent the value of the j-th eigenvalue, with the denominator being the sum of all eigenvalues ​​and the numerator being the sum of the first k eigenvalues. The value of makes and .

[0043] In some embodiments, when aligning the valve stress field tensor sequence with the pressure cyclic load spectrum along the time axis to generate a stress-load coupling feature matrix, the global stress extremum point and its corresponding grid node position at each moment in the dimension-reduced and compressed valve stress field tensor sequence are extracted. The global stress extremum point is defined as the value with the largest absolute value among the three principal stress directions of all grid nodes at a certain moment. The stress value corresponding to this extremum point is taken as the global stress extremum at that moment, and the original grid node number that generated the extremum point is recorded. The peak pressure and valley pressure corresponding to each pressure cycle in the pressure cyclic load spectrum are extracted. Each principal frequency component in the pressure cyclic load spectrum corresponds to a sinusoidal pressure fluctuation, from the peak of a complete cycle. Peak pressure values ​​are extracted from the location, and valley pressure values ​​are extracted from the trough locations of the same cycle. Peak pressure and valley pressure are combined into a pressure cycle pair. The global stress extremum points are paired with the peak pressures in the pressure cycle load spectrum according to the time sequence. Specifically, the global stress extremum point at time t is matched with the time of occurrence of the pressure cycle peak closest to t, forming a stress-pressure peak pairing sequence. Each pairing includes a time index, a global stress extremum value, and a corresponding peak pressure value. The grid node location of the global stress extremum point is spatially associated with the valve load area corresponding to the peak pressure. A mapping relationship table between grid node numbers and valve load areas is pre-established, as shown in Table 1.

[0044] Table 1: Mapping Relationship between Grid Node Numbers and Valve Load Areas

[0045] Based on the grid node number of the global stress extremum point, the corresponding valve load area name is found in the table above to obtain the spatial association marker, which is a categorical variable. The stress-pressure peak pairing sequence is merged and encoded with the spatial association marker to generate a multidimensional feature vector for each time step. The multidimensional feature vector contains the global stress extremum value, peak pressure value, valley pressure value, unique encoding of the spatial association marker, and the timestamp normalized value of the current time step. The multidimensional feature vectors of all time steps are stacked in chronological order to generate the stress-load coupling feature matrix. The number of rows in this matrix is ​​the number of time steps T, and the number of columns is the total number of dimensions of the multidimensional feature vector.

[0046] It is understandable that after the above dimensionality reduction and compression process, the valve stress field tensor sequence retains more than 95% of the energy information in the original stress field tensor sequence, while reducing the feature dimension from 37740 to k, significantly reducing the amount of data for subsequent processing. It is also understandable that after the above time axis alignment and merging encoding, the generated stress-load coupling feature matrix integrates the global extrema of the stress field, the spatial distribution of the pressure cyclic load, and temporal information into a unified matrix structure, with each time step corresponding to a row of feature vectors.

[0047] In one embodiment of the present invention, when a pre-trained recurrent gating network is invoked to perform joint temporal encoding of the stress-load coupling feature matrix and the sealing surface degradation trend curve to output the valve remaining life probability distribution, the specific operations are as follows: the stress-load coupling feature matrix is ​​expanded into a sequence of input feature vectors by time step, and the input feature vectors of each time step are sequentially fed into the input layer of the recurrent gating network; the slope value of the sealing surface degradation trend curve at the current time step is used as a state adjustment factor and synchronously input into the hidden layer of the recurrent gating network; in each time step of the recurrent gating network, the update gate activation value and the reset gate activation value are calculated based on the input feature vectors and the hidden state of the previous time step; the candidate hidden state is modulated based on the update gate activation value and the reset gate activation value to generate the encoded hidden state of the current time step; the encoded hidden state of the last time step is input into the fully connected output layer of the recurrent gating network to obtain the predicted mean and predicted variance of the valve remaining life, and the predicted mean and predicted variance are combined to form the valve remaining life probability distribution.

[0048] In each time step of the recurrent gating network, when calculating the update gate activation value and reset gate activation value based on the input feature vector and the hidden state of the previous time step, the specific operations are as follows: concatenate the input feature vector with the hidden state of the previous time step to generate a concatenated feature vector; multiply the concatenated feature vector by the update gate weight matrix to obtain the update gate linear combination value; apply a nonlinear activation operation to the update gate linear combination value to generate the update gate activation value; multiply the concatenated feature vector by the reset gate weight matrix to obtain the reset gate linear combination value; apply a nonlinear activation operation to the reset gate linear combination value to generate the reset gate activation value.

[0049] In the specific implementation, a DN400 butterfly valve installed in a chemical plant pipeline is used as the target pipeline valve. The target pipeline valve has been in service for 24 months. The stress-load coupling feature matrix and the sealing surface degradation trend curve are obtained through the data acquisition system. The stress-load coupling feature matrix has a size of 720×32, where 720 represents the number of time steps and 32 represents the feature dimension of each time step. The sealing surface degradation trend curve is formed by splicing two straight lines fitted to the wear stages. The corresponding slope value can be extracted at each time step.

[0050] In specific implementation, when calling a pre-trained recurrent gating network to jointly encode the stress-load coupling feature matrix and the sealing surface degradation trend curve to output the valve's remaining life probability distribution, the pre-trained recurrent gating network includes an input layer, a hidden layer, and a fully connected output layer. The number of neurons in the hidden layer is set to 96. The time step unfolding length of the recurrent gating network is consistent with the number of rows (720) of the stress-load coupling feature matrix. The stress-load coupling feature matrix is ​​unfolded into a sequence of input feature vectors over time steps. The input feature vector at each time step t is a row vector of size 1×32. The input feature vectors at each time step are fed into the input layer of the recurrent gating network in the order of t=1,2,…,720. The slope value of the sealing surface degradation trend curve at the current time step is used as a state adjustment factor and synchronously input into the hidden layer of the recurrent gating network. At each time step t, the instantaneous slope value corresponding to the time step is extracted from the sealing surface degradation trend curve as a state adjustment factor. The value will The input is fused with the hidden layer's state update process as a scalar input. At each time step of the recurrent gated network, the update gate activation value and reset gate activation value are calculated based on the input feature vector and the hidden state of the previous time step. The candidate hidden state is modulated based on the update gate activation value and reset gate activation value to generate the encoded hidden state of the current time step. The encoded hidden state of the last time step is input into the fully connected output layer of the recurrent gated network to obtain the predicted mean and predicted variance of the valve's remaining lifespan. The predicted mean and predicted variance are combined to form the valve's remaining lifespan probability distribution, which adopts a Laplace distribution. Optionally, the formulas for calculating the predicted mean m and the predicted scale parameter b are:

[0051] Where: m represents the predicted mean of the valve's remaining life, and b represents the prediction scale parameter for the valve's remaining life. This represents the weight matrix of the fully connected output layer. This indicates the encoded hidden state at the 720th time step, which is the last time step. Let represent the bias vector of the fully connected output layer, and 'softplus' represent the non-linear activation function, with the function form as follows: The output layer uses the softplus activation function to ensure that the prediction scale parameter is a non-negative value.

[0052] In practical implementation, at each time step of the recurrent gating network, when calculating the updated gate activation value and the reset gate activation value based on the input feature vector and the hidden state of the previous time step, the input feature vector is concatenated with the hidden state of the previous time step. The input feature vector is denoted as... The hidden state of the previous time step is recorded as The two are concatenated to generate a concatenated feature vector. Multiply the concatenated feature vector by the update gate weight matrix, denoted as . The updated gate linear combination value is obtained. A nonlinear activation operation is applied to the linear combination value of the update gate, using the sigmoid function as the nonlinear activation function, to generate the activation value of the update gate. Multiply the concatenated feature vector by the reset gate weight matrix, denoted as . The linear combination value of the reset gate is obtained. A nonlinear activation operation is applied to the linear combination value of the reset gate, using the sigmoid function as the nonlinear activation function, to generate the reset gate activation value. .

[0053] In some embodiments, when the slope value of the sealing surface degradation trend curve at the current time step is synchronously input into the hidden layer of the recurrent gating network as a state adjustment factor, the state adjustment factor... The process of embedding it into the computation of the candidate hidden state is specifically done by... The features modulated by the reset gate activation value are concatenated and then input into the candidate hidden state generation unit.

[0054] In some embodiments, when modulating the candidate hidden state according to the updated gate activation value and the reset gate activation value to generate the encoded hidden state at the current time step, the candidate hidden state is calculated as follows:

[0055] in: The weight matrix representing the candidate hidden state. This represents the weight vector of the state adjustment factor. This represents element-wise multiplication. The hyperbolic tangent activation function is used, and then the encoded hidden state at the current time step is calculated based on the update gate activation value. Optional, the parameter configuration for each layer of the cyclic gating network is shown in Table 2.

[0056] Table 2: Parameter Configuration Table for Each Layer of the Cyclic Gated Network

[0057] In this context, the input dimension 32 of the input layer corresponds to the feature dimension of the stress-load coupling feature matrix at each time step. The input dimension 96+96 of the hidden layer represents the concatenation of the dimension of the input feature vector at the current time step after linear transformation and the dimension of the hidden state at the previous time step. The output dimension 2 of the fully connected output layer corresponds to the predicted mean m and the predicted scale parameter b, respectively. It can be understood that in each time step, the gate activation values ​​are updated. Control the hidden state of the previous time step The proportion passed to the current time step resets the gate activation value. Control the hidden state of the previous time step Hidden state of candidates The degree of influence of computation is assessed, and the two parties achieve adaptive modeling of temporal dependencies through a gating mechanism.

[0058] In one embodiment of the present invention, when determining the expected failure time window of the valve based on the probability distribution of the valve's remaining life and marking the expected failure time window as the life prediction result, the specific operation is as follows: extracting the predicted mean and predicted variance from the probability distribution of the valve's remaining life, and constructing an initial confidence interval centered on the predicted mean; adjusting the width of the initial confidence interval according to the magnitude of the predicted variance to generate the expected failure time window; verifying whether the lower boundary of the expected failure time window is greater than zero, and if the lower boundary is less than or equal to zero, correcting the lower boundary to a preset minimum positive value; marking the upper and lower boundaries of the corrected expected failure time window as the latest predicted failure time and the earliest predicted failure time of the valve, respectively; and organizing the latest predicted failure time and the earliest predicted failure time together into the life prediction result.

[0059] In specific implementation, when determining the expected failure time window of the valve based on the probability distribution of the valve's remaining life and marking the expected failure time window as the life prediction result, the predicted mean and predicted variance are extracted from the probability distribution of the valve's remaining life. The predicted mean is denoted as μ, and the predicted variance is denoted as σ². An initial confidence interval is constructed with the predicted mean μ = 127.5 days as the center, with a confidence level of 95%. The corresponding confidence interval width factor is 1.96, and the lower limit of the initial confidence interval is μ - 1.96σ, and the upper limit is... The limit is μ + 1.96σ, where σ is the square root of the prediction standard deviation (i.e., the prediction variance). σ = 12 days, and the initial confidence interval is calculated to be [127.5 - 1.96 × 12, 127.5 + 1.96 × 12] = [104.0, 151.0] days. The width of the initial confidence interval is adjusted according to the magnitude of the prediction variance to generate the expected failure time window. The adjustment method is to multiply the width of the initial confidence interval by a scaling factor proportional to the prediction variance. The scaling factor is calculated using the following formula:

[0060] Where: γ represents the width adjustment coefficient. This represents the predicted variance extracted from the probability distribution of the valve's remaining life, and μ represents the predicted mean extracted from the probability distribution of the valve's remaining life. =144, μ=127.5, calculated to The initial confidence interval width of 47.0 days is multiplied by γ to obtain an adjusted width of 47.4 days. An adjusted expected failure time window is generated with the predicted mean of 127.5 days as the center, with a lower limit of 127.5 - 23.7 = 103.8 days and an upper limit of 127.5 + 23.7 = 151.2 days. The lower boundary of the expected failure time window is verified to be greater than zero. The lower boundary of 103.8 days is greater than zero and no correction is needed. The upper and lower boundaries of the corrected expected failure time window are marked as the latest predicted failure time and the earliest predicted failure time of the valve, respectively. The upper boundary of 151.2 days is marked as the latest predicted failure time, and the lower boundary of 103.8 days is marked as the earliest predicted failure time. The latest predicted failure time of 151.2 days and the earliest predicted failure time of 103.8 days are combined to form the lifetime prediction result, which includes a time interval [103.8 days, 151.2 days].

[0061] Optionally, when extracting the predicted mean and predicted variance from the valve's remaining life probability distribution, the predicted mean μ represents the central estimate of the valve's remaining life, and the predicted variance σ² represents the degree of uncertainty in the valve's remaining life estimate; a larger predicted variance indicates higher uncertainty. In some embodiments, when adjusting the width of the initial confidence interval to generate the expected failure time window based on the magnitude of the predicted variance, if the predicted variance is less than one-tenth of the predicted mean, the width adjustment coefficient γ is set to 1, meaning the initial confidence interval width is not adjusted; if the predicted variance is greater than half of the predicted mean, the width adjustment coefficient γ is set to 1.5, meaning the initial confidence interval width is increased by a factor of 1.5.

[0062] In specific implementation, when verifying whether the lower boundary of the expected failure time window is greater than zero, if the lower boundary is less than or equal to zero, the lower boundary is corrected to a preset minimum positive value, which is set to 0.5 days. That is, when the calculated lower boundary is less than or equal to 0 days, the lower boundary is forcibly set to 0.5 days to ensure that the time window boundary in the life prediction result is a valid positive number. Optionally, when the upper and lower boundaries of the corrected expected failure time window are marked as the latest predicted failure time and the earliest predicted failure time of the valve, respectively, the upper boundary corresponds to the upper limit estimate of the valve's remaining life, and the lower boundary corresponds to the lower limit estimate of the valve's remaining life. The two boundaries together define the time range in which valve failure may occur.

[0063] In some embodiments, when the latest predicted failure time and the earliest predicted failure time are combined to form the lifetime prediction result, the lifetime prediction result is output in JSON data format, containing four fields: the earliest predicted failure time field is a corrected lower boundary value, the latest predicted failure time field is a corrected upper boundary value, the predicted mean field is the predicted mean extracted from the probability distribution, and the confidence level field is the confidence level percentage used when constructing the initial confidence interval. It can be understood that the predicted mean and predicted variance extracted from the valve's remaining lifetime probability distribution are the basic parameters for generating the expected failure time window. The predicted mean determines the center position of the time window, and the predicted variance determines the initial width of the time window. It can be understood that adjusting the width of the initial confidence interval according to the magnitude of the predicted variance allows the expected failure time window to adaptively reflect the magnitude of prediction uncertainty; the time window widens when the predicted variance is large and narrows when the predicted variance is small, thereby expressing the degree of confidence in the lifetime prediction result.

[0064] The present invention has been described in detail above through embodiments, but the content described is only an exemplary embodiment of the present invention and should not be considered as limiting the scope of the present invention. The scope of protection of the present invention is defined by the claims. Any technical solutions designed by those skilled in the art using the technical solutions described in the present invention, or similar technical solutions designed by those skilled in the art under the inspiration of the technical solutions of the present invention, within the substance and scope of protection of the present invention, to achieve the above-mentioned technical effects, or equivalent changes and improvements made to the scope of the application, should still fall within the patent protection scope of the present invention. It should be noted that, for clarity, descriptions of some components and processes that are not directly and obviously related to the scope of protection of the present invention but are known to those skilled in the art have been omitted in the description of the present invention.

Claims

1. A method for predicting the lifespan of pipeline valves based on finite element analysis, characterized in that, The method includes: Step 1: Obtain a multi-time series monitoring data set of the target pipeline valve during its service life. The multi-time series monitoring data set includes time series data of valve body stress distribution, cumulative data of sealing surface wear, and time series of medium pressure fluctuations. Step 2: Perform finite element mesh node mapping processing on the time series data of the stress distribution of the valve body to generate the valve stress field tensor sequence; Step 3: Perform piecewise fitting processing on the cumulative wear data of the sealing surface to obtain the degradation trend curve of the sealing surface; Step 4: Perform frequency domain decomposition on the medium pressure fluctuation time series to extract the pressure cyclic load spectrum; Step 5: Align the valve stress field tensor sequence with the pressure cyclic load spectrum along the time axis to generate a stress-load coupling feature matrix; Step 6: Call the pre-trained recurrent gating network to perform joint temporal encoding on the stress-load coupling feature matrix and the sealing surface degradation trend curve, and output the valve remaining life probability distribution; Step 7: Determine the expected failure time window of the valve based on the probability distribution of the valve's remaining lifespan, and mark the expected failure time window as the lifespan prediction result.

2. The method according to claim 1, characterized in that, Step 2 specifically includes: Finite element meshing is performed on each frame of the stress cloud map in the time series data of the stress distribution of the valve body to obtain the stress component value corresponding to each mesh node; Extract the three principal stress direction values ​​of each grid node from the stress component values, and combine the three principal stress direction values ​​into a node stress vector; The nodal stress vectors of all mesh nodes are arranged in a matrix according to the mesh topology to form a single-frame stress field tensor; The stress field tensors of each frame at each time step are stacked in order to generate the valve stress field tensor sequence.

3. The method according to claim 1, characterized in that, Step 3 specifically includes: The cumulative wear data of the sealing surface is sorted according to the acquisition time sequence to obtain the wear time sequence; Perform a sliding window difference operation on the wear amount time series to calculate the instantaneous wear rate value within each time window; The wear amount time series is segmented according to the change range of the instantaneous wear rate value to obtain multiple wear stage sub-sequences; For each wear stage subsequence, least squares linear fitting is performed to obtain the stage wear rate straight line corresponding to each wear stage; The wear rate lines corresponding to each wear stage are spliced ​​at the segmentation points to generate the degradation trend curve of the sealing surface.

4. The method according to claim 1, characterized in that, Step 4 specifically includes: The medium pressure fluctuation time series is pre-processed to remove the static pressure baseline component, resulting in a dynamic pressure fluctuation series. Apply a discrete Fourier transform to the dynamic pressure fluctuation sequence to generate a pressure fluctuation spectrum. In the pressure fluctuation spectrum, discrete frequency components with amplitudes exceeding the background noise threshold are identified to obtain the set of dominant frequency components. Extract the waveform amplitude and phase angle corresponding to each discrete frequency component from the set of main frequency components, and combine the waveform amplitude and phase angle into a cyclic load parameter pair; The pressure cyclic load spectrum is generated by arranging all cyclic load parameter pairs in ascending order of frequency.

5. The method according to claim 1, characterized in that, Step 5 specifically includes: Extract the global stress extremum point and its grid node position corresponding to each moment in the valve stress field tensor sequence; Extract the peak pressure and valley pressure corresponding to each pressure cycle in the pressure cycle load spectrum; The global stress extremum points are paired with the peak pressures in the pressure cyclic load spectrum in chronological order to obtain a stress-pressure peak pairing sequence. Spatially correlate the grid node location of the global stress extreme point with the valve load area corresponding to the peak pressure to obtain a spatial correlation marker; The stress-pressure peak pairing sequence is merged and encoded with the spatial correlation marker to generate the stress-load coupling feature matrix.

6. The method according to claim 1, characterized in that, Step 6 specifically includes: The stress-load coupling feature matrix is ​​expanded into a sequence of input feature vectors over time steps, and the input feature vectors of each time step are fed into the input layer of the recurrent gated network in sequence. The slope value of the degradation trend curve of the sealing surface at the current time step is used as a state adjustment factor and synchronously input into the hidden layer of the cyclic gating network. In each time step of the recurrent gating network, the updated gate activation value and the reset gate activation value are calculated based on the input feature vector and the hidden state of the previous time step. The candidate hidden state is modulated based on the updated gate activation value and the reset gate activation value to generate the encoded hidden state at the current time step; The encoded hidden state of the last time step is input into the fully connected output layer of the recurrent gating network to obtain the predicted mean and predicted variance of the valve's remaining lifespan. The predicted mean and predicted variance are then combined to form the probability distribution of the valve's remaining lifespan.

7. The method according to claim 6, characterized in that, In step 6, at each time step of the recurrent gating network, the updated gate activation value and the reset gate activation value are calculated based on the input feature vector and the hidden state of the previous time step, specifically including: The input feature vector is concatenated with the hidden state of the previous time step to generate a concatenated feature vector. Multiply the concatenated feature vector by the update gate weight matrix to obtain the linear combination value of the update gate; A nonlinear activation operation is applied to the linear combination value of the update gate to generate the update gate activation value; Multiply the spliced ​​feature vector by the reset gate weight matrix to obtain the reset gate linear combination value; A nonlinear activation operation is applied to the linear combination value of the reset gate to generate the reset gate activation value.

8. The method according to claim 1, characterized in that, Step 7 specifically includes: The predicted mean and predicted variance are extracted from the probability distribution of the valve's remaining life, and an initial confidence interval is constructed with the predicted mean as the center. The width of the initial confidence interval is adjusted according to the magnitude of the predicted variance to generate the expected failure time window; Verify whether the lower boundary of the expected failure time window is greater than zero. If the lower boundary is less than or equal to zero, then correct the lower boundary to the preset minimum positive value. The upper and lower boundaries of the revised expected failure time window are marked as the latest and earliest predicted failure times of the valve, respectively. The latest predicted failure time and the earliest predicted failure time are combined to form the lifetime prediction result.

9. The method according to claim 1, characterized in that, Step 1 further includes data cleaning of the multi-time-series monitoring data set to obtain a cleaned monitoring data set, specifically including: Abnormal stress peaks are detected in the time series data of stress distribution in the valve body. Abnormal stress frames that exceed the upper limit of stress range are removed to obtain stress data after cleaning. Missing values ​​are located in the cumulative wear data of the sealing surface, and the wear values ​​at the missing time points are filled in using a linear interpolation method to obtain complete wear data. The pressure fluctuation time series of the medium is subjected to high-frequency noise filtering, and pressure spikes are removed by a sliding median filtering algorithm to obtain smooth pressure fluctuation data. The post-cleaning stress data, complete wear data, and smoothed pressure fluctuation data are combined into the post-cleaning monitoring data set.

10. The method according to claim 1, characterized in that, Before aligning the valve stress field tensor sequence with the pressure cyclic load spectrum along the time axis in step 5, the process further includes dimensionality reduction and compression of the valve stress field tensor sequence, specifically including: Perform a matrix flattening operation on each single frame stress field tensor in the valve stress field tensor sequence to obtain the stress field long vector. Calculate the covariance matrix among all stress field long vectors, and perform eigenvalue decomposition on the covariance matrix to obtain the eigenvalue sequence and eigenvector matrix; The eigenvectors corresponding to the eigenvalues ​​whose cumulative contribution rate exceeds a preset retention threshold are selected from the eigenvalue sequence to form the projection basis matrix; Perform matrix multiplication on each stress field long vector and the projection basis matrix to obtain the dimension-reduced stress field eigenvectors. All the dimensionality-reduced stress field eigenvectors are rearranged according to the original temporal order to generate a dimensionality-reduced and compressed valve stress field tensor sequence.

Images