High-temperature steam turbine valve safety detection state reconstruction method

Abstract

The application discloses a high-temperature steam turbine valve safety detection state reconstruction method, relates to the technical field of steam turbine valve operation monitoring and intelligent diagnosis, and is used for solving the problem that the valve safety state is difficult to be reliably reconstructed and warned under the condition of asynchronous multi-source monitoring data, noise missing interference, etc.; multi-source data such as valve position, driving pressure, valve front pressure or temperature, valve body metal temperature and vibration are collected, time reference alignment, abnormality and missing processing and scale unification are performed, features such as action consistency, lag, friction abnormality and leakage trend are constructed, and representation quantities are generated; under the constraint of mechanism boundary and continuity, the action of the valve, the sealing and friction state are fused and estimated, and after consistency checking, the risk level and warning information are output.

Description

Technical Field

[0001] This invention relates to the field of turbine valve operation monitoring and intelligent diagnosis technology, and more specifically, to a method for reconstructing the safety detection status of high-temperature turbine valves. Background Technology

[0002] Steam turbine high-temperature valves operate under prolonged high-temperature, high-pressure, and frequent regulation environments. Issues such as sluggish operation, jamming, seal degradation, and leakage trends can lead to decreased efficiency and safety risks. Existing online monitoring systems often rely on single signal thresholds or empirical judgments, making it difficult to simultaneously interpret multi-channel information such as valve position, driving pressure, pressure-temperature, and vibration. Furthermore, different sampling frequencies and inconsistent timestamps across channels, along with common field issues like noise spikes, communication jitter, and missing segments, can cause misjudgments when directly fused. On the other hand, valve states have physical constraints and evolutionary continuity; relying solely on data-driven models can easily lead to state drift or misinterpret long-term biases as degradation.

[0003] To address the above problems, this invention proposes a solution. Summary of the Invention

[0004] In order to overcome the above-mentioned defects of the prior art, embodiments of the present invention provide a method for reconstructing the safety detection status of valves in high-temperature steam turbines to solve the problems mentioned in the background art.

[0005] To achieve the above objectives, the present invention provides the following technical solution: In a preferred embodiment, it includes: Acquire multi-source running data and save channel identifiers and sampling timestamps. Construct a common timeline based on the timestamps, align the original sampling sequences of each channel, and generate a benchmark multi-source running dataset. Based on sampling frequency records or timestamp statistics and valve position opening and closing process records or operating procedures, time windows are determined. Within each window, safety detection characteristics are calculated and scaled and linearly weighted to generate valve action deviation characterization, valve hysteresis characterization, valve friction anomaly characterization, and valve leakage trend characterization. Within each time window, an in-window adjustment quantity is constructed using a benchmarked multi-source operational dataset and a valve safety status characterization quantity. Based on the in-window adjustment quantity, the observation mapping relationship is extended to a window adaptive mapping and updated at the window level. The observation bias correction quantity is generated by smoothing statistics of the observation residuals through bounded recursion. Using a saturation suppression loss function, the continuity and sparsity variation constraint coefficients are updated with the in-window adjustment quantity within the bounded optimization framework, and the valve action state, valve sealing state, and valve friction state are obtained recursively. The valve safety detection status reconstruction sequence is subjected to boundary consistency verification and observation consistency verification according to time window. Based on the set of reliable windows, the safety risk level and early warning information are fused according to safety criteria, and the time window identifier and threshold rule source type are output.

[0006] In a preferred embodiment, multi-source operating data such as valve position signal, actuator drive pressure signal, valve inlet pressure signal, valve inlet temperature signal, valve body metal temperature signal and vibration signal are acquired, and the channel identifier and sampling timestamp of a unified clock source are saved for each sampling point; a common time axis is constructed based on the timestamp, and the original sampling sequence of each channel is resampled to the common time axis by interpolation.

[0007] In a preferred embodiment, a sliding window robust statistical method is used to identify outliers in the aligned sequence and replace them with the midpoint of the window or by interpolation with adjacent valid points. At the same time, missing segments are identified based on the timestamp interval and interpolation is used to fill in or mark the missing intervals based on the duration of the missing data. Within the valid data interval, the mean and standard deviation or statistical upper and lower limits are calculated for each channel and scaled according to standardization or interval normalization. The processed synchronization sequences of each channel are aggregated according to a common time axis to generate a benchmarked multi-source running dataset.

[0008] In a preferred embodiment, the window length and sliding step size are determined based on sampling frequency records or timestamp statistics, valve position opening and closing process records, or operating procedures. The benchmarked multi-source operating dataset is divided into continuous time windows along a common time axis, and valve safety detection characteristics are calculated within each window using a uniform caliber. Within each window, action consistency characteristics are calculated based on the synchronization relationship between valve position and valve stem displacement or driving pressure; hysteresis characteristics are calculated based on the time shift of the cross-correlation peak between driving pressure and valve position; friction anomaly characteristics are obtained by calculating energy or root mean square based on vibration signals and optional acoustic emission signals; and leakage trend characteristics are obtained by calculating the valve inlet pressure decay rate or dispersion after selecting a closed or small opening time segment based on the valve position threshold. Then, each feature within the window is scaled according to the benchmark parameters obtained from historical normal operation or debugging data statistics, and linearly weighted according to traceable and determined weights to generate valve action deviation characterization, valve hysteresis characterization, valve friction anomaly characterization, and valve leakage trend characterization. When channel data is missing, relevant features are removed and the remaining weights are normalized.

[0009] In a preferred embodiment, an in-window adjustment quantity constructed solely from benchmarked multi-source operating data and valve safety status characterization quantities is introduced within each time window. This is achieved by using the time proportion of the valve position signal within the closed or small opening interval to form the valve closing constraint confidence, the difference between the peak and average intensity of the friction anomaly characterization quantities to form the friction impact intermittency, and the statistical analysis of the in-window standard deviation and rate of change of the valve inlet pressure and temperature signals to form the operating condition fluctuation intensity. Under the traceability constraints of the mechanism boundary parameters, valve position threshold, and stroke calibration parameters, the observation mapping relationship, observation bias correction quantity, residual suppression form, and state evolution constraint intensity are updated in a window-level linkage manner.

[0010] In a preferred embodiment, the observation mapping relationship is extended from a fixed mapping to a window adaptive mapping that varies with the adjustment amount, and the mapping intensity is adjusted within the historical normal statistical range or the drift range allowed by the procedure. The observation bias correction amount is obtained by bounded recursion through smoothed statistics of the observation residuals of the current window and the historical window to absorb low-frequency stable offset components. The residual term adopts a saturated suppression loss function that imposes a quadratic penalty on small residuals and a linear penalty on large residuals. The state solution is performed within a bounded optimization framework of observation consistency term, adjacent window continuity term, and sparse change term, and the continuity and sparse change constraint coefficients are updated with the adjustment amount. By adopting differentiated mapping adjustment, bias recursion, and constraint coefficient update paths for different modes such as synchronous increase, single anomaly, slow trend, and intermittent enhancement in different windows, a recursive reconstruction sequence of valve action state, valve sealing state, and valve friction state is generated.

[0011] In a preferred embodiment, the valve safety detection state reconstruction sequence is sequentially checked for consistency across time windows, and safety criteria conclusions are output: the upper and lower bounds of the state value range and the upper limit of the rate of change of the state in adjacent windows. For each window, the valve action state, valve sealing state, and valve friction state are checked for out-of-bounds and differential out-of-limit checks, marking out-of-bounds windows as unreliable. Then, based on the observation vector and observation mapping relationship, the observed predicted value is calculated from the state estimate within the window, and the observed residual is formed with the actual observed characteristic quantity. A threshold derived from the statistical distribution of historical normal operation residuals is used to determine if the residual exceeds the limit, and the out-of-limit window is then... Mark as untrusted; after obtaining the set of trustworthy windows, set the judgment thresholds and judgment rules from the design allowable range, operating procedures or normal operation statistical intervals for valve action state, valve sealing state and valve friction state respectively, and fuse the judgment results of the three types of states according to the priority of operating procedures or safety protection logic to obtain the safety risk level. At the same time, introduce the time consistency judgment of multiple consecutive windows meeting the triggering conditions or release conditions; when outputting the warning information, give the triggering time window identifier, the name of the triggered state component, the source type of the threshold rule and the estimated value of the state of the window, and output the untrustworthy identifier and reason category for untrustworthy windows.

[0012] The technical effects and advantages of this invention based on the high-temperature steam turbine valve safety detection state reconstruction method are as follows: This invention takes comparability, reliability, and interpretability as its main principles: First, it uses a unified time axis resampling combined with robust anomaly or missing value strategies to significantly reduce the disturbance of asynchrony and noise to diagnosis; then, it quantifies action deviation, lag, frictional impact, and the ability to maintain pressure in the closing segment into equivalent characterization quantities, so that the state estimation corresponds one-to-one with the specific risk mechanism; in the reconstruction stage, it introduces adaptive mapping or bias correction of value range, rate of change, and window to suppress spurious degradation caused by long-term offset; finally, it filters out unreliable windows through dual verification of boundary and observation consistency, outputting a more stable risk level and triggering basis, thereby improving the accuracy of early warning and engineering reproducibility. Attached Figure Description

[0013] Figure 1 This is a schematic diagram of the arrangement of high-temperature steam turbine valves and measuring points in the high-temperature steam turbine valve safety detection state reconstruction method of the present invention.

[0014] Figure 2 This is an example diagram of frictional anomaly characteristics in the safety detection state reconstruction method for high-temperature steam turbine valves based on the present invention. Detailed Implementation

[0015] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0016] Example This invention discloses a method for reconstructing the safety detection status of valves in high-temperature steam turbines, such as... Figure 1 As shown, it includes: Step 1: Acquisition of multi-source runtime data and time-series benchmarking; First, multi-source operational data acquisition is performed. For the same high-temperature turbine valve, operational data directly related to valve safety monitoring is acquired, with each type of data source limited to the actual collected values ​​from the corresponding measuring device or control feedback channel. The operational data includes at least: valve position signal, actuator drive pressure signal, valve inlet pressure signal and valve inlet temperature signal, valve body metal temperature signal, and vibration signal. When acquiring the above data, the channel identifier and sampling timestamp are simultaneously saved for each sampling point, where the sampling timestamp originates from the same control clock or the same acquisition clock. Secondly, a unified time reference alignment is performed. Since the sampling frequency and sampling time of each channel may be inconsistent, this step constructs a common timeline based on the timestamp generated by the acquisition clock, and resamples the original sampling sequence of each channel onto the common timeline. Resampling is implemented using interpolation, preferably linear interpolation: when a time on the common timeline falls between two adjacent actual sampling times of a channel, the alignment value for that time is calculated using the actual sampling values ​​of the two adjacent points. The calculation formula is as follows: ; in , For this channel at adjacent sampling times , The actual measured value, y(t) is the alignment value of time t on the common time axis, and ≤t≤ This alignment operation enables different channels to have comparable synchronous observations at the same time on a common time axis.

[0017] Next, data quality processing and scaling are performed. The aligned channel sequence may contain outliers caused by sensor glitches, communication jitter, or transient interference. This step uses robust statistical discrimination to replace outliers, preferably Hampel discrimination: within a sliding window, the window midpoint m and the median absolute deviation MAD are calculated, and then... ; As an anomaly criterion; where y(t) is the alignment value to be judged, m and MAD are calculated from the alignment value within the window, and λ is the threshold coefficient. The threshold coefficient λ is not specified arbitrarily, but is determined by the fluctuation range obtained from the statistics of historical normal operation data or debugging data; when historical data is lacking, an initial sample value can be given first and updated according to the statistics of normal samples during operation, but each update is based on traceable statistical results. For sampling points judged as abnormal, this step adopts one of two equivalent replacement rules: replacement with the window midpoint, or replacement with the value obtained by interpolation of the nearest valid sampling point before and after the abnormal point; the replacement values ​​all come from the set of real sampling points of the same channel; For handling missing data points, this step determines whether to interpolate and fill in or remove the interval based on the duration of the missing data. The duration of the missing data is directly determined by an abnormal increase in the timestamp interval. When the missing segment does not exceed the upper limit of the interruption duration allowed by the acquisition strategy or operating procedures, it is filled in using interpolation with adjacent valid points. When the missing segment exceeds this upper limit, the missing time interval is marked as unusable and removed from the dataset to avoid replacing the actual acquired value with a long-term inferred value. The upper limit of the allowed interruption duration is derived from the acquisition system configuration or operating procedures. After handling anomalies and missing data, this step performs a scaling transformation on each channel to avoid numerical incomparability between channels with different dimensions. Scaling transformation preferably employs standardization: for each channel, the mean and standard deviation are calculated within the valid data interval. Both the mean and standard deviation are obtained from the actual sampling points of that channel, and the channel's sampled values ​​are transformed by subtracting the mean and then dividing by the standard deviation. For channels with naturally occurring upper and lower limits, interval normalization can also be used, with the upper and lower limits obtained from the actual sampling points within the valid interval. All statistics used for standardization or normalization are calculated in this step and saved along with the processing records. Finally, the synchronization sequences of each channel, after time alignment, anomaly handling, missing data handling and scale consistency, are aggregated according to a common time axis to form a benchmark multi-source running dataset; Step 2: Construction of safety detection features and generation of valve safety status characterization quantities; First, time windows are determined, and segmented calculations are performed on the benchmarked multi-source operational dataset. The time window is defined by both the window length and the sliding step size, both derived from traceable engineering information: the sampling frequency comes from acquisition records or timestamp sequence statistical results, and the duration range of the valve action process comes from actual changes in the valve position signal during opening and closing or the action time range specified in the operating procedures. This method ensures that each time window contains sufficient sampling points and covers the critical stages of the valve action, thus providing stable statistical significance for the safety detection features calculated within the window. After determining the time windows, the benchmarked multi-source operational dataset is divided into continuous window sequences along a common time axis, and the same feature calculation is performed on the synchronous data of each channel within each window.

[0018] Within each time window, a set of safety detection features is constructed based on typical risk mechanisms of valve safety detection. To characterize whether inconsistencies in valve operation are caused by transmission chain loosening, hysteresis, or jamming, this step calculates valve operation consistency features. The data source for these features is the synchronous sequence of valve position signal and valve stem displacement signal within the same window. When no valve stem displacement signal is configured on-site, the valve operation consistency features are calculated based on the synchronous relationship between the valve position signal and the actuator drive pressure signal, but the data source is still limited to the acquired channel data. The preferred methods for calculating valve operation consistency features are intra-window correlation calculation and intra-window rate of change consistency calculation. Correlation calculation quantifies whether the two sequences change synchronously, while rate of change consistency calculation quantifies whether the ratio of the change amplitudes of the two sequences is stable. To characterize whether there is a delay or hysteresis in the valve response relative to the drive input, this step calculates the valve hysteresis characteristic. The data source for this characteristic is the synchronous sequence of the actuator drive pressure signal and the valve position signal within the same window. The calculation method uses cross-correlation peak time shift estimation. The cross-correlation function is defined as: ; in, This represents the k-th sampled value of the actuator-driven pressure signal sequence within the window. This represents the k-th sampled value of the valve position signal sequence within the window. and denoted by , respectively, represent the mean of the corresponding sequence within this window, and τ is the discrete-time offset. This is the cross-correlation value. This step takes... The largest As a valve hysteresis characteristic value, the physical meaning of the valve hysteresis characteristic value is the relative time delay of the transmission of the drive change to the valve position response; like Figure 2 As shown, to characterize whether there are anomalies such as increased friction, jamming release impact, or flow-induced impact during valve operation, this step calculates friction anomaly characteristics. The data source for these characteristics is vibration signals, and acoustic emission signals are further introduced as an enhanced data source when configuring acoustic emission signals. Simultaneously, valve position signals can be combined to limit the time segment of the valve's operation phase to avoid misinterpreting steady-state background vibration as friction anomalies. The calculation method for friction anomaly characteristics employs window energy calculation or root mean square (RMS) calculation. Specifically, the RMS value or energy accumulation of the vibration signal within the window is calculated to quantify the overall intensity of impact and high-frequency components within that window. When configuring acoustic emission signals, the energy accumulation or event count rate of the acoustic emission signal within the window is calculated to quantify the level of high-frequency abnormal activity. The inputs for the above characteristics are all actual sampling points within the window, and the output is the friction anomaly characteristic value. An increase in the characteristic value indicates an increase in the intensity of abnormal activity related to friction or impact.

[0019] To characterize the medium flow trend of a valve in a closed or near-closed state, thereby identifying leakage risks caused by valve seat seal degradation or valve core wear, this step calculates leakage trend characteristics. The data sources for these characteristics are at least the valve position signal and the upstream pressure signal. When a downstream pressure signal or flow signal is available, it can be used as an enhanced data source without changing the definition of the characteristic. The leakage trend characteristics are calculated using pressure holding capability assessment under valve position constraints: First, this step determines the time segment in which the valve is in the closed or small-opening range based on the valve position signal. The valve position threshold for the closed or small-opening range is derived from the valve stroke calibration results or the valve position calibration parameters of the control system. Then, within this time segment, the decay rate or fluctuation amplitude of the upstream pressure signal is calculated as the leakage trend characteristic value. The decay rate is calculated from the change of the upstream pressure signal over time within the same time segment, and the fluctuation amplitude is calculated from the statistical dispersion of the upstream pressure signal within the same time segment. By constraining the valve position conditions as described above, leakage trend characteristics are only evaluated when the valve meets the closing conditions, thereby avoiding misjudging changes in operating conditions as leakage trends when the valve is in a normal adjustment state.

[0020] After calculating various safety detection features, this step further generates a valve safety status characterization quantity and provides a unified and traceable basis for its generation method. In this scheme, the valve safety status characterization quantity is defined as a quantitative index obtained by combining multiple safety detection features within the same time window after homoscalation processing and weighting. Specifically, this step first performs homoscalation processing on each safety detection feature. The benchmark parameters required for homoscalation processing are derived from the statistical results of historical normal operation data or debugging data. The historical normal operation data or debugging data comes from the operation records of the same valve or valve of the same type under confirmed normal conditions, and the benchmark parameters are obtained by statistically analyzing the mean, standard deviation, or quantile of the corresponding feature on normal samples, thereby ensuring that the homoscalation parameters are traceable and consistent with the equipment. After completing the homoscalation processing, this step generates the valve safety status characterization quantity using a linear weighting method, the expression of which is: ; Where S represents a quantity representing the safety status of a certain valve. This represents the i-th security detection feature after being scaled up. This represents the corresponding weight, and n represents the number of features involved in the combination. Weight The acquisition method is as follows: when labeled abnormal sample data is available, the weights are determined by minimizing the classification error between abnormal and normal states by the representation quantity or by maximizing the feature discriminative power; when abnormal sample data is lacking, the initial weight values ​​are given based on the importance of the valve mechanism, and then the weights are adjusted based on the statistical stability of normal operation data, with each adjustment based on traceable statistical results. Through the above generation method, this step compresses a large number of features into a small number of valve safety state representation quantities, so that the same valve can be compared using the same set of representation quantities in different time windows.

[0021] It should be noted that, based on the unified generation logic described above, this step generates at least the following valve action deviation characterization quantities: valve hysteresis characterization quantity, valve friction anomaly characterization quantity, and valve leakage trend characterization quantity. The valve action deviation characterization quantity is composed of valve action consistency features and is used to characterize the degree of inconsistency between valve position response and displacement or drive input. The valve hysteresis characterization quantity is composed of valve hysteresis features obtained by time-shifting the cross-correlation peak and is used to characterize the delay from drive to response. The valve friction anomaly characterization quantity is composed of friction anomaly features calculated from vibration signals and acoustic emission signals and is used to characterize the intensity of frictional impact activity. The valve leakage trend characterization quantity is composed of leakage trend features obtained from the pressure holding capacity assessment under valve position constraints and is used to characterize the leakage-related trend intensity under closed conditions. If data for a certain channel is missing on-site, the relevant features of that channel are removed without introducing external data, and the weights of the remaining features are normalized to maintain consistency in the calculation of the characterization quantities.

[0022] Step 3: Data-driven state reconstruction calculation based on mechanistic constraints; First, the sources of input data and parameters for state reconstruction are clarified. The input data for state reconstruction includes: the synchronization sequence of the benchmark multi-source operational dataset from step one within the current time window, and the values ​​of valve action deviation, valve hysteresis, valve friction anomaly, and valve leakage trend from step two within the current time window. Each channel of the benchmark multi-source operational dataset originates from the actual acquired values ​​of the corresponding sensor or control feedback channel, after time alignment and quality processing. The valve safety state representation is derived from the same-scale and weighted combination calculation results of safety detection features. The input parameters for state reconstruction include: valve stroke calibration parameters, valve position thresholds for the valve closing or small opening intervals, and mechanistic boundary parameters used to limit the rate of state change. The valve stroke calibration parameters and valve position thresholds are derived from valve calibration records or control system calibration parameters, while the mechanistic boundary parameters are derived from the equipment design allowable range, operating procedures, or historical normal operation statistics. Subsequently, a state variable expression for the valve safety detection state is established, along with observation constraints, to ensure that state reconstruction has a clear state-observation correspondence. To avoid introducing unexplained new terms, this step directly represents the valve safety detection state as a state vector, which consists of three components corresponding to the valve action state, valve sealing state, and valve friction state. The observation basis for the valve action state is the valve action deviation characterization quantity and the valve hysteresis characterization quantity. This is because the valve action deviation characterization quantity reflects the degree of inconsistency between the valve position response and the displacement or drive input, while the valve hysteresis characterization quantity reflects the intensity of the time delay from the actuator drive change to the valve position response. The observation basis for the valve sealing state is the valve leakage trend characterization quantity, because the valve leakage trend characterization quantity is obtained from the pressure holding capacity assessment under valve position constraints, directly reflecting the intensity of the leakage-related trend under closing conditions. The observation basis for the valve friction state is the valve friction anomaly characterization quantity, because the valve friction anomaly characterization quantity is composed of the energy or event intensity characteristics of vibration signals and acoustic emission signals, reflecting the intensity of frictional impact activity. The purpose of the above correspondence is to ensure that each state component has clear observational support. After establishing the state variables and observation basis, this step applies mechanistic constraints to the state evolution process to ensure that the reconstruction results meet the physical rationality of valve operation. These mechanistic constraints include at least state change rate constraints and state value range constraints: the state change rate constraint limits the change in the valve's safety detection state between adjacent time windows to no more than the rate of change achievable by the equipment. This rate of change is based on the duration range of the valve's action process, the thermal inertia characteristics of the valve body metal temperature change, and historical normal operation statistics. The state value range constraint ensures that the valve's action state, valve sealing state, and valve friction state are within a predefined allowable range. This allowable range is based on the valve's design-allowed operating conditions, operating procedures, or confirmed normal sample statistical intervals.

[0023] After setting the mechanism constraints, this step uses a data-driven fusion estimation method to recursively reconstruct the valve safety detection state. The data-driven fusion estimation method is defined in this scheme as follows: taking the state vector as the estimated object, the valve safety state characterization quantity as the observation quantity, and the mechanism constraints as the boundary conditions, the state estimate is obtained by minimizing the deviation between the predicted state and the observed characterization quantity.

[0024] It should be noted that during the long-term operation of high-temperature steam turbine valves, the inlet temperature signal and the valve body metal temperature signal indicate that the heat load is in a slowly changing range. The inlet pressure signal exhibits periodic fluctuations caused by load adjustments. The valve position signal shows repeatable slight hysteresis in several small movements. The vibration signal intermittently increases during the valve's movement phase when it approaches the closed range, but the increase is not sustained and does not occur with every movement. According to the calculation results in step two, within adjacent time windows, there are windows where the valve hysteresis characteristic and the valve movement deviation characteristic increase synchronously, as well as windows where only the valve friction abnormality characteristic increases while other characteristics remain normal. Additionally, there are windows where the valve leakage trend characteristic shows a slow increase within the valve's closed or small opening range, but does not reach a significant limit.

[0025] Therefore, in this embodiment, several in-window adjustment quantities are generated within each time window based on traceable in-window operation evidence, and the observation mapping relationship, residual suppression method, state evolution constraint strength and observation bias correction quantity are updated at the window level accordingly. This allows the valve safety detection state to have differentiated responses to different types of changes while keeping the mechanistic constraint boundary unchanged, and avoids misinterpreting long-term bias as a continuous deterioration of the valve safety detection state.

[0026] Specifically, within the current time window, an in-window adjustment quantity is first constructed. The input of this in-window adjustment quantity comes only from the benchmarked multi-source operational dataset from step one and the valve safety status characterization quantity from step two. A valve closing constraint confidence score is then constructed. This score characterizes whether the valve leakage trend characterization quantity satisfies the valve position condition constraint within the current time window. The valve closing constraint confidence score is calculated based on the valve position signal and the valve position threshold within the valve closing interval or small opening interval. The valve position threshold is derived from the valve stroke calibration parameters or the control system valve position calibration parameters. Specifically, the percentage of time the valve position signal falls within the valve closing interval or small opening interval within the current time window is statistically analyzed. A higher percentage results in a higher valve closing constraint confidence score. The purpose of the valve closing constraint confidence score is to enhance the constraint effect of the valve leakage trend characterization quantity on the valve sealing state only when the closing condition is reliable. A frictional impact intermittency is constructed to characterize whether the valve friction anomaly is continuously increasing or intermittently increasing within the current time window. The input to the frictional impact intermittency is the vibration signal within the window and the valve friction anomaly calculated from the vibration signal. The calculation method compares the relative difference between the peak intensity and the average intensity of the valve friction anomaly within the window. Both the peak intensity and the average intensity are statistically obtained from actual sampling points within the time window. The purpose of the frictional impact intermittency is to allow for a short-term surge in valve friction state when a sharp impact occurs, while preventing this short-term surge from being continuously propagated into long-term changes in valve sealing state or valve operating state. A working condition fluctuation intensity is constructed to characterize the degree of fluctuation of the upstream pressure signal and the upstream temperature signal within the current time window. The input to the working condition fluctuation intensity is the upstream pressure signal and the upstream temperature signal. The calculation method is to statistically analyze the standard deviation and rate of change within the window. These statistics are calculated from actual sampling points within the window. The purpose of the working condition fluctuation intensity is to reduce the tendency to directly attribute changes in the upstream pressure signal to changes in valve sealing state when upstream working condition fluctuations are significant.

[0027] After obtaining the adjustment amount within the window, this step expands the observation mapping relationship from a fixed observation mapping matrix to a window-adaptive observation mapping matrix, and updates this window-adaptive observation mapping matrix based on the adjustment amount within the current time window. Here, the window-adaptive observation mapping matrix is ​​defined as: a mapping matrix in which some elements change with the adjustment amount within the window, while keeping the definitions of the observation vector and the valve safety detection state vector unchanged. The observation vector still consists of valve action deviation, valve hysteresis, valve leakage trend, and valve friction anomaly representations, and the valve safety detection state vector still consists of valve action state, valve sealing state, and valve friction state. The initial parameters of the window-adaptive observation mapping matrix are derived from the regression fitting results of historical normal operation data and debugging data, or from the initial value given by the mechanism importance and corrected by operating data. The update within the window only makes bounded adjustments based on these initial parameters. The bounded range is derived from the statistical interval of historical normal operation data or the parameter drift range allowed by the operating procedures, and the corresponding basis is saved in the parameter record to avoid unbounded parameter drift. The update mechanism within the window works as follows: when the confidence level of the valve closing constraint is low, the mapping strength of the valve leakage trend characterization quantity to the valve sealing state is reduced; when the intensity of operating condition fluctuations is high, the equivalent constraint strength of the valve inlet pressure signal fluctuations on the valve sealing state is reduced; when the intermittency of friction impacts is high, the mapping sensitivity of the valve friction anomaly characterization quantity to the valve friction state is maintained, while a stronger downward tendency constraint is applied to the continuous increase of the valve friction state, so as to avoid misinterpreting intermittent impacts as continuous deterioration.

[0028] After updating the observation mapping relationship within the window, this step further introduces an observation bias correction and performs a self-correcting update on the observation bias correction over the window sequence to handle situations where the observed representation exhibits a stable bias rather than a continuous deterioration of the actual state. Here, the observation bias correction is defined as a bias vector with the same dimension as the observation vector, used to describe the stable offset components that may appear in the valve safety state representation under normal operation or slow drift conditions. The observation bias correction does not come from external input but is obtained by statistical recursion of the observation residuals between the current window and historical windows. Specifically, within the current time window, the observed predicted value based on the state estimate of the previous window is first calculated, forming the observation residual vector. The observed predicted value is calculated from the valve safety detection state vector estimate of the previous window and the adaptive observation mapping matrix of the current window. Each component of the observation residual vector corresponds to the prediction error of the valve action deviation, valve hysteresis, valve leakage trend, or valve friction anomaly representation. Then, a bounded recursion is performed on the observation residual vector to obtain the observation bias correction for the current window. The recursive method satisfies two requirements: First, the observation bias correction only absorbs the low-frequency stable components of the residuals and not the peak components. This requirement is achieved by performing in-window smoothing statistics on the observation residual vector. Second, each component of the observation bias correction is constrained by an upper bound, which is derived from the statistical interval of the corresponding residual component in historical normal operation data or the measurement deviation range allowed by the operating procedures. The purpose of the observation bias correction is to preferentially interpret the long-term systematic deviation of a certain observation component as an observation bias rather than a continuous deterioration of the valve safety detection state, thereby avoiding the accumulation of bias into state drift during state reconstruction. To make this recursion have a clear and reproducible form, the observation bias correction can be expressed by the following bounded recursive method: ; in, This represents the observation bias correction amount for the previous time window, and η represents the recursive coefficient. The value of the recursive coefficient is based on the time-scale statistical results of bias changes in historical normal operation data. This represents the smoothed statistical result of the observed residual vector within the current time window, which is calculated from the actual sampling points within the current time window. This indicates projecting a vector onto the set of allowed biases. The projection operator allows the bias set to be composed of the upper and lower bounds of the bias corresponding to each observation component. The upper and lower bounds are derived from the historical normal operation statistical interval or the range allowed by the operation procedure. Through this bounded recursion, the observation bias correction changes slowly over time and will not exceed the bounds. After introducing an observation bias correction, this step solves for the valve safety detection state vector within the current time window, ensuring it simultaneously satisfies observation consistency, mechanism continuity, and mechanism boundary constraints. To reduce the impact of occasional anomalies in single observational representations on the solution, this step introduces a saturated suppression loss function to the observation residuals. This loss function maintains a quadratic penalty for small residuals and transforms into a linear penalty for large residuals, defined as follows: ; Where r represents the residual of a single observation component, This represents the residual inflection threshold, which is derived from the statistical upper bound or high quantile statistical results of the residual distribution of the corresponding characteristic quantity in historical normal operation data, and its data basis is saved in the parameter record. Based on the above loss function, this step solves the following constrained minimization problem within the current time window: ; in, This represents the valve safety detection state vector within the current time window, with its components corresponding to the valve action state, valve sealing state, and valve friction state, respectively. The j-th component of the observation vector is represented by the valve action deviation characteristic, the valve hysteresis characteristic, the valve leakage trend characteristic, and the valve friction anomaly characteristic. This represents the j-th component of the current window observation bias correction, which is obtained from the observation residual statistics through the aforementioned bounded recursion. This represents the window-adaptive observation mapping matrix of the current window; This represents the estimated value of the valve safety detection state vector in the previous time window; This represents the continuity constraint coefficient for the current window, and its purpose is to limit abrupt changes in the valve safety detection state between adjacent windows that are inconsistent with the mechanism. This represents the sparse variation constraint coefficient of the current window. Its purpose is to allow the valve safety detection status to change briefly with a small number of windows when intermittent impacts occur, without being excessively smoothed out by continuous constraints. This represents the result of the difference operator applied to the state vector, used to measure the magnitude of the state change over the window sequence. This is used to concentrate changes within a few windows to match intermittent enhancement phenomena. The continuity constraint coefficient... With sparsity constraint coefficient Instead of using fixed constants, updates are made within each window based on the aforementioned window adjustment values: when the intermittency of frictional impact is high, the sparse variation constraint coefficient is increased and the continuity constraint coefficient is adjusted accordingly, so that the valve friction state is allowed to increase in a short time but not mistakenly inferred as a long-term continuous increase; when the confidence level of valve closing constraint is high and the intensity of operating condition fluctuation is low, the fit strength of valve sealing state to the valve leakage trend characterization quantity is increased, so that the valve sealing state can track the slow evolution trend; when the intensity of operating condition fluctuation is high, the misleading sensitivity of valve sealing state to fluctuations in the upstream pressure signal is reduced.

[0029] For the aforementioned minimization problem, this step obtains the estimated valve safety detection state vector for the current time window under the constraints of the state value range and the state change rate. The state value range constraint is achieved by setting upper and lower bounds for the valve action state, valve sealing state, and valve friction state, respectively. These upper and lower bounds are derived from the valve's design allowable range, operating procedures, or normal sample statistical intervals. The state change rate constraint is achieved by setting an upper limit for the state difference components between adjacent time windows. This upper limit is derived from the duration range of the valve action process, the thermal inertia characteristics of the valve body metal temperature change, and historical normal operation statistical results. After solving, the estimated valve safety detection state vector for the current time window is used as the previous state vector estimate for the next time window for recursion. Simultaneously, the observation bias correction for the current time window is used as the previous observation bias correction for the next time window for recursion, thus forming a valve safety detection state reconstruction sequence.

[0030] Step 4: Consistency verification of state reconstruction results and output of security criteria; First, the reconstructed sequence of valve safety detection states is checked for boundary consistency. The boundary consistency check targets the estimated valve safety detection state vector within each time window. This vector consists of the valve action state, valve sealing state, and valve friction state. The boundary consistency check is based on the state value range constraints and state change rate constraints identified in step three. The upper and lower bounds of the state value range constraints are derived from the valve's design-allowed operating conditions, operating procedures, or confirmed normal sample statistical intervals. The upper limit of the state change rate constraints is derived from the duration of the valve action process, the thermal inertia characteristics of the valve body metal temperature change, and historical normal operation statistical results. This step checks whether the valve action state, valve sealing state, and valve friction state within each time window fall within their corresponding allowable ranges, and checks whether the state difference between adjacent windows exceeds the corresponding change limit. If a state component or difference exceeds the limit, the estimated valve safety detection state value for that time window is marked as an unreliable estimate. After completing the boundary consistency check, this step performs an observation consistency check on the valve safety detection state reconstruction sequence. The core basis for the observation consistency check is the observation vector and observation mapping matrix from step three. The observation vector consists of valve action deviation characterization, valve hysteresis characterization, valve leakage trend characterization, and valve friction anomaly characterization. The observation mapping matrix comes from the regression fitting results of historical normal operation data and debugging data, or from the initial value given by the mechanism importance and corrected by the operating data. In each time window, this step calculates the corresponding observation prediction value based on the valve safety detection state estimate of the current window, and compares it with the actual observation characterization of the window to form the observation residual. Here, the observation prediction value refers to the characterization prediction result obtained by substituting the valve safety detection state estimate into the observation mapping relationship, and the observation residual refers to the difference or norm of the difference between the actual observation characterization and the predicted characterization. The purpose of the observation residual is to quantify whether the reconstructed state can explain the actual observation, thereby determining whether the reconstruction result of the window is reliable. To ensure the judgment rules are well-founded, this step uses a threshold with a clearly defined source to judge the observation residuals. This threshold is derived from the statistical distribution of observation residuals in historical normal operation data; for example, the high quantile of normal sample residuals can be used as a threshold example, and the specific value is determined statistically during implementation. When the observation residuals of a certain window exceed this threshold, the valve safety detection status estimate for that window is marked as an unreliable estimate and processed according to the same elimination or replacement rules as the boundary consistency check. Through the above observation consistency check, mismatches where the state reconstruction meets the boundary conditions but cannot explain the observations can be identified without relying on subjective experience.

[0031] After obtaining the set of reliable windows through consistency verification, this step outputs safety judgment conclusions and early warning information based on safety criteria. To avoid unexplained new terms, this step defines the safety judgment conclusion as the classification result of the valve's current safety risk level, and the early warning information as the alarm content and corresponding time window identifier output when the safety risk level reaches the preset trigger condition. The sources of safety criteria are limited to three categories: limit values ​​given by the valve's design allowable range, control limits given by the operating procedures, and statistical thresholds given by the confirmed normal operation statistical interval; this step does not allow the use of thresholds whose sources cannot be traced. The execution method of safety judgment is as follows: corresponding criterion thresholds and criterion rules are set for the valve's operating state, valve sealing state, and valve friction state, and the judgment results of the three states are combined into a single safety risk level according to a clear fusion rule. The threshold criteria for valve operating status are derived from the acceptance standards for valve operating performance or the operating procedures' provisions regarding the allowable range of operating lag and deviation. Their purpose is to determine whether the valve poses a risk affecting regulation and protection actions. The threshold criteria for valve sealing status are derived from sealing performance requirements or the statistical range of confirmed normal closing pressure holding capacity. Their purpose is to determine whether the valve poses a risk of increased leakage tendency. The threshold criteria for valve friction status are derived from the statistical range of vibration and acoustic emission under normal operating conditions or the operating procedures' provisions regarding the allowable range of abnormal vibration. Their purpose is to determine whether the valve poses a risk of increased friction or jamming impact. The fusion rules originate from priority constraints or safety protection logic constraints in the operating procedures. For example, when any state reaches a severe limit exceedance, it is directly judged as a high-risk level; when multiple states simultaneously reach a moderate limit exceedance, it is judged as a higher level. To ensure stable output results and avoid fluctuations caused by single-window noise, this step introduces a time consistency judgment condition for the safety risk level. This time consistency judgment condition means that an early warning is triggered only when the same limit-crossing rule is met within multiple consecutive time windows, or the early warning is lifted only when multiple consecutive windows are below the release threshold as the risk level decreases. The number of consecutive windows is based on statistical results of the sampling frequency and the shortest reachable duration of valve state changes, and is determined after verification using historical operating data to avoid arbitrary setting.

[0032] When outputting early warning information, this step also outputs traceable evidence related to the early warning to meet the requirement of being well-founded. Specifically, the early warning information includes at least the trigger time window identifier, the name of the triggered valve safety detection status component, the corresponding threshold rule source type, and the estimated value of the valve safety detection status for that window. The time window identifier comes from the time window segmentation rule in step two. The valve safety detection status component name is fixed as valve action status, valve sealing status, or valve friction status. The threshold rule source type is fixed as one of the following three: design allowable range, operating procedure, or normal operation statistical interval. If a certain time window is marked as an unreliable estimate by the consistency check, this step does not output the safety risk level and early warning information for that window. Instead, it outputs the unreliable identifier of that window and its unreliable cause category. The cause category is limited to two types: boundary consistency exceeding the limit or observation consistency residual exceeding the limit.

[0033] The above formulas are all dimensionless calculations. The formulas are derived from software simulations based on a large amount of collected data to obtain the most recent real-world results. The preset parameters in the formulas are set by those skilled in the art according to the actual situation.

[0034] The above embodiments can be implemented, in whole or in part, by software, hardware, firmware, or any other combination thereof. When implemented using software, the above embodiments can be implemented, in whole or in part, in the form of a computer program product.

[0035] Those skilled in the art will recognize that the modules and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and inventive constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.

[0036] In addition, the functional modules in the various embodiments of this application can be integrated into one processing module, or each module can exist physically separately, or two or more modules can be integrated into one module.

[0037] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

[0038] In conclusion, the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for reconstructing the safety detection status of valves in high-temperature steam turbines, characterized in that, include: Acquire multi-source running data and save channel identifiers and sampling timestamps. Construct a common timeline based on the timestamps, align the original sampling sequences of each channel, and generate a benchmark multi-source running dataset. Based on sampling frequency records or timestamp statistics and valve position opening and closing process records or operating procedures, time windows are determined. Within each window, safety detection characteristics are calculated and scaled and linearly weighted to generate valve action deviation characterization, valve hysteresis characterization, valve friction anomaly characterization, and valve leakage trend characterization. Within each time window, an in-window adjustment quantity is constructed using a benchmarked multi-source operational dataset and valve safety status characterization quantities. Based on the in-window adjustment quantity, the observation mapping relationship is extended to a window adaptive mapping and updated at the window level. The observation bias correction is generated by bounded recursion from the smoothed statistics of the observation residuals. By employing a saturation suppression loss function, the continuity and sparsity constraint coefficients are updated with the adjustment amount within the window within the bounded optimization framework, and the valve action state, valve sealing state, and valve friction state are recursively obtained. The valve safety detection status reconstruction sequence is subjected to boundary consistency verification and observation consistency verification according to time window. Based on the set of reliable windows, the safety risk level and early warning information are fused according to safety criteria, and the time window identifier and threshold rule source type are output.

2. The method for reconstructing the safety detection state of a high-temperature steam turbine valve according to claim 1, characterized in that: By acquiring multi-source operating data such as valve position signal, actuator drive pressure signal, valve inlet pressure signal, valve inlet temperature signal, valve body metal temperature signal and vibration signal, and saving the channel identifier and sampling timestamp of a unified clock source for each sampling point; A common timeline is constructed based on timestamps, and the original sampling sequences of each channel are resampled to the common timeline using interpolation.

3. The method for reconstructing the safety detection state of a high-temperature steam turbine valve according to claim 2, characterized in that: For the aligned sequence, a sliding window robust statistical method is used to identify outliers and replace them by the middle value of the window or by interpolation with adjacent valid points. At the same time, missing segments are identified based on the timestamp interval and interpolation is used to fill in or mark the intervals to be removed based on the duration of the missing data. Within the valid data interval, the mean and standard deviation or statistical upper and lower limits are calculated for each channel and scaled according to standardization or interval normalization. The processed synchronization sequences of each channel are aggregated according to a common time axis to generate a benchmark multi-source running dataset.

4. The method for reconstructing the safety detection state of a high-temperature steam turbine valve according to claim 3, characterized in that: The window length and sliding step size are determined based on sampling frequency records or timestamp statistics, valve position opening and closing process records or operating procedures. The benchmarked multi-source operating dataset is divided into continuous time windows according to a common time axis. Valve safety detection characteristics are calculated in each window according to a uniform caliber. Within each window, the action consistency characteristics are calculated based on the synchronous relationship between valve position and valve stem displacement or driving pressure. The lag characteristics are calculated based on the time shift of the cross-correlation peak between driving pressure and valve position. The friction anomaly characteristics are obtained by calculating energy or root mean square based on vibration signals and optional acoustic emission signals. The leakage trend characteristics are obtained by calculating the valve inlet pressure decay rate or dispersion after selecting a closed or small opening time segment based on the valve position threshold. Then, the features within the window are scaled according to the baseline parameters obtained from historical normal operation or debugging data, and linearly weighted according to traceable and determined weights to generate valve action deviation, valve hysteresis, valve friction anomaly, and valve leakage trend. When channel data is missing, the relevant features are removed and the remaining weights are normalized.

5. The method for reconstructing the safety detection state of a high-temperature steam turbine valve according to claim 4, characterized in that: By introducing in-window adjustment quantities constructed solely from benchmarked multi-source operating data and valve safety status characterization quantities within each time window, the confidence level of valve closure constraint is formed by the time proportion of the valve position signal within the closed or small opening interval; the friction impact intermittency is formed by the difference between the peak and average intensity of the friction anomaly characterization quantity; and the operating condition fluctuation intensity is formed by the statistical analysis of the in-window standard deviation and rate of change of the valve inlet pressure signal and valve inlet temperature signal. Under the traceability constraints of mechanism boundary parameters, valve position threshold, and stroke calibration parameters, the observation mapping relationship, observation bias correction quantity, residual suppression form, and state evolution constraint intensity are updated in a window-level linkage manner.

6. The method for reconstructing the safety detection state of a high-temperature steam turbine valve according to claim 5, characterized in that: The observation mapping relationship is extended from a fixed mapping to a window adaptive mapping that varies with the adjustment amount, and the mapping intensity is adjusted within the historical normal statistical range or the drift range allowed by the procedure. The observation bias correction amount is obtained by smoothing the observation residuals of the current window and the historical window through bounded recursion to absorb low-frequency stable offset components. The residual term adopts a saturated suppression loss function that imposes a quadratic penalty on small residuals and a linear penalty on large residuals. The state solution is carried out within the bounded optimization framework of the observation consistency term, the continuity term of adjacent windows, and the sparse change term, and the continuity and sparse change constraint coefficients are updated with the adjustment amount. By adopting differentiated mapping adjustment, bias recursion, and constraint coefficient update paths for different modes such as synchronous increase, single anomaly, slow trend, and intermittent enhancement in different windows, a recursive reconstruction sequence of valve action state, valve sealing state, and valve friction state is generated.

7. The method for reconstructing the safety detection state of a high-temperature steam turbine valve according to claim 6, characterized in that: By performing consistency checks on the valve safety detection state reconstruction sequence one by one according to time windows and outputting safety criterion conclusions: the upper and lower bounds of the state value range and the upper limit of the state change rate of adjacent windows, the valve action state, valve sealing state and valve friction state of each window are checked for out-of-bounds and differential out-of-limit checks, and out-of-bounds windows are marked as unreliable; then, based on the observation vector and observation mapping relationship, the observed prediction value is calculated from the state estimate value within the window and formed with the actual observed characterization value to form the observation residual. The threshold derived from the statistical distribution of the residuals from historical normal operation is used to judge the residual out-of-limit and mark the out-of-limit windows as unreliable; After obtaining the set of reliable windows, the judgment thresholds and judgment rules from the design allowable range, operating procedures or normal operation statistical intervals are set for the valve action state, valve sealing state and valve friction state respectively. The judgment results of the three types of states are fused according to the priority of the operating procedures or the safety protection logic to obtain the safety risk level. At the same time, the time consistency judgment of multiple consecutive windows satisfying the triggering conditions or release conditions is introduced. When outputting early warning information, the system provides the trigger time window identifier, the name of the triggered state component, the source type of the threshold rule, and the estimated state value of the window. For untrusted windows, the system outputs an untrusted identifier and the reason category.

Images