Valve life prediction method and system based on error calibration

Abstract

The invention discloses a valve life prediction method and system based on error calibration, and relates to the technical field of valve life prediction. Comprising the following steps: acquiring a valve type and a working scene, constructing a multi-dimensional prediction index, setting a prediction index error range, sorting error sources and classifying, and constructing a valve life standard set; the method comprises the following steps: collecting full-period data of a valve, sorting the collected data, and constructing a full-period associated data set; according to the technical key points, starting from a working scene of the valve, a feature extraction technology and an error feature quantification technology are combined, and a standard index of the valve is formulated. The condition that original data is subjected to feature distortion due to noise and drifting caused by the sensor and environmental reasons is avoided, the judgment standards of abnormal types and grades are unified, the deviation caused by the experience difference of personnel is eliminated, the judgment of the valve state is more accurate, and the method has a good use prospect.

Description

Technical Field

[0001] This invention relates to the field of valve life prediction technology, specifically to a valve life prediction method and system based on error calibration. Background Technology

[0002] Pipelines are used in all aspects of life, such as water supply and heating pipes in the home, chemical raw material transportation, crude oil transportation, steam and coolant transportation in industry, and irrigation in agriculture. Pipelines are responsible for the transportation of liquids, gases and powders, and play an irreplaceable role in daily life, clothing, food, housing and transportation as well as industrial production.

[0003] In pipelines, valves are the core control components. In civil applications, valves are used to control the flow of pipelines, while in industrial applications, valves with multiple functions are used in combination to control pipelines. In pipelines, valves determine the efficiency and safety of pipeline transmission. Once a valve malfunctions, it may cause the entire pipeline to be interrupted or the medium to leak, which may lead to a great danger.

[0004] Therefore, it is necessary to understand the condition of the valves and, based on their condition and service life, to develop appropriate maintenance and replacement plans.

[0005] However, in actual use, smart instruments are usually used to collect valve working data. The subsequent analysis of the collected data still mainly relies on human experience, but this method has certain drawbacks. The main problem is the large differences in personnel experience, the lack of a unified standard for judging valve abnormality types and valve status, and the fact that the same valve may be evaluated by different personnel with different results. Moreover, the existing methods all rely on the current state of the valve for prediction without considering the impact of errors and operating conditions on the valve. This results in a large gap between the predicted valve life and the actual valve life, making the subsequent maintenance and replacement plans less economical and reliable. In summary, existing valve life prediction methods do not meet the requirements. Therefore, we propose a valve life prediction method and system based on error calibration. Summary of the Invention

[0006] To achieve the above objectives, the present invention provides the following technical solution: A valve life prediction method based on error calibration includes the following steps: Obtain valve type and working scenario, construct multi-dimensional prediction indicators, set the error range of prediction indicators, sort out and classify error sources, and construct valve life standard set; Collect valve lifecycle data and organize the collected data to construct a full lifecycle related dataset; Preprocessing and feature enhancement fusion of data in the full-cycle associated dataset are performed to form a model structured dataset; Construct and train the core prediction model, collect real-time training data of the valve, and input the real-time training data of the valve into the core prediction model to obtain preliminary prediction results; Based on the analysis of the model's structured dataset, the correlation is identified, the error propagation path and error type are clarified, an error calibration model is constructed, and the preliminary prediction results are calibrated to obtain preliminary calibration results. Acquire real-time operating data, analyze the data, correct preliminary calibration results, and formulate valve maintenance plans based on valve importance classification.

[0007] Preferably, the multidimensional prediction indicators include valve anomaly indicators, current health status, and remaining lifespan. The steps for constructing the valve lifespan standard set are as follows: Determine multidimensional predictive indicators, set valve anomaly type levels, current health status assessment values, and lifespan units to form a predictive indicator framework; Establish error ranges for valve anomaly indicators, current health status, and remaining lifespan; The sources of error throughout the entire process are identified and categorized into inherent model error, data acquisition error, and operating condition fluctuation error, and quantifiable error description indicators are established. The prediction indicators, error range, error sources, and error description indicators are summarized to construct a standard set of valve lifespan.

[0008] Preferably, the full-cycle data of the valve includes basic data and error data. The basic data includes design and manufacturing data, operation monitoring data, fault maintenance data, and other data. The design and manufacturing data includes valve model, valve material, design rated pressure, design rated flow, and design life. The operation monitoring data includes medium pressure, temperature, flow rate, valve opening and closing times, and valve stem vibration amplitude collected by sensors. The fault maintenance data includes historical fault types, fault occurrence time, fault handling measures, maintenance records, and equipment component information. Other data includes data supplemented by simulation when data is missing. The error data includes core deviation data, error correlation characteristics, and error source labels. The core deviation data includes the difference between each preliminary prediction result and the actual value. The error correlation characteristics are sensor accuracy level, data sampling frequency, operating condition fluctuation amplitude, and data missing rate. The error source labels include timestamps and deviation markings that have been tested and verified.

[0009] Preferably, the data processing includes: extracting full-cycle data of the valve, matching error data with basic data using timestamps, establishing a correlation relationship to obtain associated data, verifying the completeness of the associated data, supplementing missing error data or basic data, organizing the associated data and converting it into a structured table, attaching a unique timestamp to the associated data and numbering it to obtain a full-cycle associated dataset.

[0010] Preferably, the preprocessing and feature enhancement fusion processing of data in the full-cycle associated dataset includes: Outliers in the basic data were removed using the Raida criterion, missing data were filled using the linear imputation method, and sensor noise was smoothed using the moving average method. Extract fundamental features related to multidimensional prediction indicators from the base data after noise smoothing; The extracted basic features were processed using the standard score method; Analyze error data, calculate error quantification indicators, calculate the Pearson correlation coefficient between error quantification indicators and error descriptive indicators, and screen out error features. The basic features and error features that have undergone standardization are horizontally concatenated, and anomaly labels, status labels, lifetime labels and error labels are added to obtain fused feature data; The feature data is divided and integrated according to proportions to form a training set, a validation set, and a test set, which constitute the model's structured dataset.

[0011] Preferably, before constructing and training the core prediction model, causal inference and statistical analysis are used to clarify the impact of the increase in the anomaly type level on the state assessment value and the difference in the life loss rate under different state assessment values. The basic features related to the multidimensional prediction indicators include the time series features related to the valve anomaly identification, the state features related to the current health status, and the life features related to the remaining life. The core prediction model consists of an input layer, an anomaly detection layer, a state evaluation layer, a lifetime prediction layer, and an output layer. The input layer receives temporal features, state features, and lifetime features, which are then standardized and concatenated before being transmitted to the anomaly detection layer, state evaluation layer, and lifetime prediction layer, respectively. The anomaly detection layer takes the temporal features as input, processes them through convolutional kernels, pooling for dimensionality reduction, and fully connected neurons, and then uses the Softmax activation function for classification, outputting the anomaly type level and the corresponding confidence score. The state evaluation layer takes the confidence score corresponding to the anomaly type level and the state features as input, and uses a confidence threshold to remove low-confidence features. The system calculates the comprehensive anomaly impact value by combining the anomaly type level with the anomaly interaction mechanism. After concatenation, the values ​​are mapped to a high-dimensional space using a Gaussian kernel in a support vector machine. The system then classifies and transforms the output state evaluation value and evaluation confidence. The lifespan prediction layer takes the state evaluation value, evaluation confidence, and lifespan features as input. After concatenation, the system uses a long short-term memory network to extract features. After fully connected processing by neurons, the system outputs the preliminary remaining lifespan and lifespan fluctuation range. The output layer summarizes the anomaly type level, the confidence level corresponding to the anomaly type level, the state evaluation value, the evaluation confidence, the preliminary remaining lifespan, and the lifespan fluctuation range. After processing, the preliminary prediction results are obtained.

[0012] Preferably, the error calibration model includes an input layer, an error prediction layer, an error calibration layer, and an output layer. The input layer receives the preliminary prediction results and error features, concatenates them, standardizes them, and outputs structured error feature data. The error prediction layer takes the structured error feature data as input, uses the Gini coefficient to screen key features, and predicts and outputs anomaly type level error, state classification error probability, and lifetime deviation rate through gradient boosting decision tree. The error calibration layer uses anomaly type level error, state classification error probability, and lifetime deviation rate to calibrate the preliminary prediction results. The output layer takes the anomaly type level, state assessment value, remaining lifetime, anomaly type level error range, state assessment value error range, and lifetime error range after error calibration as input to obtain the preliminary calibration results.

[0013] Preferably, the steps for analyzing the operating data and correcting the preliminary calibration results are as follows: Real-time acquisition of on-site working condition data and sensor noise values, and comparison with the working condition data threshold range and noise threshold to determine whether there are working condition data fluctuations exceeding the working condition data threshold range or sensor noise values ​​exceeding the noise threshold range. If the fluctuation of the operating condition data exceeds the threshold range of the operating condition data, the remaining life and life error range shall be adjusted in accordance with the error propagation law. If a sensor noise value exceeds the noise threshold range, the anomaly type level and anomaly type level error range are corrected according to the error association rules, and the state evaluation value is recalculated based on the corrected anomaly type level and anomaly type level error range to obtain the corrected state evaluation value.

[0014] Preferably, the steps for developing a valve maintenance plan, based on the valve's importance classification, are as follows: Extract and correct the preliminary calibration results, find the valve importance level, determine whether it is a critical valve, and mark the maintenance priority; The preliminary calibration results were analyzed to obtain the maintenance deviation adjustment value; The preset maintenance date range is adjusted based on the maintenance deviation adjustment value to obtain the corrected maintenance date range; Summarize and correct the maintenance date range, and retrieve the monitoring strategies corresponding to the corrected status assessment values ​​to form a valve maintenance plan.

[0015] A valve life prediction system based on error calibration includes a basic analysis module, a data acquisition module, a data adjustment module, a preliminary prediction module, an error calibration module, and a real-time correction module. Basic Analysis Module: Obtain valve type and working scenario, construct multi-dimensional prediction indicators, set the error range of prediction indicators, sort out and classify error sources, and construct valve life standard set; Data acquisition module: Collects valve data throughout its entire lifecycle, processes the collected data, and constructs a full-cycle related dataset; Data adjustment module: preprocesses and enhances features in the full-cycle associated dataset to form a structured dataset for the model; Preliminary prediction module: Constructs and trains the core prediction model, collects real-time training data of the valve, and inputs the real-time training data of the valve into the core prediction model to obtain preliminary prediction results; Error calibration module: Based on the analysis of the correlation of the model's structured dataset, the error propagation path and error type are identified, an error calibration model is constructed, the preliminary prediction results are calibrated, and the preliminary calibration results are obtained; Real-time correction module: Acquires real-time operating data, analyzes the operating data, corrects the preliminary calibration results, and formulates valve maintenance plans based on the valve importance classification.

[0016] This invention provides a valve life prediction method and system based on error calibration, which has the following beneficial effects: This invention starts from the working scenario of the valve and combines feature extraction technology with error feature quantification technology to formulate standard indicators for the valve. It can effectively solve the problems in the existing technology where the subjective judgment of anomalies based on human experience is relatively strong, and the original data is distorted due to noise and drift caused by sensor itself and environmental factors. Moreover, it unifies the judgment criteria for anomaly types and levels, eliminates the deviation caused by differences in human experience, and makes the judgment of valve status more accurate and effective.

[0017] The core prediction model of this invention combines anomaly confidence screening technology with anomaly interaction mechanism analysis technology in the state assessment. This effectively solves the problem that manual judgment uses an isolated and singular approach without considering the promoting effect of anomalies, which leads to valve state assessment being detached from reality and causing inaccurate valve life prediction. By combining anomaly screening and anomaly interaction analysis of state assessment values, accurate valve state assessment is achieved, laying the foundation for subsequent valve life prediction.

[0018] This invention classifies error sources and dynamically calibrates data based on working scenarios, effectively solving the problem that simple manual correction in existing technologies cannot effectively address errors. By correcting and real-time adjusting the error of the core prediction model's predicted data, it can achieve accurate analysis of valve status and lifespan, making the solution applicable under different working conditions, with wide applicability and good performance.

[0019] This invention employs full-cycle dataset construction technology and model iterative optimization technology to achieve automatic prediction of valve status and lifespan. It effectively solves the problem that existing prediction methods rely on experience and cannot adapt to different working conditions. It realizes valve lifespan prediction for different pipeline types and working conditions. Furthermore, through comprehensive analysis, it enables adjustments to valve maintenance plans, making the solution more comprehensive, effectively reducing the workload of subsequent personnel, and demonstrating good performance and promising application prospects. Attached Figure Description

[0020] Figure 1 This is a flowchart of a valve life prediction method based on error calibration according to the present invention; Figure 2 This is a flowchart illustrating the correction of preliminary calibration results in a valve life prediction method based on error calibration according to the present invention. Figure 3 This is a structural block diagram of a valve life prediction system based on error calibration according to the present invention. Detailed Implementation

[0021] 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.

[0022] This application addresses the monitoring of valves during pipeline use, enabling status analysis and lifespan prediction of different valves within the pipeline. This allows relevant personnel to understand the valve conditions and directly implement the maintenance of the entire pipeline based on the summarized valve maintenance plan. It offers comprehensive functionality and superior performance.

[0023] Example 1: Please see Figure 1 and Figure 2 This embodiment provides a valve life prediction method based on error calibration, including the following steps: S1. Obtain valve type and working scenario, construct multi-dimensional prediction indicators, set the error range of prediction indicators, sort out and classify error sources, and construct valve life standard set; Different valve types have different data corresponding to their operation, and different working scenarios naturally result in different states and lifespans. For example, valves in humid environments are prone to damage. Therefore, it is necessary to obtain the valve type and working scenario, and construct multi-dimensional predictive indicators based on the valve type and working scenario.

[0024] The solution proposed in this application is applicable to all valves and is a general type. Therefore, the corresponding multidimensional prediction indicators include valve anomaly identification, current health status, and remaining life. Valve anomaly identification refers to the valve's anomaly type and anomaly level, i.e., the subsequent anomaly type level. The current health status is represented by a number from 0 to 10, with a larger number indicating a worse health status, 0 being the best and 10 being the worst. The remaining life is estimated in days.

[0025] The steps to construct a valve life standard set are as follows: Determine multidimensional predictive indicators, set valve anomaly type levels, current health status assessment values, and lifespan units to form a predictive indicator framework; Establish error ranges for valve anomaly indicators, current health status, and remaining lifespan. All data have a certain fluctuation range, i.e., error range. For example, the anomaly type level corresponding to the valve anomaly indicator, the status assessment value corresponding to the current health status, and the fluctuation range of the remaining lifespan. For example, the allowable error for the remaining lifespan is ±5%. The sources of error throughout the entire process are identified and categorized into inherent model error, data acquisition error, and operating condition fluctuation error, and quantifiable error description indicators are established. For example, quantifiable error description indicators include inherent error, data error, and operating condition fluctuation error. Inherent error is caused by the algorithm during data analysis and is generally expressed as root mean square error and mean absolute error. Data error is mainly caused by data acquisition equipment and is generally expressed as deviation rate and data missing rate. Operating condition fluctuation error is mainly caused by sudden changes in the working environment and is generally expressed as coefficient of variation and error fluctuation amplitude. Setting quantifiable error description indicators is mainly used for subsequent error analysis.

[0026] The prediction indicators, error range, error sources, and error description indicators are summarized to construct a standard set of valve lifespan.

[0027] Based on the valve life standard set, a unified benchmark is provided for subsequent analysis.

[0028] The valve life standard set clarifies multi-dimensional prediction indicators, enabling standardized valve analysis. It also organizes valve-related data, providing a basis for data partitioning for S2, and offering a standard for subsequent model analysis and calibration.

[0029] S2. Collect the valve's full-cycle data, organize the collected data, and construct a full-cycle correlation dataset; The full lifecycle data of a valve includes basic data and error data. Basic data includes design and manufacturing data, operational monitoring data, fault maintenance data, and other data. Design and manufacturing data is generated during valve design and production, including valve model, valve material, design rated pressure, design rated flow, and design life. Operational monitoring data is generated during valve use, including media pressure, temperature, flow rate, valve opening and closing times, and valve stem vibration amplitude collected by sensors. Fault maintenance data is historical valve maintenance data, including historical fault types, fault occurrence times, fault handling measures, maintenance records, and equipment component information. Other data includes data simulated through other methods when data acquisition errors occur, including data supplemented through simulation when data is missing, and data obtained through bench testing. The process involves testing and supplementing missing data. Error data includes core deviation data, error correlation features, and error source labels. Error data primarily consists of historical prediction data, including data from manual predictions or predictions made using other methods. Core deviation data includes the difference between each initial prediction and the actual value; for example, if the predicted lifespan is 90 days and the actual lifespan is 96 days, the difference is 6 days. Error correlation features are data that may affect the error, including sensor accuracy level, data sampling frequency, operating condition fluctuation amplitude, and data missing rate. Error source labels are mainly for facilitating data organization and retrieval, adding labels based on accurate results, including timestamps and verified deviation annotations, such as sensor data acquisition time errors, marked as sensor drift, indicating an anomaly.

[0030] The process of organizing the collected data includes: extracting the full-cycle data of the valves, matching the error data with the basic data using timestamps, establishing a correlation relationship, obtaining the associated data, verifying the completeness of the associated data, supplementing missing error data or basic data, organizing the associated data and converting it into a structured table, attaching a unique timestamp to the associated data and numbering it, so as to ensure that each piece of basic data can be traced back to the corresponding error information, which facilitates subsequent analysis and processing, and obtaining a full-cycle associated dataset.

[0031] S3. Perform preprocessing and feature enhancement fusion on the data in the full-cycle associated dataset to form a model structured dataset; Preprocessing and feature enhancement fusion of data in the full-cycle associated dataset includes: Outliers in the basic data are removed using the Raida criterion, which can remove extreme data that deviate from the normal distribution. Then, missing data is filled using the linear imputation method, which is used for data with a missing rate of less than 3% to avoid affecting the overall trend of the data. Finally, the moving average method is used to smooth sensor noise, mainly targeting high-frequency noise from vibration and pressure time series during sensor acquisition. When processing noise, the window size is set to 5. Extract fundamental features related to multidimensional prediction indicators from the base data after noise smoothing; The extracted basic features are processed using the standard score method. The basic features include abnormal features that reflect the occurrence of valve abnormalities, state features that represent the health status of valves, and life features that are related to the lifespan of valves. The abnormal features are time-series features. Abnormal characteristics can be identified using common valve data such as vibration amplitude, pressure fluctuation frequency, and temperature deviation at the valve stem seal; status characteristics can be identified using data such as the ratio of service life to design life, cumulative number of opening and closing times, average working pressure ratio, average interval between failures, and service life after the most recent maintenance; lifespan characteristics can be identified using indicators such as cumulative number of failures, average maintenance time, number of parts replaced, and deviation from the historical average lifespan of similar valves.

[0032] The standard score method transforms all basic features into standardized data with a mean of 0 and a standard deviation of 1, eliminating the influence of units and improving the stability of subsequent model training. The standard score method is a commonly used data processing method and is common knowledge, so it will not be described in detail.

[0033] Analyze error data, calculate error quantification indicators, calculate the Pearson correlation coefficient between error quantification indicators and error descriptive indicators, and screen out error features. Error quantification indicators include root mean square error, mean absolute error, and deviation rate. Root mean square error reflects the overall fluctuation of error, mean absolute error reflects the average level of error, and deviation rate reflects the magnitude of relative error. The combination of the three can comprehensively characterize error characteristics.

[0034] Pearson correlation coefficient analysis was used to analyze the linear correlation between error quantification indicators and error description indicators. Features with an absolute correlation coefficient greater than 0.6 were selected as error features. Highly correlated error features were retained to reduce data redundancy and improve model efficiency. Error features selected by Pearson correlation coefficient included sensor noise variance, coefficient of variation of operating parameters, data missing rate, sensor accuracy level, and sampling frequency.

[0035] The basic features and error features that have undergone standardization are horizontally concatenated, and anomaly labels, status labels, lifetime labels and error labels are added to obtain fused feature data; When horizontally concatenating basic features and error features, timestamps and numbers are used as unique identifiers to form an input feature vector. Labels are then added to facilitate subsequent analysis.

[0036] The feature data is divided and integrated according to proportions to form a training set, a validation set, and a test set, which constitute the model's structured dataset.

[0037] The training set is used for learning and optimizing model parameters, the validation set is used for hyperparameter tuning and preventing overfitting, and the test set is used for final evaluation of the model's generalization ability. The ratio of the training set, validation set, and test set can be 6:2:2 or 7:1.5:1.5. The specific ratio can be flexibly adjusted according to the total amount and distribution of data to ensure that the distribution of each type of sample is balanced in each dataset.

[0038] The training set should contain more than 1,000 data points to ensure that the model can fully learn the nonlinear relationships between features and avoid overfitting due to insufficient samples.

[0039] This invention starts from the working scenario of the valve and combines feature extraction technology with error feature quantification technology to formulate standard indicators for the valve. It can effectively solve the problems in the existing technology where the subjective judgment of anomalies based on human experience is relatively strong, and the original data is distorted due to noise and drift caused by sensor itself and environmental factors. Moreover, it unifies the judgment criteria for anomaly types and levels, eliminates the deviation caused by differences in human experience, and makes the judgment of valve status more accurate and effective.

[0040] S4. Construct and train the core prediction model, collect real-time training data of the valve, and input the real-time training data of the valve into the core prediction model to obtain preliminary prediction results; Before constructing and training the core prediction model, causal inference and statistical analysis were used to clarify the impact of the increase in anomaly type level on the state assessment value and the difference in life loss rate under different state assessment values. The key influence paths were identified to guide the design of the model structure. For example, through cross-analysis, it was found that for every 1 level increase in anomaly type level, the state assessment value increases by 0.68 and the life loss rate increases by 0.3%. For example, the loss rate of a certain valve is 0.6% / day when the state assessment value is 3, and 2.1% / day when the assessment value is 8. Anomalies interact with each other, thereby amplifying the loss rate, rather than simply adding them together, making the predicted life more accurate.

[0041] The basic features associated with multidimensional predictive indicators include time-series features related to valve anomaly indicators, state features related to current health status, and lifespan features related to remaining lifespan. The core prediction model consists of an input layer, an anomaly detection layer, a state evaluation layer, a lifetime prediction layer, and an output layer. The input layer receives temporal features, state features, and lifetime features, which are then standardized and concatenated before being transmitted to the anomaly detection layer, state evaluation layer, and lifetime prediction layer, respectively. The anomaly detection layer takes the temporal features as input, processes them through convolutional kernels, pooling for dimensionality reduction, and fully connected neurons, and then uses the Softmax activation function for classification, outputting the anomaly type level and the corresponding confidence score. The state evaluation layer takes the confidence score corresponding to the anomaly type level and the state features as input, and uses a confidence threshold to remove low-confidence features. The system calculates the comprehensive anomaly impact value by combining the anomaly type level with the anomaly interaction mechanism. After concatenation, the values ​​are mapped to a high-dimensional space using a Gaussian kernel in a support vector machine. The system then classifies and transforms the output state evaluation value and evaluation confidence. The lifespan prediction layer takes the state evaluation value, evaluation confidence, and lifespan features as input. After concatenation, the system uses a long short-term memory network to extract features. After fully connected processing by neurons, the system outputs the preliminary remaining lifespan and lifespan fluctuation range. The output layer summarizes the anomaly type level, the confidence level corresponding to the anomaly type level, the state evaluation value, the evaluation confidence, the preliminary remaining lifespan, and the lifespan fluctuation range. After processing, the preliminary prediction results are obtained.

[0042] In the input layer, the temporal features retain the time dimension separately, while the static features are state features and lifetime features. The static features are concatenated in batches. For the static features, min-max standardization is performed. The processed static features and the standardized temporal features maintain the same dimensions. The state features and lifetime features are then concatenated.

[0043] The anomaly detection layer mainly extracts local data from temporal features through convolutional neural networks and performs combined classification after analysis.

[0044] The anomaly detection layer contains convolutional layers, pooling layers, classification layers, and fully connected layers. The convolutional layers extract abnormal fluctuation patterns by sliding multiple convolutional kernels, and their activation function is ReLU, which extracts local abnormal features from temporal features. The pooling layers use max pooling to retain key features and reduce data computation. The fully connected layers flatten the pooled features into a one-dimensional vector and also use the ReLU activation function. The classification layer uses the Softmax function to normalize the output of the fully connected layers, generating probability distributions for various anomaly types, outputting the anomaly type level and corresponding confidence level, providing input for the state evaluation layer. The anomaly type level is a combination of anomaly type and anomaly level.

[0045] The state evaluation layer is the core layer of the model. It receives static features, anomaly type levels, and corresponding confidence levels after being concatenated by the input layer.

[0046] The state assessment layer processing flow is as follows: The confidence level corresponding to the anomaly type is compared with a preset threshold. If it is lower than the threshold, the anomaly type is removed to avoid misjudgment. After retaining the high-confidence anomaly type levels, the comprehensive anomaly impact value is calculated based on the mechanism of anomaly interaction. The specific calculation formula is as follows: ; In the formula, The total impact value of the anomaly is calculated, where K is the number of high-confidence anomaly types retained after screening. Let i be the confidence level corresponding to the i-th anomaly type. The anomaly level of the i-th anomaly type. Let be the facilitation coefficient of the j-th anomaly type on the i-th anomaly type. To avoid over-correction of coefficients and excessive calculation of the synergistic effects between anomaly types, The value is typically between 0.4 and 0.6, and is obtained through historical data analysis. This is mainly because there are two promotion scenarios during the calculation: the promotion of the j-th anomaly type to the i-th anomaly type and the promotion of the i-th anomaly type to the j-th anomaly type.

[0047] The calculated comprehensive anomaly impact value is concatenated with the state features to form a feature vector. After min-max standardization, it is mapped to a high-dimensional space through a Gaussian kernel in a support vector machine. The regression output yields the state evaluation value, and the evaluation confidence is obtained by normalizing the output probability of the support vector machine.

[0048] The lifetime prediction layer horizontally concatenates the state assessment value, assessment confidence level, and lifetime characteristics, and then reshapes them into time-series features according to the time step. After processing and analysis by the Long Short-Term Memory Network, the preliminary remaining lifetime and lifetime fluctuation range are obtained.

[0049] The core prediction model of this invention combines anomaly confidence screening technology with anomaly interaction mechanism analysis technology in the state assessment. This effectively solves the problem that manual judgment uses an isolated and singular approach without considering the promoting effect of anomalies, which leads to valve state assessment being detached from reality and causing inaccurate valve life prediction. By combining anomaly screening and anomaly interaction analysis of state assessment values, accurate valve state assessment is achieved, laying the foundation for subsequent valve life prediction.

[0050] S5. Based on the model's structured dataset, analyze the correlation, clarify the error propagation path and error type, construct an error calibration model, calibrate the preliminary prediction results, and obtain preliminary calibration results; The analysis of correlations based on the model-structured dataset involves quantifying the impact of error characteristics on error through regression analysis, and clarifying the error propagation rules and error compensation conditions. Specifically, regression analysis is used to quantify the impact of error characteristics on errors, construct a mapping relationship between characteristics and errors, then trace the transmission path of errors in the anomaly identification, state assessment and lifetime prediction stages, clarify the quantitative relationship of error transmission in each stage, clarify the error compensation method, and organize the mapping relationship, transmission path and compensation method to construct a rule condition table.

[0051] The error calibration model comprises an input layer, an error prediction layer, an error calibration layer, and an output layer. The input layer receives preliminary prediction results and error features, concatenates them, standardizes them, and outputs structured error feature data. The error prediction layer takes the structured error feature data as input, uses the Gini coefficient to filter key features, and predicts and outputs anomaly type level error, state classification error probability, and lifetime deviation rate through a gradient boosting decision tree. The error calibration layer uses anomaly type level error, state classification error probability, and lifetime deviation rate to calibrate the preliminary prediction results. The output layer takes the error-calibrated anomaly type level, state assessment value, remaining lifetime, anomaly type level error range, state assessment value error range, and lifetime error range as input to obtain the preliminary calibration results.

[0052] The input layer receives preliminary prediction results and error features. First, the received data is horizontally concatenated. Then, the box plot method is used to remove outliers in the concatenated features. The error features are processed by min-max standardization, and the preliminary prediction results are processed by standard score standardization to eliminate dimensional differences.

[0053] The error prediction layer uses the Gini coefficient to calculate the importance of features to the error, sorts them according to importance, and selects the top M features. Gradient boosting decision trees are used for error regression prediction. The model is integrated with 80-150 regression trees, each with a depth of 4-6 and a learning rate of 0.01. Mean squared error is used as the node splitting standard, and three types of core error quantification values ​​are output, which calculate the deviation of three types of data in the preliminary prediction results. Specifically, these are the difference between the anomaly type level and the true value, the difference between the state assessment value and the true value, and the relative error between the remaining life and the true value.

[0054] The error calibration layer corrects the anomaly type level, state assessment value, and remaining lifetime based on the deviation of the three types of data, and updates the assessment confidence level. The updated assessment confidence level is the assessment confidence level multiplied by the error prediction confidence level. The error prediction confidence level is obtained during the error calibration model analysis. The output layer statistically analyzes the results output by the error calibration layer, clarifies the data fields and formats, and summarizes them to obtain the preliminary calibration results.

[0055] The error calibration model and the core prediction model have the same training, validation and test sets and are trained synchronously. The loss function during training is the mean squared error. During training, the Adam optimizer is used for 100 iterations. Each iteration reads a batch of training set data, calculates the gradient through backpropagation, and updates the parameters of the error calibration model and the core prediction model synchronously to ensure that the parameters of the two are optimized together.

[0056] The training process achieves parameter co-learning through a joint loss function and the Adam optimizer; the validation process relies on the validation set to dynamically fine-tune hyperparameters and suppress overfitting; the validation set metrics are verified, and model parameters are adjusted until they meet requirements such as lifetime error; the testing process uses the test set to determine performance and ensure that the accuracy requirements of the application scenario are met.

[0057] This invention classifies error sources and dynamically calibrates data based on working scenarios, effectively solving the problem that simple manual correction in existing technologies cannot effectively address errors. By correcting and real-time adjusting the error of the core prediction model's predicted data, it can achieve accurate analysis of valve status and lifespan, making the solution applicable under different working conditions, with wide applicability and good performance.

[0058] S6. Acquire real-time operating data, analyze the operating data, correct the preliminary calibration results, and formulate a valve maintenance plan based on the valve importance classification.

[0059] The steps for analyzing operating data and correcting the preliminary calibration results are as follows: Real-time acquisition of on-site working condition data and sensor noise values, and comparison with the working condition data threshold range and noise threshold to determine whether there are working condition data fluctuations exceeding the working condition data threshold range or sensor noise values ​​exceeding the noise threshold range. Both the operating condition data threshold range and the noise threshold are obtained through historical data analysis. If the operating condition data fluctuation exceeds the operating condition data threshold range or the sensor noise value exceeds the noise threshold range, then there is an abnormal situation. If the fluctuation of the operating condition data exceeds the threshold range, adjust the remaining life and life error range according to the error propagation law, update the remaining life and life error range, and record the adjustment basis as the adjustment based on the operating condition. If a sensor noise value exceeds the noise threshold range, the anomaly type level and anomaly type level error range are corrected according to the error association rules, and the state evaluation value is recalculated based on the corrected anomaly type level and anomaly type level error range to obtain the corrected state evaluation value.

[0060] All corrected data is summarized, and corresponding real-time operating data is statistically analyzed to verify data consistency and generate a report for easy retrieval and analysis later.

[0061] The steps for developing a valve maintenance plan based on the valve's importance classification are as follows: Extract and correct the preliminary calibration results, find the valve importance level, determine whether it is a critical valve, and mark the maintenance priority; Critical valves are those on the main pipeline line, whose abnormalities will have a huge impact. Non-critical valves are those on the branch pipeline line. Valves that are irreplaceable on the branch line, difficult to repair on the branch line, or have high economic losses due to failure on the branch line can also be designated as critical valves. Correspondingly, valves on the main pipeline line that are easy to replace, have spare parts, and whose damage results in low economic losses can also be designated as non-critical valves. If the maintenance days are the same, critical valves should be maintained first.

[0062] The preliminary calibration results were analyzed to obtain the maintenance deviation adjustment value; The preset maintenance date range is adjusted based on the maintenance deviation adjustment value to obtain the corrected maintenance date range; Summarize and correct the maintenance date range, and retrieve the monitoring strategies corresponding to the corrected status assessment values ​​to form a valve maintenance plan.

[0063] For critical valves, the maintenance date can be set as the lower limit of the corrected life error range - the critical valve condition influence coefficient × condition assessment value. The higher the corrected condition assessment value, the more frequent the monitoring. For example, each level of corrected condition assessment value corresponds to a monitoring frequency. If the corrected condition assessment value is 10, the monitoring frequency is once every hour, while if the corrected condition assessment value is 5, the monitoring frequency is once every 4 hours. The frequency can be adjusted according to the valve requirements.

[0064] For non-critical valves, the maintenance date can be set to the upper limit of the corrected life error range - the critical valve status influence coefficient × status assessment value. The higher the corrected status assessment value, the more frequent the monitoring, but it is lower than the monitoring frequency of critical valves. For example, if the corrected status assessment value is 10, the monitoring frequency is once every 2 hours, while if the corrected status assessment value is 5, the monitoring frequency is once every 8 hours. Adjustments can be made according to the valve requirements.

[0065] For pipelines, all correction and maintenance date ranges can be summarized, the dates extracted, and then sorted by priority to facilitate maintenance by relevant personnel.

[0066] This invention employs full-cycle dataset construction technology and model iterative optimization technology to achieve automatic prediction of valve status and lifespan. It effectively solves the problem that existing prediction methods rely on experience and cannot adapt to different working conditions. It realizes valve lifespan prediction for different pipeline types and working conditions. Furthermore, through comprehensive analysis, it enables adjustments to valve maintenance plans, making the solution more comprehensive, effectively reducing the workload of subsequent personnel, and demonstrating good performance and promising application prospects.

[0067] Example 2: Based on Example 1, such as Figure 3 As shown, this embodiment also provides a valve life prediction system based on error calibration, including a basic analysis module, a data acquisition module, a data adjustment module, a preliminary prediction module, an error calibration module, and a real-time correction module: Basic Analysis Module: Obtain valve type and working scenario, construct multi-dimensional prediction indicators, set the error range of prediction indicators, sort out and classify error sources, and construct valve life standard set; Data acquisition module: Collects valve data throughout its entire lifecycle, processes the collected data, and constructs a full-cycle related dataset; Data adjustment module: preprocesses and enhances features in the full-cycle associated dataset to form a structured dataset for the model; Preliminary prediction module: Constructs and trains the core prediction model, collects real-time training data of the valve, and inputs the real-time training data of the valve into the core prediction model to obtain preliminary prediction results; Error calibration module: Based on the analysis of the correlation of the model's structured dataset, the error propagation path and error type are identified, an error calibration model is constructed, the preliminary prediction results are calibrated, and the preliminary calibration results are obtained; Real-time correction module: Acquires real-time operating data, analyzes the operating data, corrects the preliminary calibration results, and formulates valve maintenance plans based on the valve importance classification.

[0068] The above embodiments can be implemented, in whole or in part, by software, hardware, firmware, or any other combination thereof. When implemented in software, the above embodiments can be implemented, in whole or in part, as a computer program product. Those skilled in the art will recognize that the units 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 design constraints of the technical solution.

[0069] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment, depending on actual needs.

[0070] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any changes 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.

Claims

1. A valve life prediction method based on error calibration, characterized in that, Includes the following steps: Obtain valve type and working scenario, construct multi-dimensional prediction indicators, set the error range of prediction indicators, sort out and classify error sources, and construct valve life standard set; Collect valve lifecycle data and organize the collected data to construct a full lifecycle related dataset; Preprocessing and feature enhancement fusion of data in the full-cycle associated dataset are performed to form a model structured dataset; Construct and train the core prediction model, collect real-time training data of the valve, and input the real-time training data of the valve into the core prediction model to obtain preliminary prediction results; Based on the analysis of the model's structured dataset, the correlation is identified, the error propagation path and error type are clarified, an error calibration model is constructed, and the preliminary prediction results are calibrated to obtain preliminary calibration results. Acquire real-time operating data, analyze the data, correct preliminary calibration results, and formulate valve maintenance plans based on valve importance classification.

2. The valve life prediction method based on error calibration according to claim 1, characterized in that: Multidimensional predictive indicators include valve anomaly indicators, current health status, and remaining lifespan. The steps to construct a valve lifespan standard set are as follows: Determine multidimensional predictive indicators, set valve anomaly type levels, current health status assessment values, and lifespan units to form a predictive indicator framework; Establish error ranges for valve anomaly indicators, current health status, and remaining lifespan; The sources of error throughout the entire process are identified and categorized into inherent model error, data acquisition error, and operating condition fluctuation error, and quantifiable error description indicators are established. The prediction indicators, error range, error sources, and error description indicators are summarized to construct a standard set of valve lifespan.

3. The valve life prediction method based on error calibration according to claim 2, characterized in that: The full lifecycle data of a valve includes basic data and error data. Basic data includes design and manufacturing data, operation monitoring data, fault maintenance data, and other data. Design and manufacturing data includes valve model, valve material, design rated pressure, design rated flow, and design life. Operation monitoring data includes medium pressure, temperature, flow rate, valve opening and closing times, and valve stem vibration amplitude collected by sensors. Fault maintenance data includes historical fault types, fault occurrence time, fault handling measures, maintenance records, and equipment component information. Other data includes data supplemented through simulation when data is missing. Error data includes core deviation data, error correlation characteristics, and error source labels. Core deviation data includes the difference between each preliminary prediction result and the actual value. Error correlation characteristics include sensor accuracy level, data sampling frequency, operating condition fluctuation amplitude, and data missing rate. Error source labels include timestamps and deviation annotations that have been tested and verified.

4. The valve life prediction method based on error calibration according to claim 3, characterized in that: The process of organizing the collected data includes: extracting the full-cycle data of the valves, matching the error data with the basic data using timestamps, establishing a correlation relationship to obtain the associated data, verifying the completeness of the associated data, supplementing the missing error data or basic data, organizing the associated data and converting it into a structured table, attaching a unique timestamp to the associated data and numbering it, and obtaining the full-cycle associated dataset.

5. The valve life prediction method based on error calibration according to claim 4, characterized in that: Preprocessing and feature enhancement fusion of data in the full-cycle associated dataset includes: Outliers in the basic data were removed using the Raida criterion, missing data were filled using the linear imputation method, and sensor noise was smoothed using the moving average method. Extract fundamental features related to multidimensional prediction indicators from the base data after noise smoothing; The extracted basic features were processed using the standard score method; Analyze error data, calculate error quantification indicators, calculate the Pearson correlation coefficient between error quantification indicators and error descriptive indicators, and screen out error features. The basic features and error features that have undergone standardization are horizontally concatenated, and anomaly labels, status labels, lifetime labels and error labels are added to obtain fused feature data; The feature data is divided and integrated according to proportions to form a training set, a validation set, and a test set, which constitute the model's structured dataset.

6. The valve life prediction method based on error calibration according to claim 5, characterized in that: Before constructing and training the core prediction model, causal inference and statistical analysis were used to clarify the impact of the increase in the anomaly type level on the state assessment value and the difference in the life loss rate under different state assessment values. The basic features related to the multidimensional prediction indicators include the time series features related to valve anomaly identification, the state features related to the current health status, and the life features related to the remaining life. The core prediction model consists of an input layer, an anomaly detection layer, a state evaluation layer, a lifetime prediction layer, and an output layer. The input layer receives temporal features, state features, and lifetime features, which are then standardized and concatenated before being transmitted to the anomaly detection layer, state evaluation layer, and lifetime prediction layer, respectively. The anomaly detection layer takes the temporal features as input, processes them through convolutional kernels, pooling for dimensionality reduction, and fully connected neurons, and then uses the Softmax activation function for classification, outputting the anomaly type level and the corresponding confidence score. The state evaluation layer takes the confidence score corresponding to the anomaly type level and the state features as input, and uses a confidence threshold to remove low-confidence features. The system calculates the comprehensive anomaly impact value by combining the anomaly type level with the anomaly interaction mechanism. After concatenation, the values ​​are mapped to a high-dimensional space using a Gaussian kernel in a support vector machine. The system then classifies and transforms the output state evaluation value and evaluation confidence. The lifespan prediction layer takes the state evaluation value, evaluation confidence, and lifespan features as input. After concatenation, the system uses a long short-term memory network to extract features. After fully connected processing by neurons, the system outputs the preliminary remaining lifespan and lifespan fluctuation range. The output layer summarizes the anomaly type level, the confidence level corresponding to the anomaly type level, the state evaluation value, the evaluation confidence, the preliminary remaining lifespan, and the lifespan fluctuation range. After processing, the preliminary prediction results are obtained.

7. The valve life prediction method based on error calibration according to claim 6, characterized in that: The error calibration model consists of an input layer, an error prediction layer, an error calibration layer, and an output layer. The input layer receives preliminary prediction results and error features, concatenates them, standardizes them, and outputs structured error feature data. The error prediction layer takes structured error feature data as input, uses the Gini coefficient to screen key features, and predicts and outputs anomaly type level error, state classification error probability, and lifetime deviation rate through gradient boosting decision tree. The error calibration layer uses anomaly type level error, state classification error probability, and lifetime deviation rate to calibrate the preliminary prediction results. The output layer takes the anomaly type level, state assessment value, remaining lifetime, anomaly type level error range, state assessment value error range, and lifetime error range as input after error calibration to obtain the preliminary calibration results.

8. The valve life prediction method based on error calibration according to claim 7, characterized in that: The steps for analyzing operating data and correcting the preliminary calibration results are as follows: Real-time acquisition of on-site working condition data and sensor noise values, and comparison with the working condition data threshold range and noise threshold to determine whether there are working condition data fluctuations exceeding the working condition data threshold range or sensor noise values ​​exceeding the noise threshold range. If the fluctuation of the operating data exceeds the threshold range of the operating data, the remaining life and life error range shall be adjusted in accordance with the error propagation law. If a sensor noise value exceeds the noise threshold range, the anomaly type level and anomaly type level error range are corrected according to the error association rules, and the state evaluation value is recalculated based on the corrected anomaly type level and anomaly type level error range to obtain the corrected state evaluation value.

9. The valve life prediction method based on error calibration according to claim 8, characterized in that: The steps for developing a valve maintenance plan based on the valve's importance classification are as follows: Extract and correct the preliminary calibration results, find the valve importance level, determine whether it is a critical valve, and mark the maintenance priority; The preliminary calibration results were analyzed to obtain the maintenance deviation adjustment value; The preset maintenance date range is adjusted based on the maintenance deviation adjustment value to obtain the corrected maintenance date range; Summarize and correct the maintenance date range, and retrieve the monitoring strategies corresponding to the corrected status assessment values ​​to form a valve maintenance plan.

10. A valve life prediction system based on error calibration, characterized in that, include: Basic Analysis Module: Obtain valve type and working scenario, construct multi-dimensional prediction indicators, set the error range of prediction indicators, sort out and classify error sources, and construct valve life standard set; Data acquisition module: Collects valve data throughout its entire lifecycle, processes the collected data, and constructs a full-cycle related dataset; Data adjustment module: preprocesses and enhances features in the full-cycle associated dataset to form a structured dataset for the model; Preliminary prediction module: Constructs and trains the core prediction model, collects real-time training data of the valve, and inputs the real-time training data of the valve into the core prediction model to obtain preliminary prediction results; Error calibration module: Based on the analysis of the correlation of the model's structured dataset, the error propagation path and error type are identified, an error calibration model is constructed, the preliminary prediction results are calibrated, and the preliminary calibration results are obtained; Real-time correction module: Acquires real-time operating data, analyzes the operating data, corrects the preliminary calibration results, and formulates valve maintenance plans based on the valve importance classification.

Images