Instrument metering period prediction method based on causal inference and data volume self-adaption

Abstract

The invention provides an instrument metering period prediction method based on causal inference and data volume self-adaption, and the method comprises the steps: firstly constructing a causal graph model of instrument performance degradation, and effectively distinguishing a real driving mechanism from a false statistical association; on this basis, according to the number of currently available historical calibration samples, the modeling process is divided into three intervals of a minimum sample, a transition sample and a sufficient sample, and an adaptive heterogeneous prediction strategy is configured for each interval: each strategy independently outputs a metering period prediction value and confidence evaluation thereof; if the current strategy output confidence coefficient is lower than a preset threshold value, step-by-step degradation is automatically triggered, a more conservative modeling strategy is returned, and an artificial review alarm is sent out; if the confidence coefficient is still insufficient after degradation, a strong conservative period is adopted, and manual intervention is forcibly prompted. And finally, outputting a metering period recommendation result with high robustness and high interpretability. According to the invention, on the premise of ensuring causal rationality, adaptive, high-precision, safe and reliable metering period intelligent prediction across data volume scenes is realized.

Description

Technical Field

[0001] This invention relates to the field of instrumentation technology, and in particular to a method for predicting instrument measurement cycles based on causal inference and adaptive data volume. Background Technology

[0002] The term "instrument" as used in this article includes intelligent field devices used for industrial process monitoring and control. It covers not only traditional measuring instruments but also intelligent actuators with status sensing, self-diagnosis, or communication capabilities, such as control valves with valve positioners and intelligent on / off valves.

[0003] In modern industrial automation systems, pressure gauges, flow meters, and actuators, as core instruments for sensing pressure, flow, and performing actions, are widely used in critical process industries such as petrochemicals, power generation, and metallurgy. To ensure the accuracy of process control and system safety, these instruments must be periodically calibrated or validated. Traditional calibration cycles are usually based on national verification regulations or equipment manufacturer recommendations, using fixed time intervals or relying on human experience. Such methods ignore individual equipment differences, actual operating conditions, and dynamic environmental changes, easily leading to overcalibration or undercalibration.

[0004] In recent years, with the development of the Industrial Internet of Things (IIoT) and intelligent operation and maintenance technologies, some research has attempted to introduce data-driven methods to optimize metering cycles. For example, time series models or machine learning regressors are built based on historical calibration error sequences to predict instrument performance degradation trends and dynamically adjust calibration timing. However, these methods face three core challenges in practical deployment: First, it confuses statistical correlation with causal mechanisms, thus limiting its generalization ability.

[0005] Existing models often directly use the statistical correlation between environmental variables and measurement errors for prediction, without distinguishing whether they constitute a causal driver. For example, for electromagnetic flowmeters, the upstream valve state and measurement drift are co-variable, but the underlying cause is the change in fluid conductivity. This "spurious correlation" is prone to failure when operating conditions change, leading to prediction bias or even misjudgment.

[0006] Second, it lacks adaptive modeling capabilities for scenarios with heterogeneous data volumes.

[0007] Industrial field instruments are generally characterized by long lifecycles and sparse calibration records, resulting in huge differences in the available sample size: some devices have only 2-3 calibration records (extremely small sample), while a few key devices may have accumulated multiple data sets (sufficient sample). Existing methods mostly adopt a single model architecture, which suffers from severe overfitting under small sample sizes and fails to realize the model's potential under large data sizes, making it difficult to achieve a smooth transition and consistent performance across data volume ranges.

[0008] Third, there is a lack of effective knowledge transfer and confidence enhancement mechanisms in extremely small sample sizes.

[0009] For new equipment or low-frequency calibration instruments, existing solutions often directly adopt general rule models, which cannot make personalized predictions based on the historical experience of similar equipment. Even when transfer learning is introduced, there is a lack of quantitative evaluation of the reliability of the transfer results and a safe integration mechanism, which leads to either over-reliance on unreliable transfer results or complete abandonment of the information gain brought by transfer, making it impossible to achieve a balance between personalization and safety.

[0010] In summary, current technologies have not yet resolved three major bottlenecks: unclear causal mechanisms, weak data adaptability, and lack of small-sample knowledge transfer capabilities. Therefore, there is an urgent need for an intelligent metering cycle prediction method and system that integrates causal inference, supports adaptive data volume switching, and possesses device family transfer enhancement and dynamic confidence calibration capabilities. This would enable high-precision, robust, and reliable calibration strategy optimization while ensuring causal interpretability and decision security, and would be particularly suitable for three typical industrial instruments: pressure gauges, flow meters, and actuators. Summary of the Invention

[0011] This invention discloses a method for predicting instrument metering cycles based on causal inference and adaptive data volume, characterized by the following specific method: Acquire the operating data, environmental parameters, and maintenance metadata of the instrument to be predicted, and clean and preprocess the acquired data to form structured input data; Construct a causal graph model of instrument performance degradation to identify key variables that have a causal effect on measurement error from structured input data; Determine the current data range based on the amount of data for the key variables; Retrieve the prediction model corresponding to the data range, and the prediction model outputs the measurement period with confidence level; confidence level Perform real-time verification and decide whether to accept the prediction results or trigger a degradation based on their reliability.

[0012] Furthermore, the instrument includes any one of the following: Pressure gauge, flow meter, actuator; The pressure gauge's operating data includes: real-time pressure measurements. Historical calibration error Pressure fluctuation frequency ; The environmental parameters of the pressure gauge include: ambient temperature. Medium pH value, season; The maintenance metadata for the pressure gauge includes: calibration timestamp. Installation location, type of sensitive component; The flow meter's operating data includes: volumetric flow rate. Historical calibration error Root mean square velocity Upstream valve opening ; The environmental parameters of the flow meter include: fluid conductivity. Pipeline vibration acceleration ; The flow meter's maintenance metadata includes: calibration timestamp. Media type; The operating data of the actuators include: valve position feedback value, historical action error, and equivalent action wear factor. Control system load ; The environmental parameters of the instrument include: gas supply pressure fluctuations. ; The maintenance metadata for the instrument includes: number of actions and calibration timestamp. Pneumatic valve models.

[0013] Furthermore, operational data, environmental parameters, and maintenance metadata are represented as follows: in, Let be the feature vector of the i-th record, which contains runtime data, environmental parameters, and operation and maintenance metadata; To correspond to the calibration error, For timestamps; The acquired data is cleaned and preprocessed using the following methods: The data is preprocessed, and anomaly calibration points are identified using the Isolation Forest algorithm. Suspicious points are not included in the modeling. Missing values ​​are imputed using temporal proximity linear interpolation or LSTM prediction, and feature normalization is performed to obtain the final effective sample set. The number of samples is denoted as N.

[0014] Furthermore, key variables that have a causal effect on measurement error are identified from the structured input data, using the following specific methods: For each instrument to be predicted, define a set of variables and pre-set causal edge constraints; Solve the learned cause-effect graph; A robust ACE estimate is obtained by using causal forest for nonparametric estimation. Based on the calculated ACE values, the candidate variables are sorted and filtered, and the ACE values ​​that meet the conditions are retained.

[0015] Furthermore, the data range includes: The interval with the smallest amount of data is 1. The condition for determining the amount of data in interval 1 is: This corresponds to a conservative model; The second interval corresponds to a moderate amount of data. The condition for determining the amount of data in the second interval is as follows: This corresponds to a dual-branch hybrid model; The data volume corresponds to interval three, and the data volume judgment condition for interval three is: This corresponds to a deep time series model.

[0016] Furthermore, if the amount of data for the current key variable belongs to interval one and transfer learning is not possible, i.e. The corresponding prediction model is retrieved as a rule-guided conservative model, which uses the key variable feature set. As input, the formula for predicting the measurement cycle is: In the formula, To recommend the initial metering cycle, the operating condition severity factor α∈[1.0,2.0] and the environmental interference level β∈[0.8,1.2] are considered. α can be manually set or automatically extracted from the operation and maintenance logs. The confidence formula is: in, The baseline confidence level reflects the conservatism of the rule model; It is an environmental stability correction term; It is less than the preset confidence level.

[0017] Furthermore, if the current data volume of the key variable belongs to interval two, that is... The corresponding prediction model is a two-branch hybrid model; The dual-branch hybrid model includes a data-driven branch and a conservative branch. The conservative branch uses the measurement period and confidence level output by the interval-one model. The data-driven branch prediction measurement cycle method is as follows: A linear regression model is used to fit the historical error sequence, assuming that the error grows linearly with time: in, This is the current error. The slope is estimated using the least squares method. Let the maximum permissible error threshold be... The formula for predicting the measurement cycle is: like This indicates stable performance and an extended cycle. ; The confidence assessment of the data-driven branch is as follows: in, To regress the residual variance, For typical noise levels, when If the model is deemed unreliable, the confidence level is forcibly reduced to 0.1. The prediction period of the two-branch hybrid model is determined by both branches, as shown in the formula: The weights are adjusted by both confidence level and relevance, as shown in the formula: function The correlation decay factor is calculated using the following formula: when The decay mechanism is activated at that time; The final output confidence score takes into account both branch confidence and consensus strength. in ∈[0,0.2] is the consensus reward coefficient, which is used to moderately increase the confidence when the two branches have low correlation but high consistency in prediction period.

[0018] Furthermore, if the current key variable's data volume belongs to interval three, that is... The corresponding prediction model is retrieved as a deep time series model; Deep time series models employ LSTM, TCN, or XGBoost models; Input features include time series calibration error values Causal driving characteristics The model outputs the expected error at a certain future time. Let the maximum permissible error specified by the industry or standard be... The next forecast measurement period will be: If the prediction error is consistently lower than [a certain value] within the preset time window If so, the cycle will be extended to the preset upper limit; The Monte Carlo Dropout method is used to quantify prediction uncertainty. During the inference phase, the model is propagated M times to obtain the prediction distribution. Calculate the prediction variance The confidence level of a deep time series model is defined as: in, The variance was calibrated through historical backtesting.

[0019] Furthermore, since the data volume of the current key variable belongs to interval one, transfer learning will be applied to it using the following method: Determine if a database of causal isomorphic device families exists that shares the same causal relationship as the key variable; If it exists, then the data volume of the causal isomorphic device family database reaches the data volume of interval two, that is... Then Bayesian adaptation is performed, utilizing device family priors. A small number of calibration points for this equipment Calculate the posterior slope distribution The migration prediction period is generated by using the posterior mean as the degradation slope. With confidence level ; If it exists, then the data volume of the causal isomorphic device family database reaches the data volume of interval three, that is... Then the pre-trained PINN model weights ,by Using the prior mean, fine-tuning or Laplace approximation is performed on the new equipment data to obtain the posterior weights. The inverse solution of the prediction error trajectory is then used to obtain the posterior weights. Confidence level Calculated based on prediction variance; set up and As a threshold, if ≥ Directly adopt This is considered a high-confidence personalized rule; if ≤ < : Call the basic rule model of interval one to obtain , The two are transformed into Basic Probability Assignment (BPA), fused using Dempster's combination rule, and a conflict coefficient threshold is set. If the conflict coefficient K> Abandon integration and regress to ;like < Directly adopt ; The final output is still labeled as the interval-one strategy result, and the confidence level can be dynamically adjusted to reflect the information gain brought about by the transfer.

[0020] Furthermore, regarding confidence level Real-time verification is performed, and the decision to accept the prediction results or trigger degradation is based on their reliability. The specific method is as follows: Set the lower confidence threshold If the current policy outputs confidence level If so, then accept its prediction period; if If so, the current model is determined to be unreliable, triggering the degradation mechanism; If the confidence level is still lower than [a certain value] after downgrading Then a conservative cycle will be forced to be adopted, such as And issue an alarm; When downgrading, the minimum effective time and historical confidence moving average are used for judgment.

[0021] Due to the adoption of the above technical solutions, this application has the following beneficial effects: 1. By introducing a causal inference mechanism, the model effectively distinguishes between the true causal effects and spurious statistical associations of environmental and operating conditions on instrument errors, significantly improving the model's generalization ability and interpretability under operating condition transfer.

[0022] 2. Through an adaptive modeling strategy based on data volume, rule-based models, dual-branch hybrid models, and physical constraint deep models are configured for three typical industrial scenarios: extremely small samples, transitional samples, and sufficient samples, respectively, to achieve intelligent switching between "stable when data is scarce and accurate when data is plentiful".

[0023] 3. By constructing a family of causal isomorphic devices, the system can achieve personalized transfer reasoning with extremely small samples, and by combining evidence theory with fusion rules and transfer results, it can significantly improve the prediction accuracy of new devices while ensuring security.

[0024] 4. Through a confidence-driven dynamic verification and step-by-step degradation mechanism, the system automatically reverts to a more conservative strategy when the model output is unreliable and triggers a manual review alarm, thereby fundamentally ensuring decision security and avoiding the risk of undercalibration.

[0025] 5. The entire solution does not rely on accelerated aging tests or a large amount of historical data. Under the sparse calibration conditions commonly found in real industrial sites, it can still achieve high-precision, high-robustness, and high-reliability intelligent prediction of measurement cycles, effectively reducing the waste of resources caused by over-calibration, while eliminating measurement inaccuracies and safety accidents caused by calibration delays, and significantly improving the level of intelligence and operation and maintenance economy of the instrument's entire life cycle management.

[0026] Other advantages, objectives, and features of the invention will be set forth in part in the description which follows, and in part will be apparent to those skilled in the art from the following examination, or may be learned from practice of the invention. The objectives and other advantages of the invention can be realized and obtained through the following description. Attached Figure Description

[0027] The accompanying drawings of this invention are described below.

[0028] Figure 1 This is a schematic diagram of the overall process of the present invention. Detailed Implementation

[0029] The present invention will be further described below with reference to the accompanying drawings and embodiments.

[0030] This application provides a method for predicting instrument measurement cycles based on causal inference and adaptive data volume. The method first constructs a causal graph model of instrument performance degradation, explicitly identifying the causal effects of environmental factors and operating conditions on measurement errors, effectively distinguishing between genuine driving mechanisms and spurious statistical associations. Based on this, according to the number of available historical calibration samples, the modeling process is divided into three intervals: minimal samples, transitional samples, and sufficient samples. Adaptive heterogeneous prediction strategies are configured for each interval: a conservative model based on prior rules is used under minimal sample conditions; if causally isomorphic historical data of equipment families exists, personalized predictions are generated through Bayesian transfer learning, and the predictions are then applied in the transfer learning process. When the reliability is at a moderate level, the Dempster-Shafer evidence theory is combined with the transfer results and the local rule model to improve prediction accuracy and confidence. A hybrid model of rule-based and data-driven approaches is used for transitional samples; a deep learning model is used for sufficient samples. Each strategy independently outputs the predicted value of the measurement period and its confidence assessment. The system further introduces a dynamic verification mechanism driven by a confidence threshold: if the confidence of the current strategy output is lower than a preset threshold, a step-by-step downgrade is automatically triggered, reverting to a more conservative modeling strategy and issuing a manual review alert; if the confidence is still insufficient after downgrading, a strongly conservative period is adopted and manual intervention is forcibly prompted. The final output is a highly robust and highly interpretable measurement period recommendation result. This method, while ensuring causal rationality, achieves adaptive, high-precision, and secure intelligent prediction of measurement periods across data volume scenarios.

[0031] The steps of the instrument metering cycle prediction method are as follows: Figure 1 As shown, the details are as follows: S1. Obtain the operating data, environmental parameters, and maintenance metadata of the instrument to be predicted. Clean and preprocess the acquired data to form structured input data.

[0032] This embodiment uses a pressure gauge, an electromagnetic flow meter, and a pneumatic valve as examples, which are explained below: Pressure gauge, collecting operating data such as real-time pressure measurements. Historical calibration error Pressure fluctuation frequency (Number of pressure changes exceeding the threshold per unit time, used to calculate the fatigue cycle count of pressure-sensitive components) etc.; environmental parameters such as ambient temperature Medium pH value; maintenance metadata such as calibration timestamps Installation location, type of sensitive components, etc.

[0033] Collect operating data from the flow meter, such as volumetric flow rate. Historical calibration error Root mean square velocity Environmental parameters such as fluid conductivity Pipeline vibration acceleration Operational metadata such as calibration timestamps Media type, etc.

[0034] Collect operational data for pneumatic valves, such as valve position feedback values, historical action errors, and equivalent action wear factors; environmental parameters, such as air supply pressure; and maintenance metadata, such as the number of actions and calibration timestamps. Pneumatic valve models, etc.

[0035] Data collection for other types of flow meters and pneumatic valves can be carried out in the same manner.

[0036] All data is aligned with the calibration data to form a structured record set: in, Let be the feature vector of the i-th record, which contains process data, environmental parameter information, and operation and maintenance metadata. To correspond to the calibration error, The data is then preprocessed, using the Isolation Forest algorithm to identify outlier calibration points. Suspicious points are not included in the modeling process. Missing values ​​are filled using time-nearest linear interpolation or LSTM prediction, followed by feature normalization, ultimately yielding a valid sample set. The number of samples is denoted as N, and it will be used in subsequent modules.

[0037] S2. Construct a causal graph model of instrument performance degradation, and identify key variables that have a causal effect on measurement error from structured input data. The specific steps are as follows: S21. Construct a candidate set of causal graphs. For the three types of instruments, define variable sets and preset causal edge constraints respectively: S211, Pressure Gauge The target variable is calibration error. The difference between the pressure gauge reading and the true value is the core degradation indicator to be predicted.

[0038] The candidate variables and causal path analysis are shown in the table below: variable Physical meaning and Relationship Pressure fluctuation frequency The number of times the pressure exceeds the threshold per unit time High-frequency fluctuations → repeated deformation of elastic sensitive elements → fatigue → zero-point drift / sensitivity reduction Ambient temperature Ambient temperature at the installation location Temperature change → Change in elastic modulus of elastic sensing element → Output offset medium pH value Acidity and alkalinity of process media Strong acids / bases → corrosion of elastic sensitive elements → degradation of mechanical properties season Spring / Summer / Autumn / Winter Season itself does not directly affect pressure gauges, but summer can lead to changes in ambient temperature. An increase in humidity during winter can lead to changes in the viscosity / pH of the medium. … … … Based on the analysis in the table above, the core failure mode of the pressure gauge is fatigue of the elastic sensing element, and the fatigue life is positively correlated with the number of stress cycles. The mechanism constraint is forced edge. Season is a high-order abstract variable that does not produce physical forces or chemical reactions. To prevent the model from learning spurious correlations such as "larger errors in winter," forbidden edges are defined as follows: .

[0039] S212, Electromagnetic Flow Meter The target variable is calibration error. This refers to the difference between the flow meter's measured value and the true value. The candidate variables and causal path analysis are shown in the table below: variable Physical meaning and Relationship Fluid conductivity Dielectric conductivity Electromagnetic flowmeters are based on Faraday's law of electromagnetic induction: induced electromotive force <imgwi="20.49" he="3.56" file="O9XTFGiU8BLturGc613lEQg7BfNUqKLDv681ws5o.jpg" img-format="jpg"img-content="drawing" orientation="portrait"inline="no"> However, this is on the premise that the fluid is a conductor. If If the voltage is too low, the electrodes cannot form a stable circuit, resulting in signal distortion or zero-point drift. Pipeline vibration Mechanical vibration acceleration during pipe installation Vibration is transmitted to the flowmeter body through the support, causing micro-displacement of the coil / electrode, interfering with the uniformity of the magnetic field, and generating false signals. Upstream valve opening <imgwi="9.65" he="3.56" file="Lhel2LX1BWU4aq3kCEC78HBrTzSoibT9XLGizU5k.jpg" img-format="jpg" img-content="drawing"orientation="portrait" inline="no"> Control valve opening percentage Changes in valve opening → alterations in flow velocity distribution → may cause Local changes (such as bubble formation) may induce pipe vibration; however, the valve itself does not directly act on the electromagnetic induction process. … … … Based on the analysis of the table above, the working principle of the electromagnetic flowmeter essentially relies on conductivity, even if the data... and This causal path must also be preserved; otherwise, the model will completely fail under low conductivity conditions. Therefore, the mechanistic constraint forces the edges to be... The valve is a control system output and does not produce physical or chemical changes. To prevent the model from learning spurious correlations such as "large error when opening is 50%", a prohibited edge is set as... Other types of flow meters can be analyzed in the same way.

[0040] S213, Pneumatic valve The target variable is action error. This refers to the difference between the actual valve position and the commanded valve position. The candidate variables and causal path analysis are shown in the table below: variable Physical meaning and Relationship Equivalent wear factor <imgwi="6.94" he="4.91" file="6WZ4WPCTuR3UT4frWLMzyBWcLLIziSiQkKEb9mgj.jpg" img-format="jpg" img-content="drawing"orientation="portrait" inline="no"> Based on the wear accumulation of the combined number of actuations, pressure difference, and stroke, a wear model is constructed using Archard's wear theory. The larger the value, the more severe the irreversible mechanical degradation of components such as the valve stem, sealing ring, and spring, which will lead to increased friction, increased hysteresis, and zero-point drift. Gas supply pressure fluctuation <imgwi="8.30" he="4.23" file="Vc0k16rjWl0cvtUk6mTIbULqckUs9RK7i0PQbSpI.jpg" img-format="jpg" img-content="drawing"orientation="portrait" inline="no"> Control the standard deviation or peak-to-peak value of the gas source pressure pneumatic actuator output force ,like Fluctuations will lead to unstable output force, affecting the valve position control accuracy. The air supply pressure is the direct source of the actuator's driving force, and its stability directly affects the steady-state error. Control system load <imgwi="6.94" he="4.23" file="zPmnWn4xmSPWwPt8apqJdGTDGSBxoXpaisLPI1m4.jpg" img-format="jpg" img-content="drawing"orientation="portrait" inline="no"> DCS / PLC CPU load or communication delay High load → delayed command issuance → lag in valve position response; however, the load does not change the mechanical or pneumatic characteristics of the actuator itself. … … … Based on the analysis of the table above, irreversible mechanical wear is the main cause of actuator failure. Even if the wear effect is not significant in the data of new equipment, this degradation path is still forcibly modeled. Therefore, the mechanism constraint forces the edge to be... The control system load is an information layer variable and does not affect the physical execution layer. To avoid misjudging "large errors when the DCS is busy" as an actuator problem and to prevent incorrect maintenance, a prohibited edge is set. Other types of implementing agencies can be analyzed in the same way.

[0041] S22. NOTEARS employs a continuously optimized causal discovery method from... By learning the causal adjacency matrix, the above mechanistic constraints are transformed into a mask matrix in NOTEARS optimization. ,in Indicates prohibition , And when it is a forced edge, initialize , The magnitude reflects the strength of the causal effect between variables. ).

[0042] For each type of instrument, its set of specific variables is used to construct a data matrix. (Where N is the number of samples and d is the total number of variables), solve the NOTEARS problem with mechanistic constraints: in, is the sparsity regularization coefficient, used to encourage the obtaining of sparse causal structures; These are smoothness equality constraints that ensure the graph structure is a DAG; Represents the Hadamard product. This represents matrix exponentiation. The augmented Lagrange method is used to solve this non-convex optimization problem, ultimately yielding a continuous weight matrix. The positions of forced edges are directly preserved. Subsequently, through... Apply threshold to the remaining elements The binarized causal adjacency matrix is ​​obtained. .

[0043] This matrix is ​​the learned causal graph, which is used to subsequently screen key drivers that have a significant causal effect on instrument measurement errors.

[0044] S23. For each candidate parent node The average causal effect ACE on error E is calculated to accurately assess its actual driving force on instrument performance degradation, and key causal variables are selected accordingly. in, Represents the variable Perform intervention operations (i.e., manually set its value) ), and Represent Two typical value levels, such as high temperature and normal temperature, are usually selected from the upper / lower quartile of the data distribution or the industry standard threshold. Indicating intervention for Under the condition, error The expected value.

[0045] Since intervention experiments cannot be directly conducted in real-world scenarios, estimations must be performed using observational data and a backdoor adjustment formula, under the assumptions of "no unobserved confounding factors" or "all confounding factors have been controlled." in for and The common cause (backdoor variable); For conditional expectation; backdoor variable The marginal probability density function.

[0046] Next, we will use causal forest for nonparametric estimation. The core idea is to sort the samples in each subtree according to... The values ​​are divided into two groups, and are controlled... Under the premise of [missing information], the average effect difference between the two groups is estimated. Finally, a robust ACE estimate is obtained by averaging across multiple trees.

[0047] The causal effects of the three types of instruments, constrained by mechanistic factors, are quantified as follows. Other types of flow meter actuators can be analyzed following the same steps: S231, Pressure gauge Candidate parent node is Target error There is no direct mixed path (due to seasonal S being prohibited from direct connection). ), and Possibly covariant, therefore backdoor variable The backdoor formula is then adjusted to Next, nonparametric estimation is performed. When splitting each tree in the causal forest, the parameter is forced to... and Included as a covariate, the conditional expectation is calculated for each leaf node. Through Monte Carlo integration (on history) and (Sampling) Approximate outer layer integration, the final estimate is: and represent Two typical value levels.

[0048] S232, Electromagnetic Flowmeter Candidate parent node is Target error Upstream valves in mixed paths and ,but no and The common reason (because) Not valid, and prohibited. ), and If covariance is possible, then the backdoor variable... The backdoor formula is then adjusted to Next, nonparametric estimation is performed, following the same steps as for the pressure gauge. In the causal forest, only... As a control variable, estimate .

[0049] S233, Pneumatic valve Candidate parent node is Target error , Calculated from historical action data, it is unaffected by other observed variables, therefore it is a backdoor variable. The backdoor formula is then adjusted to Next, nonparametric estimation is performed, followed by direct fitting. right Conditional expectation, estimate .

[0050] S24. Based on the calculated ACE values, sort and filter the candidate variables, and set the minimum significant effect threshold. , , The standard deviation of historical calibration error, coefficient Based on the device's degradation sensitivity settings, a balance is struck between sensitivity and robustness. Only settings that meet the requirements are retained. For variables, the sign of ACE must be consistent with the physical degradation mechanism of the equipment. If the direction is contradictory, even if the amplitude meets the standard, it is considered as abnormal data or uncontrolled contamination and should be discarded. For core variables that inevitably lead to degradation in terms of mechanism, even if data sparsity causes degradation, the sign of ACE must be consistent with the physical degradation mechanism of the equipment. It is still mandatory to include it in the feature set to avoid misjudging key degradation paths due to short-term data bias.

[0051] Fatigue mechanism of elastic sensing element in pressure gauge: Repeated pressure fluctuations → cyclic plastic deformation or accumulation of micro-damage in the material → positive zero-point drift (higher reading), requiring... If a negative value is calculated, it may be due to incorrect calibration data labeling or unobserved contamination. In this case, it should be marked as an anomaly and removed from the variable. pH is only causally related when the membrane material is stainless steel and the medium is a strong acid / base; otherwise, it is not causally related. Usually less than Automatic rejection means that the final output will generally be... .

[0052] The mechanism of an electromagnetic flowmeter (taking the low-conductivity sensitive type as an example) is as follows: →Increased electrode polarization → Decreased induced potential → Lower measured value, therefore, it is required If it is a high conductivity type (such as concentrated salt water). If the change has a small impact and the sign may not be significant, then the amplitude is the primary factor, and the sign check is exempted. The system automatically determines whether to enable sign constraints based on the device model. This usually leads to the superposition of positive noise, therefore However, because the impact of vibration mainly increases variance, the ACE amplitude is often small, and it is necessary to ensure... The final output may be .

[0053] Mechanical wear of pneumatic valves is an irreversible physical process. This inevitably leads to the accumulation of motion errors. Even if the new equipment has only been calibrated 2-3 times, the ACE estimation variance is large and may be close to 0, so it is still forcibly retained. ,like If so, manually adjust ACE to an appropriate value. Gas supply fluctuation handling: Increased driving force leads to fluctuations in driving force, which in turn reduces positioning accuracy. It needs to meet the amplitude threshold. Therefore, the final output may be... .

[0054] Other types of flow meters, pressure gauges, and actuators can be screened and analyzed following the steps outlined above.

[0055] The above feature set , , It serves as the sole input variable for the multi-strategy modeling and confidence assessment module.

[0056] S3. Determine the interval to which the current key variable belongs.

[0057] Based on the typical data sparsity characteristics of industrial field instruments, the modeling process is divided into three data intervals: Interval 1 (very limited data): (like The data is extremely scarce and insufficient to support the training of complex models. Interval 2 (transitional data volume): (like The data volume is moderate and can support simple data-driven models; Interval 3 (sufficient data): It has abundant data and can train complex deep models.

[0058] and It can be preset according to equipment type or industry standards, or determined through historical data statistics. All interval boundary values ​​can be dynamically adjusted to adapt to different application scenarios.

[0059] To avoid decision oscillations caused by frequent model switching due to minor data fluctuations, a minimum switching interval mechanism is also introduced, which sets a minimum switching interval. That is, after the most recent model switch, at least wait Only after a certain timeframe will switching be allowed again. Switching is triggered only after new calibration data has been entered into the database and preprocessed; if the current model output confidence is too low, switching will not occur even if the timeframe has not yet been reached. This also forces a downgrade check to be triggered.

[0060] S4. If the current key variable belongs to interval one and can be matched with a causal isomorphic device family database that shares the same causal relationship as the key variable, then transfer learning is performed on the current key variable. The specific method is as follows: S41. Construction of Causal Isomorphic Device Family Database To support personalized predictions in scenarios with extremely small sample sizes, the system pre-builds a database of causal isomorphic device families. The device family partitioning follows these principles: The equipment types are consistent: such as "DN50 stainless steel pressure gauge" or "DN80 electromagnetic flowmeter"; the operating environments are similar: the installation location, medium type, and temperature / pressure range are consistent; the causal structure is the same: the causal graph topology learned through NOTEARS is consistent.

[0061] Each device family stores a set of historical calibration records. Causal-driven feature set The prior of the degenerate model is the linear slope distribution of the interval two-device family storage. The pre-trained PINN weights are stored in the three device families within the interval.

[0062] S42, When the device is in interval one ( When a matching device family exists, the system executes the following enhancement process, which is divided into two migration strategies based on the historical data volume of the device family: S421, Bayesian Quick Adaptation Scenario 1: Matched to a device family in interval two (all devices within the family are transitional samples). ) Perform Bayesian adaptation, utilizing device family priors A small number of calibration points for this equipment Calculate the posterior slope distribution: The migration prediction period is generated by using the posterior mean as the degradation slope. With confidence level .

[0063] Scenario 2: Matching a three-device family within the interval (the devices within the family are sufficient samples). ) Prior form: pre-trained PINN model weights ,by Using the prior mean, fine-tuning or Laplace approximation is performed on the new equipment data to obtain the posterior weights. The inverse solution of the prediction error trajectory is then used to obtain the posterior weights. Confidence level Calculated based on prediction variance.

[0064] Regardless of the type of prior used, the following steps are performed: set up and As a threshold, if ≥ Directly adopt This is considered a high-confidence personalized rule; if ≤ < : Call the basic rule model of interval 1 to obtain , The two are transformed into Basic Probability Assignments (BPA), fused using Dempster's combination rule, and a conflict coefficient threshold is set. If the conflict coefficient K> Abandon integration and regress to ;like < Directly adopt .

[0065] The final output is still labeled as the interval-one strategy result, but the confidence level can be dynamically adjusted to reflect the information gain brought about by the transfer.

[0066] S5. Retrieve the corresponding prediction model based on the key variables that have undergone transfer learning or those that have not. The specific method is as follows: S51, Interval 1 (very limited data), corresponds to a rule-guided conservative model. when At that time, the system activates the expert rule model. This model is not trained on data, but is built based on manufacturer specifications, industry standards, and equipment model degradation patterns.

[0067] Let the recommended initial metering period be... (From manufacturer specifications) Considering the operating condition severity factor α∈[1.0,2.0] and the environmental disturbance level β∈[0.8,1.2], the prediction period is as follows.

[0068] α is set manually based on the equipment installation location (such as high-temperature zone, vibration zone) or automatically extracted from the operation and maintenance log. β can be calculated based on the fluctuation range of environmental parameters in the past three months. This model has no training process, and the output is stable but lacks flexibility.

[0069] Due to a lack of supporting data, its confidence level is low and fixed, and its formula is as follows: in, The baseline confidence level reflects the conservatism of the rule model; This is an environmental stability correction term; if there have been no drastic fluctuations in the past three months, it will increase slightly. The total confidence level does not exceed 0.65, reflecting its uncertainty due to missing data.

[0070] S52, Interval 2 (transitional data volume), corresponding to the dual-branch hybrid model. when At this time, the system activates a two-branch hybrid model, which includes a conservative branch and a data-driven branch. Details are as follows: S521, Conservative Branch Same interval, same model, output period With confidence level .

[0071] S522, Data-Driven Branch A linear regression model is used to fit the historical error sequence. It is assumed that the error grows linearly over time. in, This is the current error. Let be the slope, estimated using the least squares method. Let the maximum permissible error threshold be... The predicted next calibration cycle is: like This indicates stable performance and an extended cycle. .

[0072] S523, the confidence level of the conservative branch remains the same as that of interval 1, that is... Data-driven branch confidence Quantification is based on the relative magnitude of the model residual variance and the system noise baseline, as shown in the following formula: in, To regress the residual variance, For typical noise levels (e.g., 1.5 times the square of sensor accuracy), when When the residuals are close to the system noise, the model is considered unreliable, and the confidence level is forcibly reduced to 0.1. This formula ensures that when the model residuals are close to the system noise, the confidence level approaches 1; when the residuals are significantly greater than the noise, the confidence level decays rapidly, reflecting the reliability of the model.

[0073] S524. If the prediction results of the two branches are always highly correlated, the benefit of fusion is not significant. To avoid redundant fusion, this invention introduces a branch prediction correlation evaluation mechanism before weighting, calculating the covariance of the prediction values ​​of the two branches within a sliding historical window, such as the two most recent predictions. (If there are not enough historical prediction values, this step is skipped, and the default fusion weights are used.) in, The correlation coefficient. If... (like This indicates that the two branches are highly consistent or redundant, and the fusion gain is limited. In this case, the weight of the data-driven branch should be reduced, and a conservative approach should be adopted. (like If both points to similar periods (e.g., relative deviation <20%), it is considered a "multi-perspective consensus" to improve overall confidence.

[0074] S525, its prediction period is determined by two branches, and the weighted formula is as follows: The weights are modulated by both confidence level and relevance: function For example, a correlation decay factor: The final output confidence score takes into account both branch confidence and consensus strength. in ∈[0,0.2] is the consensus reward coefficient, which is used to moderately increase the confidence when the two branches have low correlation but high consistency in prediction period.

[0075] S53, Interval 3 (sufficient data), corresponding to the deep time series model When the number of historical valid calibration samples At this time, the system activates a deep time-series model to perform refined modeling and prediction of the instrument performance degradation trend. Details are as follows: S531. LSTM, TCN, or XGBoost models are used for error trend prediction. Input features include time-series calibration error values. Causal driving characteristics The model outputs the expected error at a certain future time. Let the maximum permissible error specified by the industry or standard be... The predicted next calibration period is: If the prediction error is consistently lower than [a certain value] within the preset time window If so, the cycle will be extended to the preset upper limit. (The upper limit can be determined based on relevant regulations and experience).

[0076] S532. The Monte Carlo Dropout method is used to quantify prediction uncertainty. During the inference phase, the model is propagated M times (randomly discarding some neurons each time) to obtain the prediction distribution. Calculate the prediction variance The model confidence level is defined as: in, The variance is calibrated using historical backtesting (e.g., taking the 75th percentile of the historical prediction error variance).

[0077] S6. Combine key variables with the prediction model to output a measurement period with confidence level.

[0078] S7. Based on the confidence level, perform prediction downgrade and output the final result.

[0079] Step S7: Confidence level of the current activation policy output Real-time verification is performed, and the decision to accept the prediction results or trigger degradation is based on their reliability. The specific method is as follows: S71. Set the lower confidence threshold. If the current policy outputs confidence level If so, then accept its prediction period; if If the current model is deemed unreliable, a degradation mechanism will be triggered.

[0080] S72. The downgrade direction follows the principle of "retreating to a more conservative strategy that relies less on data." If the current interval is three and... Then downgrade to interval two, rerun the interval two model, and output... , Similarly, for interval two, if the confidence level is lower than the threshold, it will be downgraded to interval one; interval one is the bottom line of the rule and will never be downgraded.

[0081] S73. If the confidence level is still lower than [the previous level] after downgrading Then a conservative cycle will be forced to be adopted, such as It also issued an alert: "The model confidence level remains low. It is recommended to arrange manual calibration and evaluation."

[0082] S74. To avoid frequent degradation due to accidental noise, a minimum effective time (the same device will not be repeatedly downgraded within the minimum effective time) and a historical confidence moving average (the average of the most recent 3 confidence values ​​is used for judgment, rather than a single value) are introduced.

[0083] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit it. Although the present invention has been described in detail with reference to the above embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the specific implementation of the present invention. Any modifications or equivalent substitutions that do not depart from the spirit and scope of the present invention should be covered within the scope of protection of the claims of the present invention.

Claims

1. A method for predicting instrument metering cycles based on causal inference and adaptive data volume, characterized in that, The specific method is as follows: Acquire the operating data, environmental parameters, and maintenance metadata of the instrument to be predicted, and clean and preprocess the acquired data to form structured input data; Construct a causal graph model of instrument performance degradation to identify key variables that have a causal effect on measurement error from structured input data; Determine the current data range based on the amount of data for the key variables; Retrieve the prediction model corresponding to the data range, and the prediction model outputs the measurement period with confidence level; confidence level Perform real-time verification and decide whether to accept the prediction results or trigger a degradation based on their reliability.

2. The instrument metering cycle prediction method based on causal inference and adaptive data volume as described in claim 1, characterized in that, The instrument includes any one of the following: Pressure gauge, flow meter, actuator; The pressure gauge's operating data includes: real-time pressure measurements. Historical calibration error Pressure fluctuation frequency ; The environmental parameters of the pressure gauge include: ambient temperature. Medium pH value, season; The maintenance metadata for the pressure gauge includes: calibration timestamp. Installation location, type of sensitive component; The flow meter's operating data includes: volumetric flow rate. Historical calibration error Root mean square velocity Upstream valve opening ; The environmental parameters of the flow meter include: fluid conductivity. Pipeline vibration acceleration ; The flow meter's maintenance metadata includes: calibration timestamp. Media type; The operating data of the instrument includes: valve position feedback value, historical action error, and equivalent action wear factor. Control system load ; The environmental parameters of the instrument include: gas supply pressure fluctuations. ; The maintenance metadata for the instrument includes: number of actions and calibration timestamp. Pneumatic valve models.

3. The instrument metering cycle prediction method based on causal inference and adaptive data volume as described in claim 1, characterized in that, Operational data, environment parameters, and maintenance metadata are represented as follows: in, Let be the feature vector of the i-th record, which contains runtime data, environmental parameters, and operation and maintenance metadata; To correspond to the calibration error, For timestamps; The acquired data is cleaned and preprocessed using the following methods: The data is preprocessed, and anomaly calibration points are identified using the Isolation Forest algorithm. Suspicious points are not included in the modeling. Missing values ​​are imputed using temporal proximity linear interpolation or LSTM prediction, and feature normalization is performed to obtain the final effective sample set. The number of samples is denoted as N.

4. The instrument metering cycle prediction method based on causal inference and adaptive data volume as described in claim 1, characterized in that, The following method is used to identify key variables that have a causal effect on measurement error from structured input data: For each instrument to be predicted, define a set of variables and pre-set causal edge constraints; Solve the learned cause-effect graph; A robust ACE estimate is obtained by using causal forest for nonparametric estimation. Based on the calculated ACE values, the candidate variables are sorted and filtered, and the ACE values ​​that meet the conditions are retained.

5. The instrument metering cycle prediction method based on causal inference and adaptive data volume as described in claim 1, characterized in that, The data range includes: The interval with the smallest amount of data is 1. The condition for determining the amount of data in interval 1 is: This corresponds to a conservative model; The second interval corresponds to a moderate amount of data. The condition for determining the amount of data in the second interval is as follows: This corresponds to a dual-branch hybrid model; The data volume corresponds to interval three, and the data volume judgment condition for interval three is: This corresponds to a deep time series model.

6. The instrument metering cycle prediction method based on causal inference and data volume adaptation as described in claim 5, characterized in that, If the amount of data for the current key variable belongs to interval one and transfer learning is not possible, i.e. The corresponding prediction model is retrieved as a rule-guided conservative model, which uses the key variable feature set. As input, the formula for predicting the measurement cycle is: In the formula, To recommend the initial metering cycle, the operating condition severity factor α∈[1.0,2.0] and the environmental interference level β∈[0.8,1.2] are considered. α can be manually set or automatically extracted from the operation and maintenance logs. The confidence formula is: in, The baseline confidence level reflects the conservatism of the rule model; It is an environmental stability correction term; It is less than the preset confidence level.

7. The instrument metering cycle prediction method based on causal inference and adaptive data volume as described in claim 6, characterized in that, If the current data volume of the key variable belongs to interval two, that is... The corresponding prediction model is a two-branch hybrid model; The dual-branch hybrid model includes a data-driven branch and a conservative branch. The conservative branch uses the measurement period and confidence level output by the interval-one model. The data-driven branch prediction measurement cycle method is as follows: A linear regression model is used to fit the historical error sequence, assuming that the error grows linearly with time: in, This is the current error. The slope is estimated using the least squares method. Let the maximum permissible error threshold be... The formula for predicting the measurement cycle is: like This indicates stable performance and an extended cycle. ; The confidence assessment of the data-driven branch is as follows: in, To regress the residual variance, For typical noise levels, when If the model is deemed unreliable, the confidence level is forcibly reduced to 0.

1. The prediction period of the two-branch hybrid model is determined by both branches, as shown in the formula: The weights are adjusted by both confidence level and relevance, as shown in the formula: function The correlation decay factor is calculated using the following formula: when The decay mechanism is activated at that time; The final output confidence score takes into account both branch confidence and consensus strength. in ∈[0,0.2] is the consensus reward coefficient, which is used to moderately increase the confidence when the two branches have low correlation but high consistency in prediction period.

8. The instrument metering cycle prediction method based on causal inference and adaptive data volume as described in claim 5, characterized in that, If the current data volume of the key variable belongs to interval three, that is... The corresponding prediction model is retrieved as a deep time series model; Deep time series models employ LSTM, TCN, or XGBoost models; Input features include time series calibration error values Causal driving characteristics The model outputs the expected error at a certain future time. Let the maximum permissible error specified by the industry or standard be... The next forecast measurement period will be: If the prediction error is consistently lower than [a certain value] within the preset time window If so, the cycle will be extended to the preset upper limit; The Monte Carlo Dropout method is used to quantify prediction uncertainty. During the inference phase, the model is forwarded M times to obtain the prediction distribution. Calculate the prediction variance The confidence level of a deep time series model is defined as: in, The variance was calibrated through historical backtesting.

9. The instrument metering cycle prediction method based on causal inference and data volume adaptation as described in claim 6, characterized in that, The current key variable's data volume belongs to interval one. Transfer learning will be applied to it using the following method: Determine if a database of causal isomorphic device families exists that shares the same causal relationship as the key variable; If it exists, then the data volume of the causal isomorphic device family database reaches the data volume of interval two, that is... Then Bayesian adaptation is performed, utilizing device family priors. A small number of calibration points for this equipment Calculate the posterior slope distribution The migration prediction period is generated by using the posterior mean as the degradation slope. With confidence level ; If it exists, then the data volume of the causal isomorphic device family database reaches the data volume of interval three, that is... Then the pre-trained PINN model weights ,by Using the prior mean, fine-tuning or Laplace approximation is performed on the new equipment data to obtain the posterior weights. The inverse solution of the prediction error trajectory is then used to obtain the posterior weights. Confidence level Calculated based on prediction variance; set up and As a threshold, if ≥ Directly adopt This is considered a high-confidence personalized rule; if ≤ < : Call the basic rule model of interval one to obtain , The two are transformed into Basic Probability Assignment (BPA), fused using Dempster's combination rule, and a conflict coefficient threshold is set. If the conflict coefficient K> Abandon integration and regress to ;like < Directly adopt ; The final output is still labeled as the interval-one strategy result, and the confidence level can be dynamically adjusted to reflect the information gain brought about by the transfer.

10. The instrument metering cycle prediction method based on causal inference and adaptive data volume as described in claim 1, characterized in that, confidence level Real-time verification is performed, and the decision to accept the prediction results or trigger degradation is based on their reliability. The specific method is as follows: Set the lower confidence threshold If the current policy outputs confidence level If so, then accept its prediction period; if If so, the current model is determined to be unreliable, triggering the degradation mechanism; If the confidence level is still lower than [a certain value] after downgrading Then a conservative cycle will be forced to be adopted, such as And issue an alarm; When downgrading, the minimum effective time and historical confidence moving average are used for judgment.

Images