Real-time monitoring control system and controller based on thermal flow sensor

Abstract

The invention discloses a real-time monitoring control system based on a thermal flow sensor and a controller, and relates to the technical field of sensor monitoring. According to the real-time monitoring control system based on the thermal flow sensor and the controller, the method comprises the steps that S1, sensor monitoring data and historical statistical data are preprocessed; s2, lagging and response time dynamic characteristics are extracted; performing modal decomposition, phase estimation, peak calculation and steady-state noise analysis; s3, executing phase compensation, peak value recovery and enhancement compensation judgment; recording a residual sequence and a compensation parameter, and outputting optimized flow data; s4, constructing a credibility interval and dynamic consistency score synthesis control credibility mark; and executing security constraint and prediction filtering and archiving. The problems that an existing thermal flow sensor generates obvious phase lag and peak smoothing distortion due to the influence of thermal diffusion thermal inertia under the rapid fluctuation and pulsation working conditions, and the real-time monitoring precision and the closed-loop control stability are difficult to guarantee due to the lack of an online dynamic response analysis and compensation mechanism are solved.

Description

Technical Field

[0001] This invention relates to the field of sensor monitoring technology, specifically to a real-time monitoring and control system and controller based on a thermal flow sensor. Background Technology

[0002] Thermal gas flow sensors, due to their compact structure, wide measurement range, and ability to directly measure mass flow rate, are widely used in medical equipment, small gas systems, precision industrial process control, and various embedded gas supply devices. In such systems, flow monitoring typically works in conjunction with valve actuators, pump drive units, pipeline pressure measurement devices, and ambient temperature acquisition units to form a real-time detection and regulation chain around the gas flow state. As gas systems evolve towards rapid regulation, low flow resolution, periodic pulsating delivery, and multi-scenario coupled control, the real-time acquired flow data, control command data, pressure data, and temperature data exhibit higher dynamism and multi-source characteristics. The system's reliance on data alignment, trend identification, operating condition correlation analysis, and loop stability is constantly increasing. To ensure the reliability of flow monitoring and control, related applications generally require the use of historical statistical data, real-time multi-source signals, and gas path operating characteristics to comprehensively judge flow change behavior, thereby supporting closed-loop regulation, operation monitoring, and continuous stable operation.

[0003] For example, the invention patent with publication number CN120670710A discloses a flushing and cooling method, system, and device with automatic route identification. This method collects spatial point cloud data of the device surface using a three-dimensional laser scanning sensor and constructs a three-dimensional structural model. It combines a material identification sensor and a temperature acquisition network to obtain material properties, thermal conductivity, and multi-source temperature gradient distribution. Based on sensor data fusion, it establishes a heat load distribution map and a heat dissipation demand model, and determines the flushing point sequence and shortest route according to a data-driven path planning algorithm. Subsequently, based on the regional heat load density, a parameter adaptive algorithm generates control parameters such as flushing pressure, flow rate, and duration. During the flushing process, it monitors temperature change data in real time and dynamically adjusts the flushing parameters based on sensor feedback, thereby achieving automated flushing and cooling control based on sensor perception and data decision-making.

[0004] For example, the invention patent with announcement number CN116854156B discloses a method and system for monitoring and purifying factory wastewater quality based on the Internet of Things. This method collects monitoring data such as pollutant concentration, flow rate, and flow rate during the wastewater discharge process through a distributed water quality and flow sensor network. It uses the sensor data analysis results to construct a diffusion model of pollutants along the discharge path, and sends control commands to the purification end based on the model to determine the retention time, purification amount, and disinfectant dosage. At the same time, it combines the discharge status data recorded by the sensors with the actual operation feedback from the purification end to determine whether the working limit has been reached, and adjusts the purification operation when the limit is reached. The system as a whole takes sensor acquisition, data fusion, diffusion modeling, and control linkage as its core to realize the collaborative management of wastewater monitoring and purification processes.

[0005] Although existing sensor monitoring and data processing technologies can achieve multi-source acquisition, state identification, and simple dynamic analysis, existing thermal flow monitoring methods generally lack accurate quantification of phase lag and peak clipping distortion under rapidly fluctuating conditions. Online dynamic response analysis often relies on fixed thresholds or empirical corrections, making it difficult to achieve real-time and accurate separation and compensation for lag, peak clipping, and noise. At the same time, there is a lack of online judgment mechanisms based on residuals, dynamic consistency, and reliable boundaries, which makes it impossible to form closed-loop constraints on data reliability and control security during the compensation process, thus hindering the assurance of monitoring accuracy and control stability.

[0006] Therefore, in order to address the above problems, there is an urgent need for a real-time monitoring and control system and controller based on thermal flow sensors. Summary of the Invention

[0007] Technical problems to be solved

[0008] To address the shortcomings of existing technologies, this invention provides a real-time monitoring and control system and controller based on a thermal flow sensor. It solves the problems of existing thermal flow sensors causing significant phase lag and peak clipping distortion due to thermal diffusion and thermal inertia under rapid fluctuation and pulsating conditions, and the lack of online dynamic response analysis and compensation mechanisms, which makes it difficult to guarantee real-time monitoring accuracy and closed-loop control stability.

[0009] Technical solution

[0010] To achieve the above objectives, the present invention provides the following technical solution: a real-time monitoring and control system based on a thermal flow sensor, comprising: S1, acquiring sensor monitoring data and obtaining historical statistical data; preprocessing the sensor monitoring data and historical statistical data; S2, constructing an edge window based on the sensor monitoring data, extracting dynamic features of hysteresis and response time; and performing mode decomposition, phase estimation, peak value calculation, and steady-state noise analysis; S3, statistically generating trigger thresholds based on dynamic features and performing phase compensation, peak value recovery, and enhancement compensation determination; correcting dynamic distortion and recording residual sequences and updating compensation parameters, outputting flow data optimized for monitoring accuracy; S4, constructing a confidence interval based on the residual sequence and synthesizing a control confidence label with the dynamic consistency score; performing safety constraints and predictive filtering based on the control confidence label and completing the entire process archiving.

[0011] Furthermore, the specific process of collecting sensor monitoring data and obtaining historical statistical data is as follows: Sensor monitoring data is collected, including instantaneous flow rate data, execution sequence, sensor temperature data, heating drive data, pipeline pressure data, valve opening command data, air pump drive signal data, and sampling timestamp data; historical statistical data is obtained and a historical statistical database is created, including historical instantaneous flow rate data, historical sensor temperature data, historical valve opening command data, historical air pump drive signal data, and historical steady-state change rate; among which, heating drive data refers to the sampled values ​​of heating drive current and heating drive voltage; the sampling time interval is obtained by subtracting the timestamps.

[0012] Furthermore, the specific process for preprocessing sensor monitoring data and historical statistical data is as follows: instantaneous flow data, pipeline pressure data, and sensor temperature data are de-glitched using a sliding window midpoint algorithm to remove isolated spikes and communication jitter anomalies; the data sequences in sensor monitoring data and historical statistical data are time-aligned based on sampling timestamp data; and sensor monitoring data and historical statistical data are standardized and normalized using distribution standardization and linear normalization algorithms.

[0013] Furthermore, the specific process of constructing an edge window based on sensor monitoring data and extracting the dynamic features of hysteresis and response time is as follows: When a valve actuator is present, the valve opening command data is used as the control input reference; when a pneumatic pump actuator is present, the pneumatic pump drive signal data is used as the control input reference; Bayesian online change point detection is performed on the control input reference sequence and pipeline pressure data to extract the edge occurrence timestamp; edge detection with piecewise monotonic constraints is performed on the instantaneous flow data to extract the flow response edge timestamp; the difference between the two types of edge timestamps is calculated under the same time base of the sampling timestamp data to obtain the monitoring edge hysteresis time; the edge window is obtained by jointly determining the edge detection results of the control input reference data and pipeline pressure data under the same time base of the sampling timestamp data; within the edge window, percentage response estimation is further performed based on quantile regression; the response rise time and response fall time are obtained by locating the percentage response point of the instantaneous flow data within the same window; and the response rise time and response fall time are calculated.

[0014] Furthermore, the specific process of performing mode decomposition, phase estimation, peak value calculation, and steady-state noise analysis is as follows: Variational mode decomposition is performed on the instantaneous flow data within the pulsation window to separate the main pulsating mode and high-frequency noise mode. The pulsation window is jointly determined by the instantaneous flow data and the control input reference data under the same time base of the sampling timestamp data through periodic feature detection. For the main pulsating mode, a generalized cross-correlation phase transform is used to estimate the periodic phase lag time. A Hilbert transform is used to obtain the instantaneous phase sequence. The phase difference between the control input reference phase and the main pulsating phase of the flow is calculated, and the periodic phase lag time is output. Peak reduction distortion is calculated by comparing the reference peak amplitude with the actual peak amplitude. The steady-state window is determined based on the short-time variance and first-order difference amplitude of the instantaneous flow data and pipeline pressure data, as well as the change point detection results of the control input reference data. Historical instantaneous noise analysis is then performed. Within a steady-state window, the static amplitude mapping coefficient from the control input reference to the instantaneous flow rate is established using recursive least squares with a forgetting factor, based on the flow rate data, historical valve opening command data, and historical air pump drive signal data. Within a pulsating window, the peak value variation of the control input reference data is determined using a short-time peak detection method. The reference peak value amplitude is calculated by multiplying the static amplitude mapping coefficient by the peak value variation, and the peak hold factor is calculated by the ratio of the actual peak value amplitude of the instantaneous flow rate data to the reference peak value amplitude. Simultaneously, for the high-frequency noise mode within the steady-state window, the steady-state noise intensity is obtained by combining the sliding window standard deviation and the envelope amplitude integral. The monitoring accuracy analysis results are output, including: monitoring edge lag time, response rise time, response fall time, periodic phase lag time, peak hold factor, and steady-state noise intensity.

[0015] Furthermore, the specific process of generating trigger thresholds based on dynamic characteristics and performing phase compensation, peak recovery, and enhancement compensation is as follows: Based on the steady-state window criterion and periodic structure stability, a strategy stability judgment window constructed within a continuous pulsation window is obtained as a stable window. Based on the statistical distribution of monitoring accuracy analysis results, a Bayesian quantile estimation algorithm is used to generate dynamic response optimization thresholds, including: lag time optimization threshold, coefficient optimization threshold, and response time optimization threshold. The periodic phase lag time, peak hold coefficient, response rise time, and response fall time are compared in real time with the dynamic response optimization thresholds: when the periodic phase lag time is greater than the lag time... When the optimized threshold is met and the stable window length condition is satisfied, the phase compensation strategy is triggered; when the periodic phase lag time is less than or equal to the lag time optimization threshold, the existing phase compensation state remains unchanged and monitoring continues; when the peak retention coefficient is less than the coefficient optimization threshold and the stable window length condition is met, the peak recovery strategy is triggered; when the peak retention coefficient is greater than or equal to the coefficient optimization threshold, the original peak compensation strategy remains unchanged and monitoring continues; when both the response rise time and response fall time are greater than the response time optimization threshold, the enhanced compensation strategy is triggered; when both the response rise time and response fall time are less than or equal to the response time optimization threshold, the current compensation strategy remains unchanged.

[0016] Furthermore, the specific process of correcting dynamic distortion, recording residual sequences, updating compensation parameters, and outputting optimized flow data for monitoring accuracy is as follows: The phase compensation strategy adopts an unscented Kalman filter state observation framework: instantaneous flow data is used as the observation, and control input reference data, heating drive current data, and heating drive voltage data are used as exogenous driving quantities. A dynamic equation including hysteresis states is established. The dynamic equation adopts a two-state structure: the main flow state is updated at each sampling time point based on the changing trend of the control input reference data, the main flow state at the previous time point, and the hysteresis energy state at the previous time point; the hysteresis energy state is updated based on the heating power composed of the hysteresis energy state at the previous time point, the heating drive current data, and the heating drive voltage data; a hysteresis-corrected flow estimate is output at each sampling time point. The hysteresis-corrected flow estimate is used as the predictor, and the instantaneous flow data is used as the observation to construct the unscented Kalman filter update process. The system records the unscented Kalman filter residual sequence and updates the compensation parameters. The peak recovery strategy employs amplitude-constrained regularized deconvolution reconstruction: a pulsation transfer function is constructed using the reference peak amplitude and peak preservation coefficient for deconvolution to recover peak information in the instantaneous flow data. The equivalent amplitude response of the pulsation transfer function is taken as the peak preservation coefficient. By using the equivalent amplitude response as the basis for amplitude compensation in the deconvolution reconstruction process, total variation regularized deconvolution recovery is performed on the instantaneous flow data, with steady-state noise intensity as the upper limit constraint on amplitude. The enhancement compensation strategy employs feedforward-feedback coupling: the feedforward term is obtained online from the control input reference data via recursive least squares with a forgetting factor, and the feedback term is obtained from the closed-loop correction of the unscented Kalman filter residual. The compensation parameters are corrected using sensor temperature segmentation, with the temperature segment boundaries obtained from historical sensor temperature data through Gaussian mixture model clustering. The output monitoring accuracy is optimized by recording flow data, trigger markers, and compensation parameters.

[0017] Furthermore, the specific process of constructing a confidence interval based on the residual sequence and synthesizing a control confidence label with the dynamic consistency score is as follows: Input the unscented Kalman filter residual sequence from the compensation parameter record to construct a flow residual distribution model. Use the unscented Kalman filter residual sequence as training samples, and fit the residual conditional quantile function using ordinal quantile regression to output the upper and lower bounds of the residual confidence interval. Superimpose these on the flow data optimized for monitoring accuracy to obtain the upper and lower bounds of the flow confidence interval. Use the current sampled value of the flow data optimized for monitoring accuracy as the flow value to be determined. When the flow value to be determined falls within the flow... When the lower bound of the confidence interval is between the upper bound of the flow confidence interval, it is marked as a risk-free interval; when the flow value to be determined exceeds the upper bound of the flow confidence interval, it is marked as a level one risk interval; when the flow value to be determined is lower than the lower bound of the flow confidence interval, it is marked as a level two risk interval. Dynamic consistency score is obtained by aligning the main pulsation mode of the monitoring accuracy optimization flow data and the control input reference data through dynamic time warping. Dynamic consistency score is constructed based on the periodic phase lag time and peak hold coefficient. The flow confidence interval, dynamic consistency score and monitoring accuracy analysis results are combined to form a control confidence label and output synchronously.

[0018] Furthermore, the specific process of implementing safety constraints and predictive filtering based on the control confidence mark and completing the entire process archiving is as follows: The control confidence mark is mapped to a closed-loop control safety filtering strategy, and the execution quantity output by the controller is constrained and corrected: When the control confidence mark enters the secondary risk interval, the execution quantity change rate is calculated from the execution quantity sequence output by the controller through the first-order difference between adjacent sampling timestamps. A monotonic safety boundary construction method based on the execution quantity change rate is adopted. The safety boundary of the execution quantity change rate is constructed by statistically analyzing the quantiles of the current execution quantity change rate and the historical steady-state change rate. A control barrier function is constructed using the current execution quantity change rate and the safety boundary. A control barrier function greater than zero is used as a safety constraint on the change rate, ensuring that the output execution quantity change rate is updated within the safety boundary. When the control confidence mark enters the primary risk interval, the model predictive safety filter is activated, and the model predicts safety... The full filter employs a finite prediction domain model predictive control structure, using a deviation penalty matrix and a change penalty matrix to apply soft constraints to flow deviations and execution changes. The prediction domain length is within the range of a to b sampling periods, depending on the control cycle and dynamic characteristics of the operating conditions. The sampling period is consistent with the sampling interval of the monitoring data. The model predictive safety filter uses the flow confidence interval as a hard constraint and the dynamic consistency score as a soft constraint to solve for the safe execution quantity and output a protection trigger flag. When the control confidence flag enters the risk-free interval, no special operation is required. A real-time monitoring database is created, and event-level archiving is performed. The archived content is indexed by the sampling timestamp and includes: monitoring accuracy optimized flow data, upper bound of the flow confidence interval, lower bound of the flow confidence interval, dynamic consistency score, control confidence flag, protection trigger flag, safe execution quantity correction magnitude, and corresponding monitoring accuracy analysis results.

[0019] A second aspect of the present invention provides a real-time monitoring controller based on a thermal flow sensor, comprising: a memory, a processor, and a computer program stored in the memory and executable on the processor.

[0020] Beneficial effects

[0021] The present invention has the following beneficial effects: (1) This invention, by constructing an edge window and extracting dynamic features such as lag time, response rise time, and response fall time, realizes the structured analysis and interpretable modeling of the real-time response process of the thermal flow sensor, thereby achieving accurate characterization of the dynamic change segment and effectively solving the problem of not being able to distinguish between the real response delay and the acquisition link delay in the prior art.

[0022] (2) This invention achieves two-dimensional explicit quantification of phase lag and peak compression under pulsating conditions by performing mode decomposition, phase estimation and peak retention calculation within the pulsating window, thereby achieving accurate positioning of dynamic distortion and effectively solving the problem of difficulty in identifying the source of phase drift and peak clipping in the prior art.

[0023] (3) This invention constructs trigger thresholds such as lag, coefficient and response time based on dynamic feature statistics, realizes dynamic adaptive compensation decision, improves the compensation trigger condition from empirical judgment to statistically driven reproducible mechanism, thereby improves the reliability of compensation strategy switching, and effectively solves the problems of easy strategy oscillation and uninterpretable threshold in the prior art.

[0024] (4) This invention establishes a unified dynamic distortion correction framework through unscented Kalman filtering, regularized deconvolution and feedforward-feedback coupling compensation, thereby achieving coordinated compensation for phase lag, peak flattening and temperature drift coupling errors, and thus significantly improving the accuracy of flow monitoring. It effectively solves the problem of scattered compensation chains and the inability to uniformly optimize them in the prior art.

[0025] Of course, any product implementing this invention does not necessarily need to achieve all of the advantages described above at the same time. Attached Figure Description

[0026] Figure 1 This is a structural diagram of the real-time monitoring and control system based on a thermal flow sensor according to the present invention. Figure 2 This is a flowchart of the strategy triggering logic of the present invention; Figure 3 This is a schematic diagram of the regularized deconvolution peak recovery of the present invention; Figure 4 This is a visual bar chart of the key indicators of the dynamic response before and after compensation in this invention. Detailed Implementation

[0027] 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.

[0028] Please see Figures 1-4This invention provides a technical solution: a real-time monitoring and control system based on a thermal flow sensor, comprising the following steps: S1, collecting sensor monitoring data and obtaining historical statistical data; preprocessing the sensor monitoring data and historical statistical data; S2, constructing an edge window based on the sensor monitoring data, extracting dynamic features of hysteresis and response time; and performing mode decomposition, phase estimation, peak calculation, and steady-state noise analysis; S3, statistically generating trigger thresholds based on dynamic features and performing phase compensation, peak recovery, and enhancement compensation determination; correcting dynamic distortion and recording residual sequences and updating compensation parameters, outputting flow data optimized for monitoring accuracy; S4, constructing a confidence interval based on the residual sequence and synthesizing a control confidence label with the dynamic consistency score; performing safety constraints and predictive filtering based on the control confidence label and completing the entire process archiving.

[0029] Specifically, the process of collecting sensor monitoring data and obtaining historical statistical data is as follows: collecting sensor monitoring data, which includes instantaneous flow data, execution sequence, sensor temperature data, heating drive data, pipeline pressure data, valve opening command data, air pump drive signal data, and sampling timestamp data; obtaining historical statistical data and creating a historical statistical database, which includes historical instantaneous flow data, historical sensor temperature data, historical valve opening command data, historical air pump drive signal data, and historical steady-state change rate.

[0030] Among them, heating drive data refers to the sampled values ​​of heating drive current and heating drive voltage, used to characterize the impact of heating power changes on dynamic response; instantaneous flow data refers to the sampled values ​​of gas mass flow directly output by the thermal gas flow sensor, used to characterize the real-time changes in gas flow; sensor temperature data refers to the sampled values ​​of temperature directly output by the internal temperature measurement unit of the thermal gas flow sensor, used to characterize the thermal equilibrium state and temperature drift disturbance; pipeline pressure data refers to the sampled values ​​of pipeline pressure directly output by the differential pressure sensor or pressure sensor, used to construct a consistency verification reference for fluctuating operating conditions; valve opening command data and air pump drive signal data refer to the sampled values ​​of control quantities directly output by the controller, used to provide edge references for changes in operating conditions; sampling timestamp data refers to the time identifier directly generated by the data acquisition unit for each sampled data, used for multi-source data alignment and monitoring delay calculation; the sampling time interval used to cover the pulsating main frequency and edge changes is obtained by subtracting the timestamps, avoiding the introduction of apparent phase lag by the sampling mechanism and reducing the risk of peak clipping distortion being amplified by sampling.

[0031] In this implementation plan, a complete dynamic operating condition perception foundation is formed by collecting sensor monitoring data and historical statistical data, combined with time-domain and frequency-domain analysis. This achieves unification of the monitoring link at the time scale, dimensional scale, and noise structure levels, providing stable and reliable input conditions for subsequent dynamic feature extraction and compensation determination. This step establishes a consistent sampling time interval among multi-source data, ensuring that instantaneous flow data, execution sequence, sensor temperature data, heating drive data, pipeline pressure data, valve opening command data, and air pump drive signal data can be aligned with a common time base, allowing edge changes, pulsation frequency, and steady-state behavior to be accurately described within the same coordinate system. Simultaneously, this step analyzes the high-frequency fluctuation structure and power spectrum characteristics of the sensor monitoring data to obtain a sampling time interval covering the pulsation frequency, reducing the amplification effect of sampling lag on periodic phase lag time, peak hold coefficient, and monitoring edge lag time, thus improving the accuracy of subsequent dynamic response analysis. A monitoring data input system suitable for dynamic operating conditions is constructed, providing a continuous, accurate, and traceable data foundation for subsequent edge window construction, lag feature extraction, pulsation feature calculation, and compensation strategy triggering.

[0032] Specifically, the preprocessing of sensor monitoring data and historical statistical data involves: de-scratching instantaneous flow rate data, pipeline pressure data, and sensor temperature data using a sliding window median algorithm to remove isolated spikes and communication jitter anomalies, suppressing interference from non-physical jumps on phase lag and peak clipping index calculations. The sliding window length is tuned to 8 to 10 times the shortest physical change time constant corresponding to the sampling frequency, ensuring that the median filter only covers local short-term fluctuations without weakening the true dynamic structure. For the missing positions after filtering, a linear interpolation method maintaining monotonicity with adjacent effective sampling points is used for reconstruction. When the missing position is within the steady-state window, the mean of the steady-state segment is used as a substitute to ensure that the repaired sequence neither introduces spurious fluctuations nor compresses the true change amplitude. Using sampling timestamp data as a benchmark, sensor monitoring data and historical statistical data are time-aligned to ensure that multi-source data can be indexed synchronously at the same time, avoiding apparent lag caused by asynchrony. Linear interpolation and synchronization alignment strategies are used to form a consistent time grid for data from sources with different sampling frequencies and sampling delays before entering the analysis module. Sensor monitoring data and historical statistical data are standardized and normalized through distribution standardization and linear normalization algorithms. In the distribution standardization process, steady-state statistics are used as the quantification benchmark, with the mean based on the steady-state sample mean and the standard deviation based on the steady-state sample standard deviation. This avoids mismatch in scale estimation caused by dynamic operating condition drift, unifies the units and numerical scales, and improves the stability of subsequent monitoring accuracy assessment and optimization triggering.

[0033] In this implementation scheme, by eliminating anomalies in instantaneous flow rate data, pipeline pressure data, and sensor temperature data, and by unifying the time base and normalizing the dimensions, a stable, clean, and alignable multi-source monitoring data input environment is constructed. This ensures that the monitoring link maintains consistency and computability even under conditions of noise disturbance, time asynchrony, and dimensional differences, providing a reliable data foundation for extracting key dynamic features such as periodic phase lag time, peak hold coefficient, and monitoring edge lag time. After de-glitching, non-physical jumps no longer affect the calculation of lag and peak clipping features. After time alignment, all monitored quantities are indexed with a unified sampling timestamp, ensuring rigorous temporal consistency of edge events, pulsation behavior, and response processes. After standardization and normalization, the monitoring data from each sensor and historical statistical data are expressed in the same numerical domain, improving the robustness of the dynamic feature statistical distribution and enabling subsequent monitoring accuracy assessment and compensation strategy triggering to be determined based on a continuous and stable data structure. This step transforms multi-source monitoring data from its raw state into highly consistent analytical-grade data, laying the foundation for subsequent edge window construction, mode decomposition, dynamic feature extraction, and compensation strategy decision-making.

[0034] Specifically, the process of constructing an edge window based on sensor monitoring data and extracting dynamic features of hysteresis and response time is as follows: When a valve actuator is present, the valve opening command data is used as the control input reference; when a pneumatic pump actuator is present, the pneumatic pump drive signal data is used as the control input reference. The valve actuator refers to the regulating mechanism in the tested gas circuit system used to adjust the gas flow by changing the opening, and its control command is directly generated by the control system during operation; the pneumatic pump actuator refers to the power mechanism in the tested gas circuit system used to drive the gas flow by changing the output power, and its drive signal is directly generated by the external control system during operation. These are used to construct the edge and period reference of the operating condition change to support the phase hysteresis calculation. Bayesian online change point detection is performed on the control input reference sequence and pipeline pressure data to extract the edge occurrence timestamp. The Bayesian online change point detection uses the assumption of steady-state no abrupt change as a prior and the absolute increment of the control input gradient exceeding two to three times the steady-state standard deviation of the sequence as a candidate condition for change points. The posterior change point probability exceeds the confidence level of 0.95 to 1.Output the edge occurrence timestamp at 00:00; perform piecewise monotonicity constraint edge detection on instantaneous flow data, extract the flow response edge timestamp, and use a condition of not less than four times the standard deviation of steady-state noise for piecewise monotonicity constraint to make the monotonicity hold, replacing local samples that violate monotonicity with linear interpolation values ​​of adjacent segments to make edge positioning unaffected by local high-frequency noise; calculate the difference between the two types of edge timestamps under the same time base of the sampled timestamp data to obtain the monitoring edge lag time, and use the tenth to ninetieth percentile range between the control input reference change and the instantaneous flow change for the monitoring edge lag time. The response span is used as a lag point to ensure that the lag calculation is not affected by local perturbations, and is used to characterize the real-time performance of monitoring, covering thermal diffusion lag and acquisition link delay. Within the edge window, percentage response estimation based on quantile regression is further performed to calculate the response rise time and response fall time. Quantile regression uses quantile loss functions targeting the 20th, 50th, and 80th percentiles as optimization objectives, with L1-type quantile loss ensuring robustness to outliers. Iteration convergence is achieved through a certain number of iterations to ensure a consistent calculation benchmark for the percentage response point under different operating conditions. The edge window is defined by the control input reference data and... The edge detection results of pipeline pressure data are jointly determined under the same time base of the sampling timestamp data; Bayesian online change point detection is performed on the control input reference data to obtain the control edge occurrence time; edge detection with monotonic constraints is performed on the pipeline pressure data to obtain the gas pressure change time; edge detection with piecewise monotonic constraints is performed on the instantaneous flow rate data to obtain the instantaneous flow rate change time; an edge window is constructed with the control edge occurrence time as the starting point and the later of the gas pressure change time and the instantaneous flow rate change time as the ending point. The edge window is used to cover the changes in the actual operating conditions from the controlled edge. The entire process from system initiation to the sensor generating an observable response ensures that the calculation of response rise time and fall time is based solely on dynamically changing valid data segments. Both rise and fall times are determined by locating percentage response points within the same window of instantaneous flow data. The rise time utilizes a rise segment from the 20th percentile to the 80th percentile, while the fall time utilizes a fall segment from the 80th percentile to the 20th percentile. This ensures comparability and consistency across different operating conditions, characterizing the sensor's real-time response and thermal inertia.

[0035] In this implementation scheme, by constructing an edge detection link under a unified time base, a precise characterization of the entire process of operating condition changes is achieved, establishing a dynamic feature system that can simultaneously reflect control-side actions, gas path propagation, and sensor-side responses. Through joint edge detection of control input references, pipeline pressure data, and instantaneous flow data, a complete edge window covering the entire process from actuator action to observable flow changes is formed. This ensures that the monitored edge lag time accurately reflects the combined effects of thermal diffusion lag, gas path transmission delay, and acquisition link delay. Within the unified edge window, the response rise time and response fall time are extracted using percentage response points, strictly limiting the response process calculation to the effective dynamic segment. This avoids interference from steady-state noise and non-physical fluctuations, ensuring that the response time characteristics accurately reflect the sensor's real-time performance, thermal inertia changes, and dynamic attenuation characteristics. This step achieves the alignment and fusion of multi-source edge information, enabling lag, response speed, and dynamic change structure to be quantified in a physically consistent manner, constructing the core dynamic diagnostic foundation required for subsequent mode decomposition, phase estimation, and compensation strategy triggering.

[0036] Specifically, the process of performing mode decomposition, phase estimation, peak value calculation, and steady-state noise analysis is as follows: Variational mode decomposition is performed on the instantaneous flow data within the pulsation window. The number of modes in the variational mode decomposition is in the range of 2–4 to ensure physical decoupling between the main pulsating mode and the high-frequency noise mode. A bandwidth constraint coefficient of 1000–5000 is used to limit the bandwidth spread of each mode. The bandwidth constraint is achieved by applying a square integrable constraint to the first derivative of the analytic signal, ensuring the main mode maintains its narrowband nature and avoiding noise mode aliasing, thus separating the main pulsating mode from the high-frequency noise mode. The pulsation window is defined by… Instantaneous flow rate data and control input reference data are jointly determined through periodic feature detection under the same time base of the sampling timestamp data. First, a short-time autocorrelation function is calculated for the instantaneous flow rate data, and the peak position of the main period is identified in the autocorrelation function. Simultaneously, a short-time power spectral density is calculated for the control input reference data, and the position of the main frequency is identified in the power spectral density. The time length corresponding to the main period is used as the pulsation period, and multiple periodic segments with consecutive main periodic structures are used as the boundaries of the pulsation window to construct a pulsation window. The pulsation window is used to cover the continuous time period in which the gas flow rate exhibits periodic and quasi-periodic changes, so that the variational mode... Decomposition can separate the main pulsating mode and the noise mode under a unified structure, improving the stability of phase lag and peak clipping distortion measurements. For the main pulsating mode, a generalized cross-correlation phase transform is used to estimate the periodic phase lag time. A Hilbert transform is used to obtain the instantaneous phase sequence. The phase difference between the control input reference phase and the main pulsating phase of the flow is calculated, and the periodic phase lag time is output to stabilize and quantize the phase lag, reducing phase jitter introduced by noise. The generalized cross-correlation phase transform employs logarithmic amplitude normalization and phase spectrum preservation, requiring the sampling rate of the control input reference data to be completely consistent with that of the main pulsating mode to ensure phase stability. The results are comparable; the calculation of Hilbert envelope and instantaneous phase requires the construction of analytical signals based on the same sampling grid, and the mirror extension of the edge region is performed to reduce envelope distortion under the limited window, thereby ensuring the continuity and stability of phase transformation within the pulsating window; peak clipping distortion is calculated by comparing the reference peak amplitude with the actual peak amplitude: the steady-state window is determined based on the short-time variance, first-order differential amplitude of instantaneous flow data and pipeline pressure data, and the change point detection results of control input reference data, and historical instantaneous flow data, historical valve opening command data, and historical air pump drive signal data are used within the steady-state window;The steady-state window is used to cover the period when the gas flow rate remains stable without significant fluctuations or abrupt changes. This ensures that the recursive least squares fitting of the relationship between the control input reference and the instantaneous flow rate is based solely on the physical steady-state segment, avoiding the impact of dynamic lag and peak clipping distortion on the static mapping coefficients. A static amplitude mapping coefficient between the control input reference and the instantaneous flow rate is established using recursive least squares with a forgetting factor. The range of the forgetting factor is determined based on the variance level and time correlation of the historical steady-state rate of change sequence: the autocorrelation coefficient is calculated for the historical steady-state rate of change sequence. When the autocorrelation decays to below 0.1 within 3–5 sampling periods, this decay period is taken as the system memory length. Based on this, the exponential decay rate of the recursive least squares for samples exceeding the memory length is inferred, resulting in the usable range of the forgetting factor, which typically falls within... Between 0.95 and 0.995, to ensure algorithm stability, positive definiteness of the covariance matrix and non-divergence of the condition number are used as convergence conditions within this range, ensuring that the static amplitude mapping coefficients maintain convergence and interpretability in both dynamic fluctuations and steady-state conditions. Within the pulsation window, the peak change of the control input reference data is determined using a short-time peak detection method. The reference peak amplitude is calculated by multiplying the static amplitude mapping coefficients by the peak change. The peak retention coefficient is calculated by the ratio of the actual peak amplitude of the instantaneous flow data to the reference peak amplitude. The actual peak amplitude originates from the local extreme points of the instantaneous flow data within the pulsation window, reflecting the true output peak of the sensor within that pulsation cycle. This directly measures the degree of peak clipping and compression, corresponding one-to-one with the peak deviation problem under rapid fluctuation conditions.

[0037] Simultaneously, for high-frequency noise modes within the steady-state window, the steady-state noise intensity is obtained by combining the sliding window standard deviation with the envelope amplitude integral: The instantaneous amplitude envelope is obtained by performing a Hilbert transform on the high-frequency noise mode; the average envelope amplitude is calculated within the steady-state window to obtain the mean of the steady-state noise envelope; subsequently, the short-time standard deviation of the high-frequency noise mode is calculated using a sliding window within the same steady-state window; the average of all short-time standard deviations is taken to obtain the steady-state noise standard deviation; the weighted combination of the mean of the steady-state noise envelope and the steady-state noise standard deviation is used as the steady-state noise intensity to quantitatively characterize the noise energy level under steady-state conditions, outputting the monitoring accuracy analysis results. These results include: monitoring edge lag time, response rise time, response fall time, periodic phase lag time, peak hold coefficient, and steady-state noise intensity. This provides direct input for threshold determination and strategy selection in the monitoring accuracy compensation and peak recovery modules, avoiding secondary peak clipping caused by empirical parameter tuning.

[0038] In this implementation scheme, a structured dynamic feature system is constructed within a unified pulsation window and a steady-state window, enabling the partitioned extraction of periodically changing structures, amplitude response structures, and noise energy structures. This allows the mapping relationship between the main pulsating mode, high-frequency noise mode, and steady-state amplitude to be independently analyzed within a physically consistent framework. A stable and quantifiable periodic phase lag time is formed through variational mode decomposition and phase estimation. A peak hold coefficient is constructed through static amplitude mapping within the steady-state window, allowing the degree of peak clipping and compression to be characterized in a reproducible manner. The steady-state noise intensity is obtained by combining the envelope and short-time standard deviation, enabling accurate quantification of noise levels based on energy distribution. Overall, a complete monitoring accuracy analysis framework is established for monitoring edge lag time, response rise time, response fall time, periodic phase lag time, peak hold coefficient, and steady-state noise intensity. This provides a quantitative, stable, and traceable feature foundation for constructing dynamic response optimization thresholds and selecting compensation strategies.

[0039] Specifically, the process of generating trigger thresholds based on dynamic characteristics and performing phase compensation, peak recovery, and enhancement compensation is as follows: Based on the steady-state window criterion and the stability of the periodic structure, a stable window is obtained by constructing a strategy stability judgment window within a continuous pulsating window. Based on the statistical distribution of monitoring accuracy analysis results, including monitoring edge lag time, response rise time, response fall time, periodic phase lag time, peak hold coefficient, and steady-state noise intensity, a Bayesian quantile estimation algorithm is used to generate dynamic response optimization thresholds for triggering the optimization strategy. The training samples for Bayesian quantile estimation are derived from online sliding window samples within the stable window, supplemented by samples from the corresponding operating condition segment in the historical statistical database as priors, ensuring that threshold generation has both real-time performance and statistical stability. The quantile values ​​are taken in the range of 0.8–0.9 to enhance sensitivity to abnormal fluctuations. The threshold refresh cycle is fixed to several pulsating window lengths to ensure consistency between the update rhythm and the periodic structure. The dynamic response optimization thresholds include: lag time optimization threshold, coefficient optimization threshold, and response time optimization threshold. By reconstructing the probability density structure of the monitoring accuracy analysis results, the thresholds can reflect the dynamic drift of the distribution pattern under different operating conditions, improving the stability and robustness of threshold determination. Real-time comparison is performed between the periodic phase lag time, peak hold coefficient, response rise time, response fall time, and dynamic response optimization threshold. The execution order of the three triggering conditions follows a parallel evaluation and priority mutual exclusion principle of "phase compensation - peak recovery - enhanced compensation": phase compensation triggering has the highest priority, entering the phase compensation path first when the phase-related indicators meet the triggering conditions; the peak recovery strategy is triggered when phase compensation is not triggered or when phase compensation has converged; the enhanced compensation strategy is only executed as a supplementary path when neither of the first two strategies is triggered, thereby avoiding compensation conflicts caused by multiple strategies triggering simultaneously and ensuring that the compensation link has a clear physical dependency and temporal consistency. When the periodic phase lag time is greater than the lag time optimization threshold and the stable window length condition is met, the phase compensation strategy is triggered; when the periodic phase lag time is less than or equal to the lag time optimization threshold, the existing phase compensation state remains unchanged and monitoring continues. Based on the stability of the periodic structure within the continuous pulsation window, the phase drift trend is updated, so that the phase compensation strategy can maintain a consistent dynamic adaptation capability in subsequent cycles.

[0040] When the peak retention coefficient is less than the coefficient optimization threshold and the stable window length condition is met, the peak recovery strategy is triggered; when the peak retention coefficient is greater than or equal to the coefficient optimization threshold, the original peak compensation strategy remains unchanged and monitoring continues, and the peak confidence is corrected in conjunction with the steady-state noise intensity, so that the peak retention coefficient can maintain a quantitative reference with consistent physical meaning under different noise levels.

[0041] When both the response rise time and response fall time are greater than the response time optimization threshold, the enhanced compensation strategy is triggered; when both the response rise time and response fall time are less than or equal to the response time optimization threshold, the current compensation strategy remains unchanged. By using tiered triggering, frequent strategy switching is avoided, improving stability under fluctuating conditions. This ensures that the compensation logic remains continuous under rapid disturbances or short-cycle fluctuations, preventing overcompensation from causing reverse drift in the system under high-fluctuation conditions.

[0042] like Figure 2 The diagram shows the strategy triggering logic flowchart. Starting with the monitoring data input, after the generation of the dynamic response optimization threshold, different compensation strategy triggering paths are entered step by step according to the judgment order of hysteresis characteristics, peak characteristics, and response characteristics. The flowchart illustrates the triggering relationship of phase compensation, peak recovery, and enhancement compensation strategies in the same logical link, as well as the branching structure between compensation triggering and hold states. Furthermore, the diagram visualizes the judgment order of compensation strategies, decision nodes, and the closed-loop process of continuous monitoring to aid in understanding the execution rhythm and strategy connection of the dynamic compensation mechanism throughout the entire monitoring update cycle.

[0043] In this implementation plan, by analyzing the statistical structure of monitoring edge lag time, response rise time, response fall time, periodic phase lag time, peak hold coefficient, and steady-state noise intensity, a dynamic response optimization threshold system capable of adapting to the fluctuation characteristics of different operating conditions is formed, enabling the triggering of compensation strategies to possess data-driven stability and discriminative capabilities. Based on the linkage judgment mechanism of lag characteristics, peak characteristics, and response characteristics, the system achieves hierarchical triggering of phase compensation, peak recovery, and enhancement compensation during continuous monitoring, allowing compensation decisions to automatically adjust as monitoring accuracy changes while maintaining logical continuity. The overall process establishes a multi-feature collaborative dynamic compensation closed loop within a unified framework, enabling compensation strategies to closely follow changes in operating conditions and maintain adaptive consistency, achieving orderly connection between strategy selection, triggering, and hold during the monitoring update cycle.

[0044] Specifically, the process of correcting dynamic distortion, recording residual sequences, updating compensation parameters, and outputting optimized flow data for monitoring accuracy is as follows: The phase compensation strategy adopts an unscented Kalman filter state observation framework: instantaneous flow data is used as the observation, and control input reference data, heating drive current data, and heating drive voltage data are used as exogenous driving quantities. A dynamic equation including hysteresis states is established: the actual flow change is described as the main flow state, and the thermal inertia error formed by the accumulation of heating power is described as the hysteresis energy state. The dynamic response process of the thermal flow sensor is characterized by the joint update of the main flow state and the hysteresis energy state. The dynamic equation adopts a two-state structure: the main flow state is updated at each sampling time point according to the changing trend of the control input reference data, the main flow state at the previous time, and the hysteresis energy state at the previous time. The new, hysteresis energy state is updated based on the heating power derived from the hysteresis energy state of the previous moment, along with the heating drive current and heating drive voltage data. This update describes the response delay caused by thermal diffusion inertia. A hysteresis-corrected flow estimate is output at each sampling timestamp. The hysteresis-corrected flow estimate is used as the predictor, and the instantaneous flow data is used as the observation. An unscented Kalman filter update process is constructed. The state vector, observation vector, process noise covariance, observation noise covariance, and Sigma point generation coefficients α, β, and κ of the unscented Kalman filter are estimated from the historical steady-state noise level and transient response amplitude to ensure the numerical stability of state propagation and observation update. The unscented Kalman filter residual sequence and update compensation parameters are recorded to compensate for phase hysteresis without simply shifting the sequence, thereby reducing the risk of causal conflicts.

[0045] The peak recovery strategy employs amplitude-constrained regularized deconvolution reconstruction: a pulsation transfer function for deconvolution is constructed using a reference peak amplitude and a peak preservation coefficient. The equivalent amplitude response of the pulsation transfer function is taken as the peak preservation coefficient. The pulsation transfer function is used to recover the peak information in the instantaneous flow data: by using the equivalent amplitude response as the amplitude compensation basis for the deconvolution reconstruction process, total variation regularized deconvolution recovery is performed on the instantaneous flow data. The objective function of total variation deconvolution consists of a data fidelity term and a TV regularization term. The regularization strength is adaptively limited according to the steady-state noise intensity. The boundary processing adopts a symmetrical extension method to avoid convolution boundary artifacts. The steady-state noise intensity is used as the upper limit constraint of the amplitude, thereby compensating for peak attenuation and restoring the true peak structure while suppressing noise amplification.

[0046] like Figure 3 The diagram illustrates peak recovery using regularized deconvolution, showcasing the differences between the real signal, the degraded signal after convolution and noise interference, and the signal recovered using a deconvolution strategy with amplitude constraints. The real signal exhibits a complete peak structure and high-frequency details; the degraded signal, under the influence of thermal inertia and noise, shows significant phase lag, peak attenuation, and edge blunting dynamic distortion; the recovered signal, under amplitude constraints, suppresses noise amplification through deconvolution while compensating for the attenuated peaks, making the overall waveform closer to the real signal's shape, with significantly improved peak prominence and edge variation tracking. This diagram visually demonstrates the peak recovery strategy's effect on correcting dynamic response distortion.

[0047] The enhanced compensation strategy employs a feedforward-feedback coupling: the feedforward term is obtained by online updating of the control input reference data via recursive least squares with a forgetting factor. The forgetting factor value is determined based on the autocorrelation decay characteristics of the historical steady-state rate of change, ensuring that the memory length is consistent with the physical change cycle of the system. The feedback term is obtained by closed-loop correction of the unscented Kalman filter residual. The compensation parameters are corrected using sensor temperature segmentation. The temperature segment boundaries are obtained by clustering historical sensor temperature data using a Gaussian mixture model. The number of cluster segments is determined by minimizing the Bayesian information criterion. Inter-segment switching uses exponential smooth transition to avoid compensation jumps caused by temperature segment jitter, thereby reducing compensation mismatch caused by temperature drift and improving long-term stability. The feedforward and feedback terms are used together to update the phase compensation parameters and peak compensation parameters in the compensation parameter record and output monitoring accuracy optimized flow data. The output monitoring accuracy optimized flow data, optimized trigger flags, and compensation parameter records provide a unified basis for reliable flow output and reliable output and control protection of the closed-loop control protection module, and support full-link archiving and traceability.

[0048] In this implementation scheme, the coordinated execution of phase compensation, peak recovery, and enhancement compensation strategies achieves systematic repair of dynamic response distortion under the influence of hysteresis, peak clipping, and temperature drift. This enables the compensation process to adaptively adjust the compensation intensity based on the real-time data structure and maintain long-term stability. Phase compensation relies on unscented Kalman filtering to jointly estimate the main flow state and hysteresis energy state, dynamically canceling hysteresis errors within the prediction domain. Peak recovery relies on a regularized deconvolution mechanism with amplitude constraints, allowing the weakened peak value to be controllably compensated under steady-state noise intensity limitations. Enhancement compensation continuously corrects compensation parameters within a temperature-segmented framework through a feedforward-feedback coupling mechanism, ensuring that the optimized monitoring accuracy and flow data maintain consistency and dynamic reliability across the entire operating range. This enables continuous transmission of monitoring results to reliable output and control / protection logic.

[0049] Specifically, the process of constructing a confidence interval based on the residual sequence and synthesizing a control confidence label with the dynamic consistency score is as follows: Input the unscented Kalman filter residual sequence from the compensation parameter record to construct a flow residual distribution model. Use the unscented Kalman filter residual sequence as training samples, fit the residual conditional quantile function using ordinal quantile regression, and output the upper and lower bounds of the residual confidence. Superimpose these onto the flow data optimized for monitoring accuracy to obtain the upper and lower bounds of the flow confidence interval. Output the upper and lower bounds of the flow confidence interval. Use the current sampled value of the flow data optimized for monitoring accuracy as the flow value to be determined. When the flow value to be determined falls between the lower and upper bounds of the flow confidence interval, it is marked as a risk-free interval; when the flow value to be determined exceeds the upper bound of the flow confidence interval, it is marked as a level-one risk interval; when the flow value to be determined exceeds the upper bound of the flow confidence interval, it is marked as a level-one risk interval. When the value is below the lower bound of the flow confidence interval, it is marked as a secondary risk interval. This is used to expand the compensated flow value into a feedback quantity with an uncertain boundary, solving the problem that the controller cannot determine the confidence level when there is still residual lag and peak-shaving error in the rapid fluctuation segment. At the same time, a dynamic consistency score is constructed based on the periodic phase lag time and the peak value retention coefficient. The dynamic consistency score is calculated by aligning the main pulsation mode of the monitoring accuracy-optimized flow data and the control input reference data through dynamic time warping. This is used to quantify whether the dynamic correspondence between the feedback and the operating condition input is valid. The flow confidence interval, dynamic consistency score and monitoring accuracy analysis results are combined to form a control confidence label, which is output synchronously with the monitoring accuracy-optimized flow data. This allows the controller to obtain directly usable data quality constraints, avoiding treating the peak value after peak shaving as the real peak value and the lag edge as the real edge.

[0050] In this implementation scheme, by constructing a flow residual distribution model and a dynamic consistency score, the reliability of the flow data is enhanced to improve the accuracy of monitoring. This ensures that the compensated flow sequence not only has a single-value output form but also quantifiable uncertainty boundaries and dynamic consistency constraints, enabling the control system to clearly distinguish between real changes and residual distortions. The residual distribution model establishes the flow reliability interval through ordinal quantile regression, allowing the lag and peak-shaving residuals to be expressed in a boundary-based manner under rapid fluctuation conditions. The dynamic consistency score quantifies the correspondence between the feedback and control input response by aligning with the main pulsating mode, enabling the compensated signal to obtain reliability verification at the structural level. These two factors, together with the monitoring accuracy analysis results, synthesize a control reliability label, giving the output a quality label that can distinguish the risk level. This allows the control system to directly identify and safely use the data reliability, thereby ensuring that subsequent control strategies can obtain a stable and reliable feedback basis under various disturbance scenarios.

[0051] Specifically, the process of implementing safety constraints and predictive filtering based on control confidence markers and completing the entire process archiving is as follows: The control confidence markers are mapped to a closed-loop control safety filtering strategy, and the execution quantities output by the controller are constrained and corrected to form an engineering-feasible anti-oscillation protection link. When the control confidence markers enter the secondary risk range, the execution quantity change rate is calculated from the first-order difference between adjacent sampling timestamps of the execution quantity output sequence. A monotonic safety boundary construction method based on the execution quantity change rate is adopted. The safety boundary of the execution quantity change rate is constructed using the quantile statistics of the current execution quantity change rate and the historical steady-state change rate. A control barrier function is constructed using the current execution quantity change rate and the safety boundary, and a control barrier function greater than zero is used as... The rate-of-change safety constraint ensures that the rate of change of the output execution quantity is updated within a safe boundary, thereby avoiding reverse following and over-adjustment caused by residual phase lag. The controller output execution quantity maintains the principle of minimum modification while satisfying the safety constraint, which is used to suppress reverse following and over-adjustment caused by residual phase lag. When the control confidence flag enters the first-level risk range, the model prediction safety filter is activated. The model prediction safety filter adopts a finite prediction domain model prediction control structure, and uses deviation penalty matrix and change penalty matrix to apply soft constraint penalties to flow deviation and execution quantity change. The prediction domain length is within the range of a to b sampling periods according to the control cycle and the dynamic characteristics of the operating conditions. The range is determined based on the dynamic response characteristics of the gas path system under typical operating conditions: First First, autocorrelation analysis is performed on the historical steady-state rate of change sequence, monitoring edge lag time, response rise time, and response fall time to identify the effective response span required for the dynamic correlation to decay to below the significance threshold. Then, this effective span is converted into the corresponding number of sampling periods, ensuring the prediction domain covers a complete dynamic response process while retaining necessary look-ahead margin. The effective response span ranges from 10 to 20 sampling periods, with the sampling period consistent with the monitoring data sampling interval. The model prediction safety filter uses the flow confidence interval as a hard constraint and the dynamic consistency score as a soft constraint to solve for the safety execution quantity and output a protection trigger flag. This is used to maintain closed-loop control during periods of significant peak-shaving distortion and significant lag, avoiding measurement errors. The error-adjustment-larger fluctuation amplification link; when the control confidence mark enters the risk-free zone, no special operation is required; at the same time, a real-time monitoring database is created, and event-level archiving is performed. The archived content is indexed by the sampling timestamp and includes monitoring accuracy optimization traffic data, upper bound of the traffic confidence interval, lower bound of the traffic confidence interval, dynamic consistency score, control confidence mark, protection trigger mark, safety execution amount correction magnitude, and corresponding monitoring accuracy analysis results. The minimum retention period is jointly determined by the duration distribution of historical oscillation events and the convergence time statistics of the compensation strategy: the ninth percentile of the event duration is calculated for the historical event sequence, and this percentile is used as the lower bound of the minimum retention period, resulting in a retention range of not less than one million milliseconds;The backtracking window is determined by the autocorrelation decay length of the oscillation event interval and the minimum sample size required for parameter tuning. By calculating the autocorrelation function of the event trigger sequence and locating the hysteresis length at which it decays to below 0.1, and then combining this with the minimum sample size required for parameter tuning, a reasonable range of 50 to 100 event windows is obtained. This range is used to reproduce experiments and trace the triggering cause and effect of each oscillation suppression, supporting tuning and algorithm iteration, and forming a complete chain of research and development evidence.

[0052] In this embodiment, Table 1 is a comparison table of monitoring accuracy analysis results before and after compensation. It details the core monitoring accuracy indicators under different operating conditions, including periodic phase lag time (before and after compensation), peak value retention coefficient (before and after compensation), and monitoring edge lag time (before and after compensation). These indicators are used to quantify the optimization effect of the compensation strategy on dynamic response accuracy under different disturbance conditions. Specifically: In operating condition 1 (steady-state minor disturbance), the periodic phase lag time decreased from 18.4 to 7.2, the peak value retention coefficient increased from 0.78 to 0.93, and the monitoring edge lag time decreased from 22.1 to 11.3; in operating condition 2 (moderate disturbance, periodic fluctuation), the periodic phase lag time decreased from 33.7 to 14.9, the peak value retention coefficient increased from 0.66 to 0.89, and the monitoring edge lag time decreased from 41.3 to 19.4; in operating condition 3 (high-frequency disturbance), the periodic phase lag time decreased from 49.5 to 21.3, and the peak value retention coefficient increased from 0.78 to 0.93. The retention factor improved from 0.51 to 0.83, and the monitoring edge lag time decreased from 62.7 to 28.6. In condition 4 (abrupt edge), the periodic phase lag time decreased from 57.2 to 25.7, the peak retention factor improved from 0.46 to 0.81, and the monitoring edge lag time decreased from 79.3 to 34.1. In condition 5 (strong noise + high fluctuation), the periodic phase lag time decreased from 66.4 to 29.5, the peak retention factor improved from 0.39 to 0.77, and the monitoring edge lag time decreased from 102.4 to 41.8. This clearly demonstrates the extent to which the compensation strategy improves dynamic response performance under different disturbance types.

[0053] Table 1 Comparison of Monitoring Accuracy Analysis Results Before and After Compensation

[0054] like Figure 4The bar chart shows the key dynamic response indicators before and after compensation. Combined with Table 1, it can be seen that the periodic phase lag time significantly decreased across all operating conditions, the peak value retention coefficient significantly improved, and the monitoring edge lag time converged to a shorter delay interval. For example, in operating condition 5, after compensation, the periodic phase lag time decreased from 66.4 to 29.5, the peak value retention coefficient increased from 0.39 to 0.77, and the monitoring edge lag time decreased from 102.4 to 41.8, indicating that the compensation mechanism remains effective even under strong noise and high fluctuation environments. Operating conditions 1 to 4 also show a continuous trend of shortened lag time and enhanced amplitude retention capability. The visualized bar chart of dynamic response indicators intuitively presents the optimization effect of the compensation strategy under different disturbance scenarios and can serve as an important basis for monitoring accuracy analysis and control strategy adjustment.

[0055] In this implementation scheme, a safety filtering strategy driven by control credibility markers is used to enable the monitoring accuracy-optimized flow data to form a clear risk-level response logic in the closed-loop control link, achieving constrained correction of the execution quantity and engineering-level anti-oscillation protection. The secondary risk interval limits the magnitude of execution quantity changes through a rate-of-change safety boundary and a control barrier function, ensuring that the adjustment action maintains a controllable monotonic trend even with residual phase lag. The primary risk interval generates safe execution quantities through a model-predicted safety filter under the dual constraints of credibility intervals and dynamic consistency scores, ensuring the controller maintains dynamic stability during peak-shaving distortion and significant lag stages, avoiding error amplification. The risk-free interval maintains the original process, ensuring that normal adjustment efficiency is not affected. The entire process synchronously archives monitoring accuracy-optimized flow data, flow credibility intervals, control credibility markers, and protection triggering processes, providing reproducible and interpretable evidence for each risk suppression action, constructing a complete closed-loop evidence system for both field and R&D purposes.

[0056] A second aspect of the present invention provides a real-time monitoring controller based on a thermal flow sensor, comprising: a memory, a processor, and a computer program stored in the memory and executable on the processor.

[0057] In this implementation scheme, a real-time monitoring controller based on a thermal flow sensor is employed to achieve full-link dynamic sensing and adaptive control of flow changes. This enables monitoring data to be collected, processed, and fed back to the control unit within milliseconds, significantly improving the real-time response and control accuracy under changing operating conditions. The system relies on the coordinated operation of memory, processor, and executable program to achieve continuous processing of monitoring data, online updates of compensation strategies, and safety constraints on control outputs. This effectively suppresses sensor hysteresis errors, peak clipping distortion, and noise disturbances, thereby ensuring the stability and predictability of closed-loop control even under rapidly fluctuating operating conditions. Through these mechanisms, the reliability of flow measurement and control execution can be maintained in complex dynamic scenarios, significantly improving overall monitoring and control performance.

[0058] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0059] The preferred embodiments of the present invention disclosed above are merely illustrative of the invention. These preferred embodiments do not exhaustively describe all details, nor do they limit the invention to the specific implementations described. Clearly, many modifications and variations can be made based on the content of this specification. This specification selects and specifically describes these embodiments to better explain the principles and practical applications of the invention, thereby enabling those skilled in the art to better understand and utilize the invention. The invention is limited only by the claims and their full scope and equivalents.

Claims

1. A real-time monitoring and control system based on a thermal flow sensor, characterized in that, Includes the following steps: S1, Collect sensor monitoring data and obtain historical statistical data; preprocess the sensor monitoring data and historical statistical data; S2, construct an edge window based on sensor monitoring data, and extract dynamic features of hysteresis and response time; Modal decomposition, phase estimation, peak value calculation, and steady-state noise analysis are performed. S3 generates trigger thresholds based on dynamic characteristics and performs phase compensation, peak recovery, and enhancement compensation determination; corrects dynamic distortion, records residual sequences, updates compensation parameters, and outputs flow data optimized for monitoring accuracy; S4, construct a confidence interval based on the residual sequence and synthesize a control confidence label with the dynamic consistency score; Security constraints and predictive filtering are performed based on the control confidence level, and the entire process is archived.

2. The real-time monitoring and control system based on a thermal flow sensor according to claim 1, characterized in that: The specific process for acquiring historical statistical data from sensor monitoring data is as follows: Collect sensor monitoring data, which includes: instantaneous flow data, execution sequence, sensor temperature data, heating drive data, pipeline pressure data, valve opening command data, air pump drive signal data, and sampling timestamp data; Acquire historical statistical data and create a historical statistical database. The historical statistical data includes: historical instantaneous flow rate data, historical sensor temperature data, historical valve opening command data, historical air pump drive signal data, and historical steady-state change rate. Among them, the heating drive data refers to the sampled values ​​of heating drive current and heating drive voltage; the sampling time interval is obtained by subtracting the timestamps.

3. The real-time monitoring and control system based on a thermal flow sensor according to claim 1, characterized in that: The specific process for preprocessing sensor monitoring data and historical statistical data is as follows: The instantaneous flow rate data, pipeline pressure data, and sensor temperature data are de-glitched using a sliding window midpoint algorithm to remove isolated spikes and communication jitter anomalies; the data sequences in the sensor monitoring data and historical statistics are time-aligned based on the sampling timestamp data; and the sensor monitoring data and historical statistics are standardized and normalized using distribution standardization and linear normalization algorithms.

4. The real-time monitoring and control system based on a thermal flow sensor according to claim 1, characterized in that: The specific process of constructing an edge window based on sensor monitoring data and extracting dynamic features of hysteresis and response time is as follows: When a valve actuator is present, the valve opening command data is used as the control input reference; when a pneumatic pump actuator is present, the pneumatic pump drive signal data is used as the control input reference; Bayesian online change point detection is performed on the control input reference sequence and pipeline pressure data, and the edge occurrence timestamp is extracted. Edge detection is performed on instantaneous flow data with segmented monotonicity constraints to extract flow response edge timestamps; The difference between the two types of edge timestamps is calculated under the same time base of the sampling timestamp data to obtain the monitoring edge lag time. The edge window is jointly determined by the edge detection results of the control input reference data and pipeline pressure data under the same time base of the sampling timestamp data. Within the edge window, the percentage response is estimated based on quantile regression. The response rise time and response fall time are obtained by locating the percentage response point of the instantaneous flow data within the same window. The response rise time and response fall time are then calculated.

5. The real-time monitoring and control system based on a thermal flow sensor according to claim 1, characterized in that: The specific process of performing mode decomposition, phase estimation, peak value calculation, and steady-state noise analysis is as follows: Variational mode decomposition is performed on the instantaneous flow rate data within the pulsation window to separate the main pulsating mode and the high-frequency noise mode. The pulsation window is jointly determined by the instantaneous flow rate data and the control input reference data under the same time base of the sampling timestamp data through periodic feature detection. For the main pulsating mode, a generalized cross-correlation phase transform is used to estimate the periodic phase lag time. A Hilbert transform is used to obtain the instantaneous phase sequence. The phase difference between the control input reference phase and the main pulsating phase of the flow rate is calculated, and the periodic phase lag time is output. Peak clipping distortion is calculated by comparing the reference peak amplitude with the actual peak amplitude: based on the instantaneous flow rate data and pipeline pressure data... The steady-state window is determined by the short-time variance, first-order difference amplitude, and change point detection results of the control input reference data. Using historical instantaneous flow rate data, historical valve opening command data, and historical air pump drive signal data, a static amplitude mapping coefficient from the control input reference to the instantaneous flow rate is established within the steady-state window using recursive least squares with a forgetting factor. Within the pulsating window, the peak value change of the control input reference data is determined using a short-time peak value detection method. The reference peak value amplitude is calculated by multiplying the static amplitude mapping coefficient by the peak value change. The peak value retention coefficient is calculated by the ratio of the actual peak value amplitude of the instantaneous flow rate data to the reference peak value amplitude. Meanwhile, for the high-frequency noise modes within the steady-state window, the steady-state noise intensity is obtained by combining the sliding window standard deviation with the envelope amplitude integral. Output monitoring accuracy analysis results, which include: monitoring edge lag time, response rise time, response fall time, periodic phase lag time, peak hold factor, and steady-state noise intensity.

6. The real-time monitoring and control system based on a thermal flow sensor according to claim 1, characterized in that: The specific process of generating trigger thresholds based on dynamic feature statistics and performing phase compensation, peak recovery, and enhancement compensation determination is as follows: Based on the steady-state window criterion and periodic structure stability, a stable window is obtained by constructing a strategy stability judgment window within a continuous pulsating window. Based on the statistical distribution of monitoring accuracy analysis results, a Bayesian quantile estimation algorithm is used to generate dynamic response optimization thresholds, including: lag time optimization threshold, coefficient optimization threshold, and response time optimization threshold. Real-time comparison is performed between the periodic phase lag time, peak hold coefficient, response rise time, response fall time, and dynamic response optimization thresholds. When the periodic phase lag time is greater than the lag time optimization threshold and the stable window length condition is met, the phase compensation strategy is triggered; when the periodic phase lag time is less than or equal to the lag time optimization threshold, the existing phase compensation state remains unchanged and monitoring continues. When the peak retention coefficient is less than the coefficient optimization threshold and the stable window length condition is met, the peak recovery strategy is triggered; when the peak retention coefficient is greater than or equal to the coefficient optimization threshold, the original peak compensation strategy remains unchanged and monitoring continues. When both the response rise time and response fall time are greater than the response time optimization threshold, an enhanced compensation strategy is triggered; when both the response rise time and response fall time are less than or equal to the response time optimization threshold, the current compensation strategy remains unchanged.

7. The real-time monitoring and control system based on a thermal flow sensor according to claim 1, characterized in that: The specific process of correcting dynamic distortion, recording the residual sequence and updating compensation parameters, and outputting optimized flow data for monitoring accuracy is as follows: The phase compensation strategy employs an unscented Kalman filter state observation framework: instantaneous flow rate data is used as the observation, and control input reference data, heating drive current data, and heating drive voltage data are used as exogenous driving quantities. A dynamic equation including hysteresis states is established, and the dynamic equation adopts a two-state structure: the main flow rate state is updated at each sampling time point based on the changing trend of the control input reference data, the main flow rate state at the previous time point, and the hysteresis energy state at the previous time point; the hysteresis energy state is updated based on the heating power composed of the hysteresis energy state at the previous time point, the heating drive current data, and the heating drive voltage data; a hysteresis-corrected flow rate estimate is output at each sampling time point. The hysteresis-corrected flow rate estimate is used as the predictor, and the instantaneous flow rate data is used as the observation to construct the unscented Kalman filter update process, and the unscented Kalman filter residual sequence and update compensation parameters are recorded. The peak recovery strategy employs amplitude-constrained regularized deconvolution reconstruction: a pulsation transfer function for deconvolution is constructed using the reference peak amplitude and peak preservation coefficient to recover the peak information in the instantaneous flow data. The equivalent amplitude response of the pulsation transfer function is taken as the peak preservation coefficient. By using the equivalent amplitude response as the basis for amplitude compensation in the deconvolution reconstruction process, total variation regularized deconvolution recovery is performed on the instantaneous flow data, with steady-state noise intensity as the upper limit constraint of amplitude. The enhanced compensation strategy adopts a feedforward-feedback coupling: the feedforward term is obtained by online updating of the control input reference data through recursive least squares with a forgetting factor, and the feedback term is obtained by closed-loop correction of the unscented Kalman filter residual; The compensation parameters are corrected by segmenting the sensor temperature. The temperature segment boundaries are obtained by clustering historical sensor temperature data using a Gaussian mixture model. The output monitoring accuracy is optimized by recording flow data, trigger markers, and compensation parameters.

8. The real-time monitoring and control system based on a thermal flow sensor according to claim 1, characterized in that: The specific process of constructing a confidence interval based on the residual sequence and synthesizing a control confidence label with the dynamic consistency score is as follows: The unscented Kalman filter residual sequence in the input compensation parameter record is used to construct the flow residual distribution model. The unscented Kalman filter residual sequence is used as the training sample. The conditional quantile function of the residual is fitted by the ordinal quantile regression. The upper and lower bounds of the residual are output and superimposed on the flow data for monitoring accuracy optimization to obtain the upper and lower bounds of the flow confidence interval. The current sampled value of the traffic data is optimized based on the monitoring accuracy and used as the traffic value to be judged. When the traffic value to be judged falls between the lower bound and the upper bound of the traffic confidence interval, it is marked as a risk-free interval. When the traffic value to be determined exceeds the upper limit of the traffic confidence interval, it is marked as a level 1 risk interval; When the traffic value to be determined is lower than the lower bound of the traffic confidence interval, it is marked as a level 2 risk interval; Dynamic consistency score is obtained by aligning the main pulsation mode of the monitoring accuracy optimization flow data and the control input reference data through dynamic time warping. The dynamic consistency score is constructed based on the periodic phase lag time and peak hold coefficient. The flow confidence interval, dynamic consistency score and monitoring accuracy analysis results are combined to form a control confidence label and output synchronously.

9. The real-time monitoring and control system based on a thermal flow sensor according to claim 1, characterized in that: The specific process of performing security constraints and predictive filtering based on control confidence tags and completing the entire process archiving is as follows: The control confidence mark is mapped to a closed-loop control safety filtering strategy to perform constraint-based correction on the execution quantity output by the controller: When the control confidence mark enters the secondary risk interval, the execution quantity change rate is calculated from the execution quantity sequence output by the controller through the first-order difference of adjacent sampling timestamps. A monotonic safety boundary construction method based on the execution quantity change rate is adopted. The safety boundary of the execution quantity change rate is constructed by statistically analyzing the quantiles of the current execution quantity change rate and the historical steady-state change rate. A control barrier function is constructed by the current execution quantity change rate and the safety boundary. The control barrier function being greater than zero serves as a safety constraint on the change rate, ensuring that the output execution quantity change rate is updated within the safety boundary. When the control confidence mark enters the primary risk interval, the model prediction safety filter is activated. The model prediction safety filter adopts a finite prediction domain model prediction control structure. The deviation penalty matrix and the change penalty matrix apply soft constraint penalties to the flow deviation and execution quantity change. The prediction domain length is within the range of a to b sampling cycles according to the control cycle and the dynamic characteristics of the operating condition. The sampling cycle is consistent with the sampling interval of the monitoring data. The model prediction safety filter uses the flow confidence interval as a hard constraint and the dynamic consistency score as a soft constraint to solve for the safe execution quantity and output the protection trigger mark. When the confidence level flag enters the risk-free zone, no special action is required; Create a real-time monitoring database and perform event-level archiving. The archived content is indexed by the sampling timestamp and includes: monitoring accuracy optimization traffic data, upper bound of the traffic confidence interval, lower bound of the traffic confidence interval, dynamic consistency score, control confidence flag, protection trigger flag, security execution amount correction magnitude, and corresponding monitoring accuracy analysis results.

10. A real-time monitoring controller based on a thermal flow sensor, characterized in that... The system includes: a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement a monitoring and control system for the controller as described in any one of claims 1-9.

Images