Dew-point meter measurement data processing method and device

Abstract

The invention discloses a dew-point meter measurement data processing method and equipment, relates to the technical field of industrial automatic measurement and control and fault diagnosis, and aims to construct a dynamic sensing window with a variable form to preprocess dew-point meter time sequence data, calculate fault probability entropy in real time and detect instantaneous abnormality. And a window form is decided in combination with the fault-window form mapping library, and adaptive switching is carried out to obtain an anomaly recognition result. And selecting a corresponding algorithm from the correction strategy library to correct the abnormity. According to the method, a dynamic sensing window with a variable form is constructed, and adaptive switching among panoramic, sniping and multi-focus forms is realized by taking the fault probability entropy calculated in real time as a drive. By means of a built-in fault-window form mapping library, abnormal symptoms and fault physical mechanisms are associated, multiple concurrent faults are separated, root causes are deduced, and decision space explosion is avoided. And on the basis of a diagnosis result, a correction strategy library targeted algorithm is called to correct abnormity, and an identification-diagnosis-correction intelligent closed loop is formed.

Description

Technical Field

[0001] This invention relates to the field of industrial automation measurement and control and fault diagnosis technology, specifically to a dew point meter measurement data processing method and equipment. Background Technology

[0002] With the rapid advancement of automation and intelligence in industrial processes, dew point meters, as key gas humidity measurement devices, are playing an increasingly important role in high-precision industrial fields such as natural gas processing, compressed air drying, and semiconductor manufacturing. However, in actual operation, dew point meter measurement systems, as typical dynamic systems, are often subject to coupled interference from various complex factors, including drastic fluctuations in the pressure and temperature of process gases, gradual accumulation of contaminants on the mirror surface, slow performance drift of core sensor components, and transient noise interference from the electrical environment. These faults, with their different physical natures, have varying time constants and manifestations, and are often interconnected and mutually influential, forming complex fault chains that pose a severe challenge to the reliability and accuracy of measurement data.

[0003] Existing dew point meter data anomaly handling solutions mostly employ fixed-time analysis windows or single diagnostic strategies. These methods lack the ability to perceive and respond to the inherent physical characteristics of faults, resulting in a rigid analytical perspective. When faced with complex anomalies caused by multiple concurrent faults and intertwined over time, existing technologies cannot effectively separate and attribute different types of anomalies, leading to the inherent problem of decision space explosion in diagnostic systems—that is, the number of possible fault combinations grows exponentially with the increase in monitoring dimensions. This makes it difficult for the system to quickly locate the most probable root cause combination from a massive number of fault possibilities within limited real-time computing resources. Ultimately, this results in either providing only general anomaly alerts without guiding precise maintenance, or being too computationally complex to meet real-time requirements, often missing the optimal maintenance window. Furthermore, data correction based on erroneous or ambiguous diagnostic results inevitably lacks specificity, significantly reducing the effectiveness of corrections. Summary of the Invention

[0004] The purpose of this invention is to provide a method and device for processing dew point meter measurement data to address the shortcomings in the prior art.

[0005] To achieve the above objectives, the present invention provides the following technical solution: a dew point meter measurement data processing method, comprising: A dynamic sensing window is constructed to preprocess the time-series measurement data of the dew point meter, and preprocessed data is obtained. Based on the preprocessed data, the fault probability entropy of the data stream is then calculated in parallel and instantaneous abnormal events are detected, obtaining the fault probability entropy calculation results and instantaneous abnormal event detection results, respectively. The fault probability entropy calculation result and the instantaneous abnormal event detection result are input together into the fault-window shape mapping library constructed based on the fault physical time constant to obtain the window shape decision result; Based on the window shape decision result, the dynamic perception window is controlled to adaptively switch between panoramic mode, sniper mode and multi-focus mode to obtain the abnormal recognition result after the window switching. Based on the anomaly identification results after the window switching, a corresponding correction algorithm is selected from the correction strategy library to correct the abnormal data in the preprocessed data, and the corrected data is obtained. Verify the effect of the correction algorithm on the corrected data, obtain the verification result, and simultaneously feed the verification result back to the fault-window shape mapping library and the correction strategy library.

[0006] In a preferred embodiment, the fault probability entropy is a comprehensive index obtained by weighting the data uncertainty represented by the Shannon entropy of the data within the dynamic perception window and the model uncertainty represented by the prediction confidence of the diagnostic model for the current data input. The formula for calculating the fault probability entropy is: Hfault=α·Hdata+β·Hmodel; Wherein, Hdata is the Shannon entropy calculated based on the data within the dynamic perception window, Hmodel is the prediction confidence of the diagnostic model, and α and β are adaptive weighting coefficients based on the device health status. The dynamic adjustment of the weighting coefficients α and β includes: When the sensor health is below a preset threshold, increase the model confidence weight β and decrease the data Shannon entropy weight α. When the intensity of environmental interference is higher than the preset threshold, increase the data Shannon entropy weight α and decrease the model confidence weight β.

[0007] In a preferred embodiment, the form of the dynamic sensing window includes at least panoramic mode, sniper mode, and multi-focus mode; The shape switching of the dynamically perceived window follows these rules: When the fault probability entropy continues to be higher than the first threshold, the window switches to the panoramic mode and the width is expanded to the maximum value. When a transient data spike is detected and its amplitude exceeds the dynamic threshold, the window switches to the sniper mode, shrinks to the minimum width, and anchors around the spike event. When the system detects multiple suspected abnormal segments that are not discontinuous in time, the window switches to the multi-focus mode, forming multiple non-discontinuous sub-windows that surround each abnormal segment.

[0008] In a preferred embodiment, the implementation of the multi-focus mode includes the following steps: Identify the start and end points of multiple anomalous segments in time-series data; Generate independent sub-windows centered on each abnormal fragment; Feature extraction and anomaly analysis are performed in parallel on the data within each sub-window. The analysis results are then merged and input into the diagnostic engine.

[0009] In a preferred embodiment, detecting transient anomalous events based on the preprocessed time-series data includes: Real-time monitoring of the first and second differences of data points; When the first-order difference exceeds the threshold dynamically calculated based on historical data, and the second-order difference indicates that the change is an instantaneous spike rather than the start of a trend, it is determined to be an instantaneous abnormal event.

[0010] In a preferred embodiment, the fault-window shape mapping library includes the following mapping relationships: For sensor-sensitive detection surface contamination and sensor detection circuit drift-related faults, map to panoramic mode; For electromagnetic interference and power ripple-related faults, map to sniper mode; For pressure over-release caused by intermittent ultra-high flow at the dew point and intermittent sensor connection stability failure, the system is mapped to a multi-focus mode.

[0011] In a preferred embodiment, the correction strategy library includes the following correction algorithms, and the parameters of the correction algorithms are dynamically adjusted according to the severity of the fault: To address the contamination fault on the sensor's sensitive detection surface, a historical data trend compensation and correction algorithm is employed. To address the drift fault in the sensor detection circuit, a self-calibration algorithm based on the built-in reference capacitor is adopted. To address the pressure over-release fault caused by intermittent ultra-high flow rate at the dew point, a dynamic correction algorithm for the pressure-dew point relationship is adopted. To address intermittent connection stability issues in sensors, a sliding window mid-range smoothing correction algorithm is employed. Specifically, a conservative compensation coefficient is used for minor faults; an aggressive compensation coefficient is used for severe faults; and a gradual compensation strategy is used for persistent faults.

[0012] In a preferred embodiment, the verification correction effect includes: Short-term verification: Check whether the corrected data has returned to the normal fluctuation range; Interim validation: Evaluate the stability of the device output after correction; Long-term validation: Analyze the impact of correction strategies on equipment lifespan; The verification results are quantified into an effectiveness score, which is used to optimize the algorithm weights in the correction strategy library.

[0013] The present invention also provides a dew point meter measurement data processing device, configured to implement the above-described method, comprising: The dynamic sensing window module is configured to construct a dynamic sensing window to preprocess the time-series measurement data of the dew point meter and obtain preprocessed data. The data stream calculation module is used to calculate the fault probability entropy of the data stream in parallel based on the preprocessed data and detect instantaneous abnormal events, and obtain the fault probability entropy calculation result and the instantaneous abnormal event detection result, respectively. The window shape decision module is used to input the fault probability entropy calculation result and the instantaneous abnormal event detection result into the fault-window shape mapping library constructed based on the fault physical time constant to obtain the window shape decision result. The window switching module is used to control the dynamic sensing window to adaptively switch between panoramic mode, sniper mode and multi-focus mode according to the window shape decision result, so as to obtain the abnormal recognition result after window switching. The correction module is used to select a corresponding correction algorithm from the correction strategy library based on the anomaly identification result after the window switching to correct the abnormal data in the preprocessed data and obtain the corrected data. The verification module is used to verify the correction effect of the correction algorithm on the corrected data, obtain the verification result, and simultaneously feed the verification result back to the fault-window shape mapping library and the correction strategy library.

[0014] The technical effects and advantages provided by the present invention in the above technical solution are as follows: This invention constructs a dynamically changing perception window. Using a comprehensive index—the fault probability entropy—calculated in real time, the window adaptively switches between panoramic, sniper, and multi-focus modes. Through a built-in fault-window mode mapping library, the system correlates detected abnormal symptoms with potential fault physical mechanisms, achieving preliminary separation and root cause inference in multi-fault concurrent scenarios. This decomposes the complex diagnostic problem into multiple manageable sub-problems, effectively avoiding decision space explosion. Furthermore, a correction strategy library based on the diagnostic results can invoke the most targeted algorithms to correct abnormal data, forming an intelligent closed loop from identification and diagnosis to correction. Attached Figure Description

[0015] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments recorded in this invention. For those skilled in the art, other drawings can be obtained based on these drawings.

[0016] Figure 1 This is a flowchart of the method of the present invention. Detailed Implementation

[0017] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, 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, 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.

[0018] Example 1, please refer to Figure 1 As shown in this embodiment, a dew point meter measurement data processing method includes: S1. Construct a dynamic sensing window to preprocess the time-series measurement data of the dew point meter and obtain preprocessed data; S2. Based on the preprocessed data, the fault probability entropy of the data stream is then calculated in parallel and instantaneous abnormal events are detected, and the fault probability entropy calculation results and instantaneous abnormal event detection results are obtained respectively. S3. Input the fault probability entropy calculation result and the instantaneous abnormal event detection result into the fault-window shape mapping library constructed based on the fault physical time constant to obtain the window shape decision result. S4. Based on the window shape decision result, control the dynamic perception window to adaptively switch between panoramic mode, sniper mode and multi-focus mode to obtain the abnormal recognition result after window switching. S5. Based on the anomaly identification result after the window switching, select the corresponding correction algorithm from the correction strategy library to correct the abnormal data in the preprocessed data and obtain the corrected data; S6. Verify the correction effect of the correction algorithm on the corrected data, obtain the verification result, and simultaneously feed the verification result back to the fault-window shape mapping library and the correction strategy library.

[0019] Existing dew point meter data anomaly handling solutions mostly employ fixed-time analysis windows or single diagnostic strategies. These methods lack the ability to perceive and respond to the inherent physical characteristics of faults, resulting in a rigid analytical perspective. When faced with complex anomalies caused by multiple concurrent faults and intertwined over time, existing technologies cannot effectively separate and attribute different types of anomalies, leading to the inherent problem of decision space explosion in diagnostic systems—that is, the number of possible fault combinations grows exponentially with the increase in monitoring dimensions. This makes it difficult for the system to quickly locate the most probable root cause combination from a massive number of fault possibilities within limited real-time computing resources. Ultimately, this results in either providing only general anomaly alerts without guiding precise maintenance, or being too computationally complex to meet real-time requirements, often missing the optimal maintenance window. Furthermore, data correction based on erroneous or ambiguous diagnostic results inevitably lacks specificity, significantly reducing the effectiveness of corrections.

[0020] As described in S1-S6 above, a dynamic perception window with variable shape is constructed. Using a comprehensive index—the fault probability entropy—calculated in real time, the window adaptively switches between panoramic, sniper, and multi-focus modes. Through a built-in fault-window shape mapping library, the system correlates detected abnormal symptoms with potential fault physical mechanisms, achieving preliminary separation and root cause inference in multi-fault concurrent scenarios. This decomposes the complex diagnostic problem into multiple manageable sub-problems, effectively avoiding decision space explosion. Furthermore, a correction strategy library based on the diagnostic results can invoke the most targeted algorithms to correct abnormal data, forming an intelligent closed loop from identification and diagnosis to correction.

[0021] In one embodiment, in step S3, the fault physical time constant refers to the characteristic duration range of various faults, which is determined by the physical failure mechanism corresponding to the fault, specifically including three types of quantization ranges: millisecond level (1-50ms), second level (1-60s), and hour / day level (5 hours-90 days). The physical time constant of the fault was determined through a combination of fault mechanism analysis and experimental statistical calibration. The specific steps are as follows: (1) Collect failure data of typical dew point meter malfunctions: including laboratory simulation failure data and industrial field failure records; (2) Extracting fault feature duration: For the time series data of each type of fault, the PELT mutation point detection algorithm is used to identify the fault start point and end point, and the duration t is calculated; (3) Statistical quantification interval: The duration t of the same type of fault is fitted with a normal distribution, and the 95% confidence interval is taken as the physical time constant interval of the fault type; (4) Dynamic calibration: When a new fault type is added or the field data shows that the original interval is not applicable, the interval boundary is adjusted based on the verification results of the effect evaluation feedback module; The correspondence between the quantization range of the fault physical time constant and the fault type is as follows: Millisecond level (1-50ms): corresponds to electromagnetic interference and power supply ripple type faults; Second-level (1-60s): corresponds to intermittent gas leakage and data transmission interruption faults; Hourly / Day Level (5 hours - 90 days): Corresponds to mirror contamination and sensor aging-related faults.

[0022] In one embodiment, the fault probability entropy is a comprehensive index obtained by weighting the data uncertainty represented by the Shannon entropy of the data within the dynamic perception window and the model uncertainty represented by the prediction confidence of the diagnostic model for the current data input. The formula for calculating the fault probability entropy is: Hfault=α×Hdata+β×Hmodel; Wherein, Hdata is the Shannon entropy calculated based on the data within the dynamic perception window, Hmodel is the prediction confidence of the diagnostic model, and α and β are adaptive weighting coefficients based on the device health status. Data uncertainty (Shannon entropy Hdata): Focuses on the inherent fluctuation characteristics of dew point meter measurement data, covering scenarios such as airflow fluctuations, 485 bus transmission errors, and sensor electrical signal noise. The calculation first preprocesses the time-series data within the dynamic sensing window, then establishes the data probability distribution through histogram statistics, and finally calculates it using the Shannon entropy formula. Hdata=- × ; Where P(xi) is the frequency of a certain data value within the window, and the higher the entropy value, the more drastic the data fluctuation and the stronger the uncertainty.

[0023] Model uncertainty (prediction confidence Hmodel): Addressing the limitations of diagnostic models in complex dew point meter scenarios, such as the inability to accurately distinguish between sensor mirror contamination drift and gradual changes in ambient temperature, and identification biases in novel intermittent pressure relief faults. Its value is calculated custom-based according to the model type to ensure a true reflection of the reliability of the model output—the variance of the output from multiple inferences via dropout in neural network models, the posterior distribution variance in Bayesian models, and the discrepancy measure of the prediction results from the base learners in ensemble learning models.

[0024] Weighted fusion mechanism: By linearly weighting, the two types of uncertainty are integrated into a unified index Hfault=α×Hdata+β×Hmodel, where α+β=1, to ensure the normalization characteristics of the index and facilitate subsequent threshold judgment and window shape triggering.

[0025] The dynamic adjustment of the weighting coefficients α and β includes: When the sensor health is below a preset threshold, increase the model confidence weight β and decrease the data Shannon entropy weight α. When the intensity of environmental interference is higher than the preset threshold, increase the data Shannon entropy weight α and decrease the model confidence weight β; The adjustment logic of the weighting coefficients is completely bound to the actual operating status of the dew point meter. Based on quantifiable equipment health indicators and environmental interference parameters, it achieves real-time matching of state and weights, rather than fixed coefficients or general adaptive rules. (1) Weight adjustment driven by sensor health; Sensor health is assessed using a multi-dimensional set of indicators, rather than relying on a single parameter. Health assessment indicators include sensor indication error, zero drift rate, response time T90, and repeatability standard deviation. The health score is calculated from 0 to 100 by weighting each indicator.

[0026] Threshold setting basis: The preset threshold is based on the sensor's factory calibration parameters and industry standards. For example, the health threshold of a cold mirror dew point meter is set to 60 points. A value lower than this indicates that the mirror surface is contaminated or that the thermoelectric cooling system has degraded performance.

[0027] Adjustment rules: When the health score is below 60, the reliability of sensor data decreases significantly (e.g., electrical signal noise from aging sensors may be misjudged as data fluctuation). In this case, β is increased from the default 0.5 to 0.7-0.8, and α is simultaneously reduced to 0.2-0.3, prioritizing the use of model confidence to avoid the impact of data distortion.

[0028] (2) Weight adjustment driven by environmental disturbance intensity The environmental interference intensity is designed to address the strong coupling characteristics of dew point meters, quantifying the degree of external interference in industrial scenarios: Interference intensity assessment indicators: covering ambient temperature fluctuation range, air pressure change rate, electromagnetic interference intensity, vibration amplitude, etc., and obtained through normalization to obtain an interference intensity score of 0-100.

[0029] Threshold setting basis: Referencing the dew point meter working environment standard, when the interference intensity is higher than 70 points, it indicates that environmental factors have seriously affected the stability of the data (such as electromagnetic interference in chemical workshops causing data spikes, and temperature fluctuations causing dew point reading coupling offset).

[0030] Adjustment rules: At this time, α is increased to 0.7-0.9 and β is decreased to 0.1-0.3. Prioritize capturing real fluctuation characteristics through data Shannon entropy to avoid misjudgment caused by insufficient environmental adaptation of the model (such as misjudging temperature coupling offset as normal data).

[0031] As mentioned above, the system can significantly improve the accuracy of anomaly identification and greatly reduce the false alarm rate, especially providing effective early warning for progressive faults. Under complex operating conditions such as sensor performance degradation and strong electromagnetic interference, the system exhibits excellent robustness, effectively maintaining diagnostic reliability and overcoming the industry challenge of performance degradation when traditional methods experience fluctuations in data quality. Based on a parallel processing architecture with a dynamic perception window, the system achieves significant optimization of computational efficiency, fully meeting the stringent real-time requirements of industrial sites. In terms of engineering applications, this solution can effectively reduce unplanned downtime, extend equipment service life, lay a solid technical foundation for building a predictive maintenance system, promote the intelligent transformation of operation and maintenance models, and has broad industry application prospects.

[0032] In one embodiment, the form of the dynamic sensing window includes at least panoramic mode, sniper mode, and multi-focus mode; The shape switching of the dynamically perceived window follows these rules: When the fault probability entropy continues to be higher than the first threshold, the window switches to the panoramic mode and the width is expanded to the maximum value. When a transient data spike is detected and its amplitude exceeds the dynamic threshold, the window switches to the sniper mode, shrinks to the minimum width, and anchors around the spike event. When the system detects multiple suspected abnormal segments that are not discontinuous in time, the window switches to the multi-focus mode, forming multiple non-discontinuous sub-windows that surround each abnormal segment. The panoramic mode specifically includes: Window width definition: The maximum value is dynamically set based on the sampling frequency of the dew point meter. When the sampling frequency is 1Hz, the window width in panoramic mode is 30 minutes to 2 hours (i.e., 1800-7200 sampling points); when the sampling frequency is 10Hz, the width is 3-12 minutes (i.e., 1800-7200 sampling points), ensuring coverage of the minimum time period for trending faults (referring to the physical characteristics of the daily average 0.1℃ drift of mirror contamination).

[0033] First threshold setting: The mean μ0 and standard deviation σ0 of the fault probability entropy are calculated using historical normal operating data (at least 30 days of continuous operation without faults). The first threshold is set to μ0 + 1.5σ0. If there is no historical data for the first operation of the equipment, the default first threshold is 0.7 (entropy value normalized to the range of 0-1).

[0034] Duration standard: The duration of the fault probability entropy being higher than the first threshold is ≥5 minutes (can be adjusted according to industry needs, set to 5 minutes for chemical scenarios and 10 minutes for laboratory scenarios) to avoid false triggering due to short-term fluctuations.

[0035] Applicable fault scenarios: Specifically address faults with time constants on the order of hours / days, such as mirror contamination (slow shift in condensation temperature due to contaminant adhesion), sensor aging (baseline drift caused by performance degradation of sensitive elements), and long-term power supply voltage deviation (output voltage of power supply module decreases at a rate of 0.01V / hour).

[0036] The sniper mode specifically includes: Dynamic threshold calculation: Based on the statistical characteristics of the 100 normal sampling points before triggering, the dynamic threshold = mean μ + 3 × standard deviation σ (3σ principle), and the baseline mean and standard deviation are updated every 50 sampling points to ensure adaptation to the slow changes in data distribution.

[0037] Instantaneous spike determination: The spike amplitude must simultaneously meet the following requirements: exceeding the dynamic threshold and absolute deviation > 0.5℃ (dew point measurement accuracy requirement), duration < 100 milliseconds (determined by the timestamp difference between adjacent sampling points, such as ≤ 100 consecutive points when sampling at 1 kHz).

[0038] Window width and anchoring: The minimum value is 10-50 milliseconds (10-50 points for 1kHz sampling). Centered on the peak point, anchor 5 points forward (covering the rising edge) and 5 points backward (covering the falling edge) to form a fixed anchoring range of 5 points before the peak + 1 point after the peak + 5 points after the peak (a total of 11 points) to ensure complete capture of transient features.

[0039] Applicable fault scenarios: clearly define faults with time constants in the millisecond range, such as electromagnetic interference (pulse signals generated by the start-up and shutdown of industrial power equipment, causing data to jump by more than 2°C within 10 milliseconds), power supply ripple (voltage spikes caused by instantaneous fluctuations in the power supply module, which are mapped to jumps in dew point readings), and transient breakdown of sensor circuits (abnormal instantaneous readings caused by occasional sparks).

[0040] Multi-focus modes specifically include: Non-continuous abnormal segment determination: ≥2 abnormal segments are detected within 10 minutes, and the duration of normal data between any two segments is >30 seconds (determined by the time difference between the segment termination point and the next segment start point); each segment must meet the following conditions: fault probability entropy > second threshold (second threshold = μ0 + 0.8σ0, lower than the first threshold) and segment duration 5-60 seconds (distinguishing between instantaneous spikes and trend anomalies).

[0041] Sub-window generation: The PELT (PrunedExactLinearTime) mutation point detection algorithm is used to identify the start and end points of the segment (tstart is the moment when the data first deviates from the normal range, and tend is the moment when it returns to normal); the sub-window width = segment duration + 2 seconds of redundancy (1 second before and after) to ensure coverage of the transition features before and after the segment (such as the precursor fluctuations before the anomaly occurs).

[0042] Parallel processing logic: Each sub-window independently extracts features (including abnormal amplitude, duration, deviation from the normal mean, and fluctuation frequency), and outputs a comprehensive abnormal result through weighted fusion (weight = segment fault probability entropy / sum of all segment entropy values), avoiding the dominance of single segment features in the judgment.

[0043] Applicable fault scenarios: Clearly identify faults with time constants on the order of seconds, such as intermittent gas leakage (poor valve sealing causing leakage once every 5 minutes, each lasting 10 seconds, accompanied by a 0.3℃ change in dew point reading), intermittent data transmission (5 seconds of data loss and instantaneous retransmission caused by wireless module signal blockage), and periodic condensation of sensors (condensation once every 3 minutes, lasting 8 seconds, caused by fluctuations in ambient humidity).

[0044] Cooperative constraints for form switching include: Mode priority: Sniper mode (instantaneous spikes) > Multi-focus mode (discrete fragments) > Panoramic mode (trend anomalies), ensuring that millisecond-level urgent anomalies are captured first.

[0045] Switching back mechanism: After completing the peak analysis in sniper mode (time < 1 second), it automatically returns to the original mode (such as panoramic mode); in multi-focus mode, if a sub-window meets the triggering conditions of panoramic mode, the sub-window is upgraded to panoramic mode, and the other sub-windows remain in multi-focus analysis.

[0046] Mapping library constraints: All switching must conform to the fault-window shape mapping library (e.g., when a preliminary assumption of mirror contamination is detected, the panoramic mode is forcibly triggered, ignoring the current entropy value).

[0047] As mentioned above, by clearly defining quantifiable parameters such as window width, threshold setting, trigger duration, and PELT algorithm, it is ensured that those skilled in the art can directly reproduce the shape design and switching logic of the dynamic sensing window. At the same time, by supplementing the correspondence between specific fault scenarios such as mirror contamination, electromagnetic interference, and pressure relief and shape parameters such as time constant and window width, the customized design logic of the three modes is clearly presented. This enhances the creativity of the solution in adapting to the specific fault characteristics of dew point meters, thus differentiating it from general dynamic window technology. In addition, the content such as priority, return mechanism, and mapping library constraints eliminates the ambiguity of mode switching, ensuring that the window shape is stably and reasonably adaptively adjusted under different scenarios, improving the reliability of the technical solution. Specific parameters such as 3σ dynamic threshold and 5-minute duration provide clear standards for subsequent experimental verification, making it easy to intuitively demonstrate the advantages of this solution in multi-scale anomaly identification through comparative testing.

[0048] In one embodiment, the implementation of the multi-focus mode includes the following steps: Identify the start and end points of multiple anomalous segments in time-series data; Identification of the start and end points of abnormal segments includes: The PELT mutation point detection algorithm is adopted, and the algorithm penalty coefficient λ is set to 10 (based on the variance characteristics calibration of the dew point meter time series data, which can be adjusted within the range of 5-20 according to the intensity of data fluctuation) to balance the detection sensitivity and false detection rate of abnormal segments.

[0049] The normal data range is calculated based on the absence of abnormal data in the previous hour: mean μnormal ± 2σnormal (2σ principle, covering 95% of normal data distribution). Data exceeding this range is considered a suspected abnormality.

[0050] The start point (tstart) is defined as the moment when the data first exceeds the normal range for three consecutive sampling points; the end point (tend) is defined as the moment when the data first returns to the normal range for five consecutive sampling points, thus avoiding misjudgment of segment boundaries due to fluctuations in a single sampling point.

[0051] Only segments with a duration of 5-60 seconds and an interval of >30 seconds from the previous segment are retained as valid abnormal segments, while invalid segments that are too short (<5 seconds) or too close together (interval <30 seconds) are removed.

[0052] Generate independent sub-windows centered on each abnormal fragment; The generation of independent child windows includes: The sub-window width adopts a fragment adaptive + fixed redundancy design: the actual width of the sub-window = the duration of the abnormal fragment + 2 seconds of redundancy (1 second before and after), to ensure coverage of the precursor fluctuations before the anomaly occurs and the recovery transition characteristics after it ends.

[0053] Example: If an abnormal segment lasts for 10 seconds (t1-t11), the corresponding sub-window time range is from t1-1 seconds to t11+1 seconds; if the segment lasts for 5 seconds (t20-t25), the sub-window range is from t19 to t26, to avoid feature loss due to the segment being too short.

[0054] The sub-windows are numbered sequentially according to the appearance time of the segments (e.g., sub-window 1, sub-window 2, etc.), and each sub-window independently stores the corresponding time series data and segment attributes (start and end times, initial judgment results).

[0055] Feature extraction and anomaly analysis are performed in parallel on the data within each sub-window. Parallel feature extraction and anomaly analysis include: Each sub-window performs feature extraction in parallel, and the core features extracted are clearly defined as follows: Abnormal amplitude: The maximum deviation of the data within a segment from the normal mean μnormal (unit: °C dew point temperature); Duration: The actual duration of the segment (accurate to milliseconds, calculated using sampling point timestamps); Bias: The proportion of the average deviation of all data within a segment from μnormal to the normal range (2σnormal) (normalized to 0-1); Fluctuation frequency: The number of times data within a segment crosses μnormal ± 0.1σnormal (the threshold for slight fluctuations).

[0056] Anomaly analysis logic: Compare each feature with preset thresholds (amplitude threshold = 0.3℃, deviation threshold = 0.5, fluctuation frequency threshold = 3 times / second). If ≥2 features exceed the standard, it is determined as a confirmed anomaly; otherwise, it is marked as a suspected anomaly, and the non-compliant feature items are recorded.

[0057] The analysis results are then merged and input into the diagnostic engine; The analysis results fusion and output include: The fusion weight is calculated based on the fault probability entropy (Hfault) of each sub-window: the weight of a single sub-window is Wi=Hfaulti / Σ(Hfault1~Hfaultn) (n is the number of effective sub-windows). The higher the entropy value, the greater the weight of the segment, highlighting the impact of high-confidence anomalies.

[0058] The fusion output consists of two parts: Overall anomaly confidence: Σ(Wi×sub-window anomaly confidence) (confirmed anomaly confidence = 1, suspected anomaly confidence = 0.5), threshold is set to 0.7, if it is higher than the threshold, it is judged as a multifocal anomaly; Feature vector set: Contains four core features and weights for all sub-windows, input into the diagnostic engine in a structured data format (such as JSON) to provide complete feature support for subsequent fault type matching.

[0059] As mentioned above, by clarifying quantifiable details such as PELT algorithm parameters, sub-window size rules, feature extraction terms, and fusion weight logic, the implementation process of the multi-focus mode is ensured to be completely open and reproducible. Through targeted feature extraction terms and fusion mechanisms, the discrete features of intermittent faults in dew point meters are accurately captured, avoiding the dilution of abnormal features by continuous windows. At the same time, parallel processing logic improves the efficiency of multi-segment analysis, ensuring real-time response in industrial scenarios. Ultimately, this improves the accuracy and reliability of anomaly identification in the multi-focus mode, providing comprehensive and high-quality input data for the diagnostic engine.

[0060] In one embodiment, detecting transient anomalous events based on the preprocessed time-series data includes: Real-time monitoring of the first and second differences of data points; When the first difference exceeds the threshold dynamically calculated based on historical data, and the second difference indicates that the change is an instantaneous spike rather than the start of a trend, it is judged as an instantaneous abnormal event. The characteristics of the preprocessed data include: The preprocessed time series data are dew point temperature values ​​that have been denoised by moving average (window size = 5 sampling points), with a fixed sampling frequency of 1kHz and a data format of timestamp + temperature value to ensure that the time base for differential calculation is consistent.

[0061] The rules for calculating first-order and second-order differences include: First-order difference (Δ1): Characterizes the instantaneous rate of change of the data, and is calculated using the following formula: Δ1(t) = x(t) - x(t-1); Where x(t) is the data of the current sampling point, and x(t-1) is the data of the previous sampling point, with the unit being ℃ / ms (reflecting the temperature change per millisecond).

[0062] Second-order difference (Δ2): Characterizes the trend of the rate of change, and is calculated using the following formula: Δ2(t) = Δ1(t) - Δ1(t-1); Used to distinguish the changing characteristics of instantaneous spikes from the beginning of trends.

[0063] The dynamic threshold calculation and update mechanism includes: First-order difference threshold (ThΔ1): Determined based on the statistical characteristics of the first-order difference of the first 1000 normal sampling points. Normal data must meet the requirement that no anomaly judgment is triggered for 1000 consecutive points. The calculation rule is: ThΔ1=μΔ1+3σΔ1, where μΔ1 is the mean of the normal first-order difference and σΔ1 is the standard deviation.

[0064] For example, under normal operating conditions, μΔ1 = 0.002℃ / ms and σΔ1 = 0.005℃ / ms, then ThΔ1 = 0.002 + 3 × 0.005 = 0.017℃ / ms.

[0065] Threshold update: The historical normal data window is updated every 500 sampling points (i.e., every 0.5 seconds) by rolling updates (removing the earliest 500 points and including the latest 500 points), and μΔ1 and σΔ1 are recalculated to ensure that the threshold adapts to the data baseline drift during long-term operation of the equipment.

[0066] The criteria for distinguishing between instantaneous spikes and trend initiations (based on second-order differences) include: Instantaneous peak determination: The second-order difference Δ2(t) must satisfy: ① Absolute value |Δ2(t)|>ThΔ2 (ThΔ2=μΔ2+2σΔ2, where μΔ2 and σΔ2 are the mean and standard deviation of the normal second difference, for example, set to 0.008℃ / ms²). ② The signs of Δ2(t) and Δ1(t) are opposite (e.g., if Δ1(t) is positive and Δ2(t) is negative, it indicates that the data has changed from a rapid rise to a rapid fall, which is consistent with the characteristics of a sharp rise and fall). ③ The duration of this feature is ≤5 sampling points (i.e. ≤5ms, to avoid misjudging short-term trends as spikes).

[0067] Trend initiation exclusion: If the second difference Δ2(t) and Δ1(t) have the same sign (e.g., if Δ1(t) is positive, Δ2(t) is also positive), or |Δ2(t)|≤ThΔ2, it indicates that the rate of change is continuously increasing or leveling off, which belongs to the beginning of a trend (e.g., the initial stage of slow sensor drift), and is not judged as an instantaneous anomaly.

[0068] The final criteria for determining transient abnormal events include: A transient abnormal event is determined when all of the following conditions are met: The absolute value of the first-order difference |Δ1(t)| > the dynamic threshold ThΔ1; The second-order difference conforms to the instantaneous spike characteristics (opposite sign, absolute value exceeding ThΔ2, duration ≤5ms); This event is not covered by the current analysis logic in multi-focus mode or panoramic mode (to avoid duplicate judgments).

[0069] As mentioned above, by clarifying reproducible details such as the calculation formulas for first-order / second-order differences, the quantization rules for dynamic thresholds, and the distinction criteria between instantaneous spikes and trend initiations, the detection logic for instantaneous anomalies is fully disclosed. At the same time, by accurately distinguishing between instantaneous anomalies and trend changes through the feature analysis of second-order differences, the problem of misjudging the two types of changes by the traditional single threshold method is solved, significantly improving the accuracy of instantaneous anomaly detection in dew point meters, especially suitable for the identification needs of millisecond-level transient interference in industrial scenarios.

[0070] In one embodiment, the fault-window shape mapping library includes the following mapping relationships: For sensor-sensitive detection surface contamination and sensor detection circuit drift-related faults, map to panoramic mode; For electromagnetic interference and power ripple-related faults, map to sniper mode; For pressure over-release caused by intermittent ultra-high flow at the dew point and intermittent sensor connection stability failure, the system is mapped to a multi-focus mode.

[0071] The mapping library uses a key-value pair structured storage (in JSON format). The core fields include "fault type ID", "fault name", "fault physical characteristics", "associated window shape", "window parameter configuration", "trigger priority" and "conflict arbitration rules", which facilitates quick retrieval and calling by the diagnostic engine. The mapping details between various faults and window shapes include: (1) Mirror contamination, sensor aging faults → Panoramic mode; Quantification of physical characteristics of faults: Time constant: 5-24 hours for mirror contamination (contaminant adhesion causes slow drift in condensation threshold), and 30-90 days for sensor aging (performance degradation of sensitive elements causes baseline drift). Data characteristics: All data exhibit "non-instantaneous, continuous trend changes", with no obvious peaks or discrete segments, and the maximum offset within 24 hours is ≤1℃ (minor fault) to >2℃ (serious fault). Judgment criteria: Quantitative identification based on sensor reflectivity deviation (mirror contamination) and indication error (sensor aging) (refer to the fault severity index in the correction strategy library).

[0072] Panoramic mode parameter binding: The window width is dynamically adapted according to the sampling frequency (30 minutes to 2 hours at 1Hz, 3 to 12 minutes at 10Hz), the sliding step is 1 / 6 of the window width, and the preprocessing uses Gaussian filtering (σ=5) to ensure that trend features are highlighted and high-frequency noise is suppressed.

[0073] Triggering constraint: When the diagnostic engine initially determines this type of fault, it will force the window mode to switch to panoramic mode, ignoring the temporary fluctuations of the current fault probability entropy, and continue to run for at least 1 hour after switching (to ensure that the complete trend is captured).

[0074] (2) Electromagnetic interference, power ripple faults → Sniper mode; Quantification of physical characteristics of faults: Time constant: 1-50 milliseconds (1-10 milliseconds for electromagnetic interference, 10-50 milliseconds for power supply ripple); Data characteristics: "sharp rise and fall of instantaneous peaks", peak amplitude > 0.5℃, duration ≤ 100 milliseconds, first-order difference > 0.017℃ / ms (refer to the dynamic threshold for instantaneous anomaly detection). Judgment criteria: Joint detection using first-order and second-order differences (Δ1 exceeds the threshold and Δ2 has the opposite sign to Δ1).

[0075] Sniper mode parameter binding: the window width is fixed at 10-50 milliseconds (10-50 points corresponding to 1kHz sampling), anchored with the peak value as the center (first 5 points + last 5 points), and the smoothing filter is turned off in the preprocessing to retain the original transient features.

[0076] Triggering constraints: This type of fault triggering priority is level 1 (highest). Even if the current mode is panoramic or multi-focus, it will immediately interrupt and switch to sniper mode. After completing the peak capture, it will return to the original mode (interruption duration ≤ 1 second).

[0077] (3) Intermittent gas leakage and data transmission interruption faults → Multifocal mode Quantification of physical characteristics of faults: Time constant: 1-60 seconds (10-60 seconds for intermittent gas leakage, 1-10 seconds for intermittent data transmission); Data characteristics: "Discontinuous discrete segments", with ≥2 segments appearing within 10 minutes, segment interval > 30 seconds, and data fluctuation amplitude > 0.3℃ within each segment; Judgment criteria: The start and end points of the segment are identified by the PELT mutation point detection algorithm, combined with the fault probability entropy judgment (Hfault>μ0+0.8σ0).

[0078] Multi-focus mode parameter binding: sub-window width = segment duration + 2 seconds of redundancy (1 second before and after), extract 4 types of core features in parallel (abnormal amplitude, duration, deviation, fluctuation frequency), and the fusion weight is calculated based on the fault probability entropy.

[0079] Trigger constraints: The trigger priority is level 3. If it conflicts with the sniper mode, the instantaneous spike will be processed first before switching to the multi-focus mode. If the fragment continues to expand to meet the conditions of the panoramic mode, the corresponding child window will be upgraded to panoramic mode.

[0080] The construction and update mechanism of the mapping library: Initial setup: Based on factory failure simulation data of dew point meters and typical failure cases in the industry, an initial mapping relationship was established through manual annotation, and the parameter configuration was optimized through three rounds of simulation verification.

[0081] Dynamic Update: When the diagnostic engine identifies a new fault type (not included in the mapping library), or when the feedback error of the correction effect is greater than 0.1℃ (exceeding the accuracy requirement), the mapping library is automatically updated. The new fault type is classified according to its physical characteristics and bound to the corresponding window form. After the update, it must be verified by 10 sets of test data before it can take effect.

[0082] (1) Mapping table between parameters and fault severity

[0083] (2) Example of progressive compensation (taking persistent mirror contamination failure as an example) Initial stage (0-5 min): ΔR=6% (minor fault), k1=0.02℃ / %, correction formula Tcorrection=Tmeasured-0.02×ΔR; Gradient increment (min 5-15): ΔR remains ≥3%, k1 is increased by 0.005℃ / % every 5 min, specifically 0.025℃ / % and 0.03℃ / % respectively. Stable convergence (after 15 minutes): k1 is increased to 0.03℃ / % and held. If ΔR drops below 3%, k1 is lowered back to 0.02℃ / %; Boundary constraints: k1 should not exceed 0.05℃ / %, to avoid overcompensation leading to data distortion. As mentioned above, by clearly defining the structured storage format of the mapping library, the quantitative indicators of the physical characteristics of the fault, the parameter binding rules of the window shape, and the trigger priority constraints, the implementation process of the mapping relationship is ensured to be completely open and reproducible. At the same time, the mapping library is designed based on the physical time constant and data characteristics of the dew point meter's specific faults, achieving precise binding of "fault essential characteristics - window observation mode" and avoiding shape mismatch caused by general mapping rules. In addition, the trigger priority and conflict arbitration rules eliminate the confusion of mode switching in the scenario of multiple faults superimposed, ensuring that all kinds of faults can be captured by the optimal window shape, ultimately improving the efficiency and accuracy of anomaly identification.

[0084] In one embodiment, the correction strategy library includes the following correction algorithms, and the parameters of the correction algorithms are dynamically adjusted according to the severity of the fault: To address the contamination fault on the sensor's sensitive detection surface, a historical data trend compensation and correction algorithm is employed. To address the drift fault in the sensor detection circuit, a self-calibration algorithm based on the built-in reference capacitor is adopted. To address the pressure over-release fault caused by intermittent ultra-high flow rate at the dew point, a dynamic correction algorithm for the pressure-dew point relationship is adopted. To address intermittent connection stability issues in sensors, a sliding window mid-range smoothing correction algorithm is employed. Specifically, a conservative compensation coefficient is used for minor faults; an aggressive compensation coefficient is used for severe faults; and a gradual compensation strategy is used for persistent faults.

[0085] Specifically, a conservative compensation coefficient is used for minor faults; an aggressive compensation coefficient is used for severe faults; and a gradual compensation strategy is used for persistent faults. Criteria for quantifying the severity of faults: The severity of each fault type is determined by specific quantitative indicators, with threshold values ​​set based on the dew point meter's factory calibration standards and industrial application error tolerance. The specific implementation and parameter adjustment of the correction algorithm: (1) Mirror contamination fault: Image enhancement and reflectivity compensation algorithm Image enhancement process: The mirror image is acquired by the built-in CMOS camera of the dew point meter. The CLAHE (contrast-limited adaptive histogram equalization) algorithm is used to enhance the image details. The kernel size is set to 8×8 and the contrast limit parameter is set to 2.0 to highlight the grayscale difference between the contaminated area and the clean area.

[0086] Reflectivity compensation formula: Tcorrection = Tmeasured - k1 × ΔR; Wherein, k1 is the compensation coefficient. For minor faults, k1 = 0.02℃ / % (conservative compensation to avoid overcorrection), and for severe faults, k1 = 0.05℃ / % (active compensation to offset the impact of severe pollution).

[0087] Adaptation to persistent contamination: A gradual compensation method is adopted. Initially, it is executed at a minor fault rate of k1=0.02℃ / % and ΔR is recalculated every 5 minutes. If ΔR is still ≥3%, k1 is increased in increments of 0.01℃ / % (maximum not exceeding 0.05℃ / %) to avoid data fluctuations caused by a one-time correction.

[0088] (2) Sensor drift fault: self-calibration algorithm based on precision resistor Self-calibration logic: Using the PT1000 precision resistor (accuracy ±0.1%) built into the dew point meter as a reference, the voltage signal U of the sensor temperature measurement circuit is collected in real time and compared with the voltage signal U reference corresponding to the reference resistor.

[0089] Correction formula: T_correction = T_measured × (U_baseline / U_measured) × k2, where k2 is the correction coefficient. For minor faults, k2 = 0.98-1.02 (conservative range, small correction), and for severe faults, k2 = 0.95-1.05 (positive range, significantly offsetting drift).

[0090] Continuous drift adaptation: The U-baseline and k2 are updated every 30 minutes, following a gradual process of "initial calibration → error verification → coefficient fine-tuning". Each fine-tuning increment does not exceed ±0.01, ensuring a smooth calibration process.

[0091] (3) Pressure relief failure: Dynamic correction algorithm for pressure-dew point relationship Pressure correlation model: Based on the Clausius-Clapeyron equation correction, the mapping relationship between dew point temperature and gas pressure is established: T_correction = T_measured + k3 × (P_standard - P_measured), where P_standard is the rated pressure of the gas line (default 0.4 MPa) and k3 is the pressure compensation factor.

[0092] Parameter adjustment: for minor leakage, k3=0.3℃ / MPa (conservative compensation, adapting to small pressure fluctuations); for severe leakage, k3=0.8℃ / MPa (active compensation, offsetting large pressure deviations).

[0093] Continuous leakage adaptation: "Segmented compensation + trend prediction" is adopted. The average leakage rate is calculated every minute, and k3 is finely adjusted in advance according to the increasing trend of leakage rate. Each adjustment is ≤0.1℃ / MPa to avoid lag error.

[0094] (4) Data transmission interference: Sliding window midpoint smoothing correction algorithm; Window configuration: The sliding window size is set to 15 sampling points (15 seconds when the sampling frequency is 1Hz). Smoothing is only performed on data points with detected interference jumps, and normal data points are output directly.

[0095] Correction logic: Replace the jump point value with the median of the 14 data points within the window excluding the jump point. The compensation coefficient k4 controls the smoothing intensity. When there is slight interference, k4=0.7 (70% of the original data features are retained, conservative smoothing) and when there is severe interference, k4=0.3 (30% of the original data features are retained, aggressive smoothing).

[0096] Persistent interference adaptation: Gradually expand the window to 25 sampling points, while gradually increasing k4 to 0.5 (intermediate value) to balance the smoothing effect and data response speed.

[0097] Parameter adjustment triggering and update mechanism: Triggering timing: The correction algorithm parameters are linked with the severity of the fault in real time. The fault quantification index is re-evaluated every 10 seconds, and the corresponding coefficient is updated immediately when the index jumps through levels.

[0098] Boundary constraints: All compensation coefficients are subject to limits to prevent correction distortion caused by parameter overflow.

[0099] As mentioned above, by clearly defining the quantitative indicators of fault severity, the specific formulas and parameter ranges of the correction algorithm, and the execution logic of progressive compensation, the implementation process of the correction strategy library is ensured to be completely open and reproducible. At the same time, dynamic parameter adjustment based on fault type and severity avoids the "one-size-fits-all" defects of traditional fixed parameter correction. Minor faults are conservatively corrected to prevent excessive intervention, severe faults are actively corrected to ensure error cancellation, and continuous faults are progressively corrected to ensure data stability. Ultimately, this significantly improves the accuracy, adaptability, and stability of abnormal data correction for dew point meters, keeping the data error after correction within ±0.1℃, meeting the requirements of industrial-grade measurement accuracy.

[0100] In one embodiment, the verification correction effect includes: Short-term verification: Check whether the corrected data has returned to the normal fluctuation range; Interim validation: Evaluate the stability of the device output after correction; Long-term validation: Analyze the impact of correction strategies on equipment lifespan; The verification results are quantified into performance scores, which are used to optimize the algorithm weights in the correction strategy library. Short-term validation: a check to restore the normal range of the corrected data; Verification time window: Starts immediately after the correction strategy is executed and lasts for 5 minutes (corresponding to 300 sampling points when the sampling frequency is 1Hz) to ensure coverage of the initial stable phase after correction.

[0101] Definition of normal fluctuation range: The benchmark is calculated based on the equipment's historical normal operation data (at least 30 days of no fault records): mean μbase ± 2σbase (2σ principle, covering 95% of normal data distribution), where μbase is the benchmark mean of dew point temperature and σbase is the standard deviation of normal fluctuation.

[0102] Judgment criteria: If within 5 minutes after correction, 30 consecutive sampling points (i.e., 30 seconds) all fall within the range of μbase±2σbase, and the maximum deviation of a single point is ≤0.08℃ (stricter than the normal range threshold to ensure recovery quality), then the short-term verification is passed; otherwise, it is marked as "short-term deviation", and the number of points out of range and the maximum deviation value are recorded.

[0103] Interim validation: Stability assessment of the device output after correction; Validation timeframe: Initiated after short-term validation is passed, lasting 24 hours, covering one full shift of equipment operation, to assess the sustainability of the correction effect.

[0104] Stability quantification indicators: Variance fluctuation (σcorr): Calculates the standard deviation of the corrected data within 24 hours, which must satisfy σcorr≤1.2×σbase (the maximum allowable deviation from normal fluctuation is 1.2 times, to control stability decay). Trend drift (ΔTtrend): Linear fitting of 24-hour data, with an absolute slope |k| ≤ 0.005℃ / hour (ensuring no significant quadratic drift); Anomaly recurrence rate: The number of times the same anomaly reappears after correction is ≤ 1 time / 24 hours (to evaluate the suppression effect of correction on the root cause of the fault).

[0105] Judgment criteria: If σcorr≤1.2×σbase, |k|≤0.005℃ / hour, and anomaly recurrence rate≤1 time / 24 hours are simultaneously met, the interim verification is passed; otherwise, it is marked as "interim instability" and the specific values ​​of the non-compliant indicators are recorded.

[0106] Long-term validation: Analysis of the impact of correction strategies on equipment lifespan; Validation timeframe: Initiated after successful interim validation and lasting for 3 months to assess the long-term security of the mitigation strategy.

[0107] Quantitative indicators of lifespan impact: Sensor loss rate: Compare the response time T90 (time to reach a stable 90% humidity value) of the sensor before and after correction. The increase in T90 after correction is ≤10% (avoid over-correction to prevent the sensor from aging). Mirror cleaning cycle: The interval between when the mirror contamination causes a reflectivity deviation ΔR ≥ 5% after correction, which is shortened by ≤ 15% compared to the historical cycle when no correction strategy was used (to control the additional damage to the mirror caused by correction). Circuit module temperature rise: When the correction algorithm is running, the temperature rise of the core control module is ≤2℃.

[0108] Judgment criteria: If the sensor loss rate is ≤10%, the mirror cleaning cycle is shortened by ≤15%, and the circuit temperature rise is ≤2℃, then the long-term verification is passed; otherwise, it is marked as "long-term loss exceeds the standard", and the specific changes in the loss indicators are recorded.

[0109] Performance scoring and strategy optimization logic; Scoring and Quantification Rules: A 100-point scale is used, with short-term, medium-term, and long-term verifications accounting for 30%, 40%, and 30% of the total score, respectively. Each verification is scored according to its degree of achievement. Short-term verification: 30 points for a perfect pass; 20 points for ≤5 points exceeding the range; 0-10 points for >5 points exceeding the range (proportional conversion). Mid-term verification: 40 points for meeting all 3 indicators; 20 points for failing to meet 1 indicator; 0-10 points for failing 2 or more indicators. Long-term verification: 30 points are awarded if all 3 indicators are met; 15 points are awarded if 1 indicator is not met; and 0-5 points are awarded if 2 or more indicators are not met.

[0110] Strategy optimization mechanism: For correction algorithms with an effect score ≥ 80, their weight in the correction strategy library is increased by 10%; the weight of algorithms with a score of 60-79 remains unchanged; the weight of algorithms with a score < 60 is reduced by 20%, and algorithm parameter recalibration is triggered; for algorithms with a score < 60 for 3 consecutive times, they are temporarily stored in the strategy library and a replacement algorithm test is started.

[0111] As described above, this embodiment ensures that the verification process of the correction effect is completely open and reproducible by clearly defining the time windows, quantitative indicators, judgment criteria, and effect scoring rules for short-term / medium-term / long-term verification. At the same time, the three-level verification system comprehensively evaluates the correction effect from immediate recoverability and continuous stability to long-term security, avoiding the one-sidedness of traditional single verification. In addition, the strategy optimization mechanism based on scoring realizes the dynamic iteration of the correction algorithm, enabling the strategy library to continuously adapt to changes in equipment status and fault characteristics, ultimately improving the reliability, adaptability, and long-term economy of the correction strategy, and ensuring that the accuracy of the corrected data and the lifespan of the equipment both meet the standards.

[0112] Example 2: A dew point meter measurement data processing device, comprising: The dynamic sensing window module is configured to construct and control the shape switching of the dynamic sensing window based on the received dew point meter time-series measurement data to identify anomalies. The intelligent correction execution module is configured to execute the corresponding correction strategy based on the anomaly identification result. The effect evaluation and feedback module is configured to verify the correction effect and optimize the identification and correction strategies. The output of the dynamic perception window module is connected to the input of the intelligent correction execution module, and the input of the effect evaluation feedback module collects data from the dynamic perception window module and the intelligent correction execution module respectively, and the output is fed back to the two modules mentioned above.

[0113] The intelligent correction execution module includes: The correction strategy selection submodule is configured to select the optimal correction algorithm based on the fault type. The parameter adaptive adjustment submodule is configured to dynamically adjust the correction parameters according to the severity of the fault. The correction effect monitoring submodule is configured to monitor the correction process in real time and prevent overcorrection.

[0114] The dynamic sensing window module includes: The panorama processing submodule is configured to process wide-window data in panorama mode and identify systematic drift faults. The sniper processing submodule is configured to process narrow-window data in sniper mode and capture instantaneous interference features; The multi-focus processing submodule is configured to process multiple non-continuous sub-window data in parallel in multi-focus mode to analyze intermittent anomalies. The morphological arbitration submodule is configured to determine the optimal window shape based on multi-source input.

[0115] Overall hardware architecture and connection method of the device; Hardware platform: Based on an industrial-grade ARM Cortex-A9 processor, integrating a 16-bit ADC data acquisition unit, DDR3 memory, Ethernet / 485 communication interface and Flash storage.

[0116] Physical connection between modules: The dynamic perception window module and the intelligent correction execution module are connected through an internal high-speed data bus to realize the real-time transmission of anomaly identification results; the effect evaluation feedback module is connected to the former two through bidirectional GPIO interfaces respectively, with a data acquisition delay of ≤100ms and a feedback control signal response time of ≤50ms.

[0117] Data interaction format: Data between modules is encapsulated in JSON format, including fields such as "timestamp", "fault ID", "abnormal feature vector", "corrected data", and "effect score", to ensure that the information is complete and parsable.

[0118] Configuration of each submodule of the dynamic sensing window module; (1) Panoramic processing submodule Hardware configuration: Includes an independent data buffer and a low-pass filter circuit.

[0119] Core functionality implementation: Wide window data processing: The buffer is dynamically allocated according to the principle of "window width of 30 minutes to 2 hours when sampling frequency is 1Hz", and a sliding operation is performed every 5 minutes; Systematic drift identification: The slope (k) of the data within the window is calculated by linear regression. If |k| > 0.01℃ / hour and lasts for 5 minutes, it is determined to be a drift fault. Output features: including drift start time, current offset, and drift rate, sent to the morphological arbitration submodule in feature vector format.

[0120] (2) Sniper processing submodule Hardware configuration: Includes a high-frequency data acquisition unit and a transient feature extraction circuit.

[0121] Core functionality implementation: Narrow window data processing: Fixed window width 10-50ms, anchored with the peak value as the center; Transient interference capture: The first-order differential is monitored in real time through a comparator circuit. When Δ1 > 0.017℃ / ms, capture is triggered, and the peak amplitude and rise time are recorded synchronously. Output characteristics: including peak time, amplitude, and duration, which are sent to the morphological arbitration submodule in the form of interrupt signal + data frame.

[0122] (3) Multi-focus processing submodule Hardware configuration: Includes 4 parallel data processing channels and a dedicated FPGA for segment recognition.

[0123] Core functionality implementation: Non-contiguous sub-window processing: The start and end points of segments are identified using the PELT algorithm, and each channel is allocated an independent buffer; Intermittent anomaly analysis: Extract four types of features from each sub-window in parallel: "abnormal amplitude, duration, deviation, and fluctuation frequency"; Output features: including feature vectors of each segment and comprehensive anomaly confidence, sent to the morphological arbitration submodule.

[0124] (4) Morphological Arbitration Submodule Multiple input sources include fault probability entropy (Hfault), output features of each processing submodule, and association rules of the fault-window shape mapping library.

[0125] Decision-making logic: Priority order: Sniper mode > Multi-focus mode > Panoramic mode; Conflict arbitration: When the triggering conditions of multiple modes are met simultaneously, they are executed according to priority. The low-priority mode is paused and its current state is cached, and the high-priority mode resumes after completion. Output: Sends "start / pause" commands to each processing submodule, and sends the final anomaly identification result to the intelligent correction execution module via the SPI bus.

[0126] Configuration of each submodule of the intelligent correction execution module; (1) Correction Strategy Selection Submodule Policy matching rules: Based on the "fault ID" output by the dynamic perception window module, search for matching items in the correction policy library. The matching response time is ≤10ms.

[0127] Optimal algorithm determination: When there are multiple candidate algorithms, select the algorithm with a historical performance score of ≥80. If the scores are the same, prioritize the algorithm with the smaller computational cost.

[0128] (2) Parameter adaptive adjustment submodule Parameter adjustment basis: Dynamically adjusted according to the quantitative indicators of fault severity, with the adjustment step size calculated as "severity level × base step size".

[0129] Example: When there is a minor mirror contamination fault, the compensation coefficient k1 starts from the default 0.02℃ / % and increases by 0.005℃ / % every 5 minutes; when there is a serious fault, k1 is directly set to 0.05℃ / % and maintained.

[0130] (3) Correction effect monitoring submodule Overcorrection threshold: If the absolute value of the deviation between the corrected data and the standard value is greater than 0.1℃, or if the deviation of three consecutive sampling points shows an increasing trend, it is judged as overcorrection.

[0131] Intervention mechanism: When overcorrection is detected, the current correction algorithm is immediately paused, conservative mode is enabled, and an alarm signal is sent to the effect evaluation feedback module.

[0132] Configuration of the effect evaluation feedback module; Data collection points: Raw data and anomaly identification results are collected from the dynamic perception window module; Collect corrected data, correction parameters, and overcorrection alarms from the intelligent correction execution module; Feedback content: For the dynamic sensing window module: output window shape parameter adjustment suggestions and fault-mapping library optimization items; For the intelligent correction execution module: output the updated weight values ​​of the correction algorithm and the correction of the parameter adjustment range; Feedback cycle: short-term verification feedback, medium-term verification feedback, long-term verification feedback, and immediate feedback in case of emergency alarms.

[0133] A dew point meter measurement data processing device, configured to implement the above method, includes: The dynamic sensing window module is configured to construct a dynamic sensing window to preprocess the time-series measurement data of the dew point meter and obtain preprocessed data. The data stream calculation module is used to calculate the fault probability entropy of the data stream in parallel based on the preprocessed data and detect instantaneous abnormal events, and obtain the fault probability entropy calculation result and the instantaneous abnormal event detection result, respectively. The window shape decision module is used to input the fault probability entropy calculation result and the instantaneous abnormal event detection result into the fault-window shape mapping library constructed based on the fault physical time constant to obtain the window shape decision result. The window switching module is used to control the dynamic sensing window to adaptively switch between panoramic mode, sniper mode and multi-focus mode according to the window shape decision result, so as to obtain the abnormal recognition result after window switching. The correction module is used to select a corresponding correction algorithm from the correction strategy library based on the anomaly identification result after the window switching to correct the abnormal data in the preprocessed data and obtain the corrected data. The verification module is used to verify the correction effect of the correction algorithm on the corrected data, obtain the verification result, and simultaneously feed the verification result back to the fault-window shape mapping library and the correction strategy library.

[0134] As mentioned above, by clearly defining the equipment hardware architecture, the connection methods between modules, the specific configurations of each sub-module, and the data interaction format, the implementation process of the anomaly identification and correction equipment is ensured to be completely open and reproducible. At the same time, the parallel processing of multiple sub-modules and the morphological arbitration mechanism of the dynamic perception window module enable efficient identification of multi-scale anomalies; the strategy selection and parameter adjustment logic of the intelligent correction execution module ensures correction accuracy; and the bidirectional optimization mechanism of the effect evaluation feedback module enables the equipment to have self-iterative capabilities, ultimately forming a closed-loop system of "identification-correction-evaluation-optimization", which significantly improves the measurement reliability and long-term stability of the dew point meter in industrial scenarios.

[0135] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A dew point meter measurement data processing method, characterized by, The method comprises the following steps: Pretreatment of the dew point instrument timing measurement data by constructing a dynamic perception window, to obtain pretreatment data; Based on the pretreatment data, the fault probability entropy of the data stream is calculated in parallel and transient abnormal events are detected, to obtain the fault probability entropy calculation result and the transient abnormal event detection result respectively; The fault probability entropy calculation result and the transient abnormal event detection result are jointly input into a fault-window morphology mapping library constructed based on the fault physical time constant, to obtain a window morphology decision result; According to the window morphology decision result, the dynamic perception window is adaptively switched between panoramic mode, sniper mode and multi-focus mode, to obtain an abnormal recognition result after window switching; Based on the abnormal recognition result after window switching, a corresponding correction algorithm is selected from a correction strategy library to correct the abnormal data in the pretreatment data, to obtain corrected data; The correction effect of the correction algorithm on the corrected data is verified, to obtain a verification result, and the verification result is fed back to the fault-window morphology mapping library and the correction strategy library.

2. The dew point meter measurement data processing method according to claim 1, characterized in that: The fault probability entropy is a comprehensive index obtained by weighting the data uncertainty represented by the Shannon entropy of the data in the dynamic perception window and the model uncertainty represented by the prediction confidence of the current data input into the diagnostic model; The calculation formula of the fault probability entropy is: Hfault=α×Hdata+β×Hmodel; Wherein, Hdata is the Shannon entropy calculated based on the data in the dynamic perception window, Hmodel is the prediction confidence of the diagnostic model, and α and β are weight coefficients that are self-adaptive according to the equipment health status, wherein α+β=1; The dynamic adjustment of the weight coefficients α and β includes: When the sensor health degree is lower than the preset threshold, increase the model confidence weight β and reduce the data Shannon entropy weight α; When the environmental interference intensity is higher than the preset threshold, increase the data Shannon entropy weight α and reduce the model confidence weight β.

3. The dew point instrument measurement data processing method according to claim 2, characterized in that: The form of the dynamic perception window at least includes panoramic mode, sniper mode and multi-focus mode; The form switching of the dynamic perception window follows the following rules: When the fault probability entropy is continuously higher than the first threshold, the window is switched to the panoramic mode, and the width is expanded to the maximum value; When a transient data spike is detected and its amplitude exceeds the dynamic threshold, the window is switched to the sniper mode, the width is contracted to the minimum value and anchored around the spike event; When the system detects multiple time discontinuous suspected abnormal segments, the window is switched to the multi-focus mode, forming multiple non-continuous sub-windows to surround each abnormal segment.

4. The dew point instrument measurement data processing method according to claim 3, characterized in that: The implementation of the multi-focus mode includes the following steps: Identify the start and end points of multiple abnormal segments in the time series data; Generate independent sub-windows centered on each abnormal segment; Perform feature extraction and abnormal analysis on the data in each sub-window in parallel; Fuse the analysis results and input them into the diagnostic engine.

5. The dew point instrument measurement data processing method of claim 1, wherein: the subsequent parallel calculation of the failure probability entropy of the data stream and the detection of transient abnormal events based on the preprocessed data comprises: real-time monitoring of the first-order difference and the second-order difference of the data points; when the first-order difference exceeds the threshold value dynamically calculated based on the historical data, and the second-order difference indicates that the change is a transient spike rather than a trend start, it is determined to be a transient abnormal event.

6. The dew-point meter measurement data processing method of claim 1, wherein: The failure-window morphology mapping library contains the following mapping relationships: For sensor sensitive detection surface pollution, sensor detection circuit drift and other faults, it is mapped to the panoramic mode; For electromagnetic interference, power ripple and other faults, it is mapped to the sniper mode; For intermittent super-large flow of dew point leading to pressure over-discharge, intermittent connection stability fault of sensor, it is mapped to the multi-focus mode.

7. The dew point instrument measurement data processing method of claim 1, wherein: The correction strategy library contains the following correction algorithms, and the parameters of the correction algorithms are dynamically adjusted according to the severity of the fault: For sensor sensitive detection surface pollution fault, a historical data trend compensation correction algorithm is used; For sensor detection circuit drift fault, a self-correcting algorithm based on precision resistance is used; For pressure over-discharge fault caused by intermittent super-large flow of dew point, a pressure-dew point relationship dynamic correction algorithm is used; For intermittent connection stability fault of sensor, a sliding window median smoothing correction algorithm is used; For minor faults, a conservative compensation coefficient is used; for serious faults, an aggressive compensation coefficient is used; for persistent faults, a gradual compensation strategy is used.

8. The dew-point meter measurement data processing method of claim 1, wherein: The verification of the correction effect of the correction algorithm on the corrected data includes: Short-term verification: check if the corrected data returns to the normal fluctuation range; Medium-term verification: evaluate the stability of the output of the corrected device; Long-term verification: analyze the impact of the correction strategy on the service life of the device; The verification results are quantified as effect scores, which are used to optimize the algorithm weights in the correction strategy library.

9. A dew point meter measurement data processing device, characterized by, The configuration is to realize the method in any one of claims 1-8, comprising: a dynamic perception window module configured to build a dynamic perception window to preprocess the dew point instrument time series measurement data and obtain preprocessed data; a data stream calculation module for parallel calculation of the failure probability entropy of the data stream and the detection of transient abnormal events based on the preprocessed data, respectively obtaining the failure probability entropy calculation result and the transient abnormal event detection result; a window morphology decision module for inputting the failure probability entropy calculation result and the transient abnormal event detection result into a failure-window morphology mapping library based on the failure physical time constant, and obtaining a window morphology decision result; a window switching module for controlling the dynamic perception window to adaptively switch between the panoramic mode, the sniper mode and the multi-focus mode according to the window morphology decision result, and obtaining the abnormal recognition result after window switching; a correction module for selecting a corresponding correction algorithm from a correction strategy library based on the abnormal recognition result after window switching to correct the abnormal data in the preprocessed data, and obtaining corrected data. The verification module is used to verify the correction effect of the correction algorithm on the corrected data, obtain the verification result, and simultaneously feed the verification result back to the fault-window shape mapping library and the correction strategy library.

Images