A method for quantitatively monitoring blockage of an air preheater based on differential pressure time sequence characteristics

Abstract

This invention relates to the field of air preheater process control technology in thermal power units, and discloses a method for quantitative monitoring of air preheater blockage based on differential pressure time-series characteristics. The method includes: collecting differential pressure signals at the air preheater inlet and outlet to form raw differential pressure time-series data; generating absolute differential pressure time-series data after denoising, and then generating fluctuation-normalized differential pressure time-series data; extracting differential pressure level features, fluctuation morphology features, and spectral energy features from a continuous data window, constructing adjacent window evolution features, and screening key time-series features; inputting the key time-series features into a support vector regression model to obtain a quantitative index of blockage degree; and using soot blowing feedback events to form model correction samples and update the model. This invention achieves continuous quantitative monitoring and graded early warning of air preheater blockage, improving its resistance to operating condition interference and long-term monitoring stability.

Description

Technical Field

[0001] This invention relates to the field of process control technology for air preheaters in thermal power units, and specifically to a quantitative monitoring method for air preheater blockage based on differential pressure timing characteristics. Background Technology

[0002] An air preheater is a heat exchange device in the waste heat recovery system of flue gas at the tail end of a thermal power unit boiler. When flue gas passes through the heat exchange elements of the air preheater, it transfers heat to the primary and secondary air entering the boiler, thereby reducing the exhaust gas temperature and increasing the temperature of the air entering the furnace. After a selective catalytic reduction (SCR) denitrification device is installed in a thermal power unit, sulfur oxides, ammonia escape components, water vapor, and fly ash in the flue gas can combine to form adhesive deposits under suitable temperature conditions. These deposits adhere to the surface of the heat exchange elements in the low-temperature region of the air preheater and capture fly ash, gradually narrowing the flow channels between the heat exchange elements and increasing the flow resistance of the flue gas through the air preheater.

[0003] Increased flow resistance in the air preheater manifests as a larger pressure differential between the inlet and outlet. Therefore, during operation, pressure measurements at the air preheater inlet and outlet can be used to assess blockage changes. Pressure differential signals are characterized by simple measurement equipment, easy integration into distributed control systems, and continuous data acquisition. However, the pressure differential between the air preheater inlet and outlet is not solely affected by the degree of blockage. Changes in boiler load cause variations in flue gas volume, changes in coal quality alter flue gas composition and fly ash characteristics, fan adjustments cause short-term changes in flue gas flow rate, and pressure measurement devices and pressure-sensing pipelines also introduce random fluctuations. Therefore, the same blockage state can exhibit different instantaneous pressure differentials at different operating times, and the same instantaneous pressure differential may correspond to different stages of blockage evolution.

[0004] If only a fixed differential pressure threshold is used for blockage detection, it is difficult to utilize the fluctuation structure and evolution information of the differential pressure signal over continuous time. If zero-mean standardization is directly applied to each data window to eliminate dimensional differences, and then the average differential pressure level characteristics are calculated from the standardized data, the flow resistance information contained in the average pressure difference of the window is eliminated, making it impossible to simultaneously preserve information reflecting the absolute level of blockage and information reflecting short-term fluctuation patterns.

[0005] Machine learning regression models can be used to establish a nonlinear mapping between monitoring features and equipment state variables. However, if the model input features are merely routine statistics at the current moment or within a single data window, the model is still prone to interpreting short-term operating disturbances as blockage changes. Furthermore, after soot blowing in the air preheater, the surface deposition state of the heat exchange elements changes, and the coal quality, the aging degree of the heat exchange elements, and the differential pressure measurement baseline will gradually change over long-term operation. Training the model using only initial historical samples and fixing it for an extended period will result in a gradual deviation between the model's state scale and the actual equipment state. Summary of the Invention

[0006] The purpose of this invention is to provide a quantitative monitoring method for air preheater blockage based on differential pressure time-series characteristics. This method solves the technical problems of existing air preheater blockage monitoring methods, which are unable to utilize differential pressure signals to simultaneously retain absolute flow resistance information and fluctuation structure information, and whose quantitative monitoring models cannot adaptively correct with soot blowing operations, leading to a long-term decline in monitoring accuracy.

[0007] To solve the above-mentioned technical problems, the technical solution adopted by the present invention is as follows:

[0008] A quantitative monitoring method for air preheater blockage based on differential pressure time-series characteristics, comprising:

[0009] Collect the differential pressure signals at the inlet and outlet of the air preheater to form raw differential pressure time-series data;

[0010] The original differential pressure time series data is denoised by moving average, and the denoised absolute differential pressure time series data is retained. The fluctuation-normalized differential pressure time series data is then formed from the absolute differential pressure time series data.

[0011] In the current data window and the preceding data window, pressure difference level features are extracted from the absolute pressure difference time series data, and fluctuation morphology features and spectral energy features are extracted from the fluctuation normalized pressure difference time series data. Based on the changes of the corresponding features between the current data window and the preceding data window, adjacent window evolution features are formed. The pressure difference level features, fluctuation morphology features, spectral energy features and adjacent window evolution features are combined into a candidate time series feature set, and key time series features are selected from the candidate time series feature set.

[0012] Key time-series features are input into a support vector regression model trained based on congestion level calibration samples to obtain a quantitative indicator of congestion level.

[0013] After the dust removal feedback event, model calibration samples are formed according to the same feature construction process. The model calibration samples are used to update the support vector regression model, and the updated support vector regression model is used for subsequent data window congestion calculation.

[0014] As a preferred technical solution of the present invention, the air preheater inlet and outlet pressure difference signals are sampled at a frequency of 1Hz to form the original pressure difference time series data. The window length of the current data window and the previous data window is 10min, and the current data window and the previous data window are continuous in time.

[0015] As a preferred technical solution of the present invention, the moving average denoising adopts a moving window with 5 sampling points; the absolute pressure difference time series data directly retains the magnitude of the filtered pressure difference; the fluctuation normalized pressure difference time series data is obtained by subtracting the mean of the corresponding data window from the absolute pressure difference time series data and then scaling it according to the standard deviation of the corresponding data window.

[0016] As a preferred embodiment of the present invention, the differential pressure level characteristics include at least two of the following: mean, variance, peak value, and peak-to-peak value; the fluctuation morphology characteristics include at least one of the following: zero-crossing rate, waveform factor, and impulse factor; the spectral energy characteristics include at least one of the following: dominant frequency, preset frequency band energy proportion, and power spectral density distribution, wherein the dominant frequency is the frequency corresponding to the peak value of the power spectral density; the process of extracting the spectral energy characteristics includes: obtaining the absolute differential pressure mean of the absolute differential pressure time series data within the current data window; calculating the adaptive cutoff frequency based on the absolute differential pressure mean, the preset Strouhal number of the air preheater heat exchange element, the characteristic length of the heat exchange element, and the flue gas density inside the air preheater, wherein the adaptive cutoff frequency is the square root of the quotient of the Strouhal number divided by the characteristic length and then multiplied by twice the absolute differential pressure mean divided by the flue gas density; dividing the integral interval of the power spectral density function into a low-frequency band and a high-frequency band based on the adaptive cutoff frequency; and extracting the discrete spectral energy proportion of the fluctuation normalized differential pressure time series data in the low-frequency band and the high-frequency band as the spectral energy characteristics.

[0017] As a preferred technical solution of the present invention, the adjacent window evolution characteristics are formed by comparing the pressure difference level characteristics, fluctuation pattern characteristics or spectral energy characteristics with the same name and physical meaning in the current data window and the previous data window. The adjacent window evolution characteristics include at least the change direction and change amount of the corresponding characteristics.

[0018] As a preferred technical solution of the present invention, the selection of key time series features from the candidate time series feature set includes: calculating the Pearson correlation coefficient between the candidate time series features and the congestion degree calibration value based on historical congestion degree calibration samples, and determining the candidate time series features with an absolute value of correlation coefficient greater than 0.7 as key time series features.

[0019] As a preferred technical solution of the present invention, the support vector regression model adopts the radial basis kernel function, and the regularization parameter, kernel parameter and insensitive loss parameter are determined by the cross-validation results of the congestion degree calibration samples. The congestion degree calibration samples adopt a uniform congestion degree calibration method within the same model training period.

[0020] As a preferred technical solution of the present invention, the air preheater inlet and outlet differential pressure signals are tested for data validity before forming key timing features; when data that exceeds the range, remains constant for a long time, or has a rate of change that exceeds the allowable range of the equipment is detected, a data quality mark is generated, and continuous monitoring is completed using backup differential pressure sensor data or historical valid data; data windows containing abnormal substitute data are not used as model calibration samples.

[0021] As a preferred technical solution of the present invention, after the soot blowing feedback event ends and the system enters a stable operating state, the time series data of the pressure difference after soot blowing is collected. Model calibration samples are formed according to the same noise reduction, dual-path data construction, window feature extraction, and key time series feature screening process as online blockage monitoring. The process of updating the support vector regression model using the model calibration samples includes: extracting the mean absolute pressure difference before soot blowing during the stable operating phase before the soot blowing feedback event; extracting the mean absolute pressure difference after soot blowing during the stable operating state after the soot blowing feedback event; calculating the soot blowing efficiency evaluation coefficient, which is equal to the difference between the mean absolute pressure difference before soot blowing and the mean absolute pressure difference after soot blowing divided by the difference between the mean absolute pressure difference before soot blowing and the benchmark reference pressure difference under completely clean conditions; calculating the sample weighting factor of the corresponding model calibration sample based on the soot blowing efficiency evaluation coefficient; and using a constraint condition with the sample weighting factor as the upper limit of the penalty coefficient to solve the dual problem of the support vector regression model to complete the model update; the update of the support vector regression model ends when the calibration sample processing is completed and the model verification error meets the preset accuracy conditions.

[0022] As a preferred technical solution of the present invention, the original differential pressure time series data, key time series characteristics, quantitative indicators of blockage degree, data quality markers, soot blowing feedback events and early warning information are written into the air preheater blockage monitoring database according to the timestamp association.

[0023] As a preferred technical solution of the present invention, the quantitative index of the degree of blockage is sent to the distributed control system of the thermal power unit, and the corresponding early warning threshold group is selected according to the load status of the thermal power unit and the historical blockage records under the same operating conditions. The quantitative index of the degree of blockage is compared with the early warning threshold group to form graded early warning information.

[0024] As a preferred technical solution of the present invention, the air preheater is a rotary air preheater for thermal power units. The quantitative index of the degree of blockage and the online soot blowing event are displayed in the distributed control system according to time correlation. When the quantitative index of the degree of blockage reaches the corresponding early warning condition, a soot blowing maintenance prompt is output.

[0025] Meanwhile, the present invention also discloses a quantitative monitoring system for air preheater blockage based on differential pressure timing characteristics, including a differential pressure signal acquisition interface, an industrial processor, a memory, and a distributed control system communication interface;

[0026] The differential pressure signal acquisition interface receives pressure signals from the air preheater inlet and outlet pressure measuring devices and outputs the air preheater inlet and outlet differential pressure signals to the industrial processor.

[0027] The memory stores the original differential pressure timing data, blockage degree calibration samples, support vector regression models, and program instructions;

[0028] The industrial processor executes program instructions to perform moving average denoising on the raw differential pressure time series data and generate absolute differential pressure time series data and fluctuation-normalized differential pressure time series data. It generates differential pressure level characteristics, fluctuation morphology characteristics, spectral energy characteristics, and adjacent window evolution characteristics from the current data window and previous data windows. It selects key time series characteristics and inputs them into the support vector regression model to generate quantitative indicators of blockage degree. After receiving a soot blowing feedback event, it generates model correction samples and updates the support vector regression model. The industrial processor uses the updated support vector regression model for subsequent data windows and outputs quantitative indicators of blockage degree and early warning information through the distributed control system communication interface.

[0029] Compared with the prior art, the present invention has the following beneficial effects:

[0030] This invention employs a quantitative monitoring method for air preheater blockage based on differential pressure time-series characteristics. This method enables continuous and accurate quantitative monitoring of air preheater blockage levels using only existing inlet and outlet differential pressure signals without the need for complex internal imaging equipment. It solves the problems of existing technologies where a single differential pressure threshold is susceptible to operational interference and cannot distinguish between blockage and load fluctuations. The core solution constructs two parallel processing paths: absolute differential pressure time-series data and fluctuation-normalized differential pressure time-series data. This simultaneously preserves the differential pressure magnitude reflecting the absolute level of flue gas flow resistance and the fluctuation pattern reflecting changes in the microstructure of ash accumulation, avoiding misjudgments caused by the elimination of key resistance features in traditional standardization processing. By introducing the evolution characteristics of adjacent windows between the current and previous data windows, the static statistics of a single window are expanded into a description of state changes across time dimensions. This allows the assessment of blockage levels to consider both the current state level and the state evolution trend, significantly enhancing the algorithm's resistance to short-term operational disturbances and random fluctuations. After inputting the multi-dimensional features into a support vector regression model, continuous quantitative indicators of blockage levels are output, providing a direct basis for graded early warning and precise operation and maintenance.

[0031] Furthermore, in terms of spectral energy feature extraction, an adaptive cutoff frequency based on the dynamic calculation of the mean absolute pressure difference and its correlation with flue gas velocity was used to accurately classify the frequency domain energy of pressure difference fluctuations within a wide load adjustment range of the unit. This overcomes the shortcomings of fixed frequency thresholds in terms of feature distortion under low or high loads, ensuring that the spectral energy features always accurately reflect changes in ash accumulation morphology. Regarding long-term model maintenance, a quantitative sootblowing efficiency evaluation mechanism was established. Sample weighting factors were calculated using sootblowing efficiency evaluation coefficients, and the upper limit of the penalty coefficient during model updates was constrained by these weighting factors. This effectively identified and suppressed the negative pull of invalid sootblowing samples caused by sootblowing gun jamming or insufficient steam parameters on the regression hyperplane, maintaining the purity of the support vector distribution in the feature space. This ensured long-term accurate alignment between the digital model and the physical state of the air preheater, improving the stability and maintainability of the monitoring system during long-term operation. Attached Figure Description

[0032] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the embodiments will be briefly introduced below. It should be understood that the following drawings only show some embodiments of the present invention and should not be regarded as a limitation of the scope. For those skilled in the art, other related drawings can be obtained from these drawings without creative effort.

[0033] Figure 1 This is a flowchart illustrating the overall process of the method of the present invention.

[0034] Figure 2 This is a flowchart of the sub-process for preprocessing differential pressure data of the air preheater and constructing dual-path data in this invention.

[0035] Figure 3 This is a flowchart of the adaptive spectral energy feature extraction sub-process of the present invention.

[0036] Figure 4 This is a flowchart of the model correction and update sub-process based on soot blowing efficiency evaluation of the present invention. Detailed Implementation

[0037] In the following description, only certain exemplary embodiments are briefly described. As those skilled in the art will recognize, the described embodiments can be modified in various ways without departing from the spirit or scope of the embodiments of the invention. Therefore, the drawings and description are considered to be exemplary in nature and not restrictive.

[0038] The following is in conjunction with the appendix Figures 1-4 The embodiments of the present invention will be described in detail below.

[0039] All data processing in this invention revolves around the physical signals generated by the inlet and outlet pressure measuring devices of the air preheater in thermal power units. The input to the support vector regression model is the time-series characteristics formed based on the air preheater differential pressure signal, and the output is a quantitative index of the degree of blockage defined according to a unified calibration method.

[0040] The core concept of this invention lies in constructing a dual-channel feature extraction pipeline that includes both absolute magnitude paths and fluctuation normalization paths. Simultaneously, it utilizes feature evolution information between adjacent windows to ensure that the characterization of blockage states does not lose absolute flow resistance information while capturing dynamic fluctuation structures related to ash accumulation morphology. Based on this, model correction is triggered by soot blowing events to achieve long-term alignment between the monitoring model and the physical state of the equipment, thereby forming a self-consistent data loop.

[0041] The overall process includes differential pressure signal acquisition, data quality detection, moving average denoising, dual-path differential pressure characterization, current data window and preceding data window division, time series feature extraction, adjacent window evolution feature construction, key time series feature screening, blockage quantitative mapping, early warning output, and soot blowing feedback model correction.

[0042] The differential pressure signal acquisition stage obtains time-synchronized pressure data from the pressure measurement points on the inlet and outlet sides of the air preheater, and generates a differential pressure signal between the inlet and outlet of the air preheater at each sampling moment.

[0043] In one implementation, raw differential pressure time-series data is continuously generated at a sampling frequency of 1 Hz. The current data window and the previous data window are each 10 minutes long, and the two windows are continuous in time.

[0044] The raw differential pressure time series data undergoes range detection, constant value detection, rate of change detection, and pressure measurement device status detection. Data quality flags are generated when anomalies are detected. When a backup pressure measurement device is configured, the system switches to the backup data source; if no backup data source is available, the system can maintain screen continuity for a limited time using statistics from the most recent valid sampling point. Windows containing anomalous substitute data do not form model correction samples.

[0045] The raw differential pressure time series data is processed by moving average of 5 sampling points. The filtered result directly forms the absolute differential pressure time series data.

[0046] Absolute differential pressure time series data are processed separately in the fluctuation processing path. Within each data window, the sampled value is subtracted from the window mean, and then scaled according to the window standard deviation to form fluctuation-normalized differential pressure time series data. The core purpose of fluctuation normalization is to remove the amplitude scaling effect caused by changes in operating conditions, so that the subsequently extracted fluctuation morphology and spectral energy characteristics only reflect the internal structure of differential pressure fluctuations.

[0047] Pressure differential level characteristics are derived from absolute pressure differential time series data, including window mean, window variance, peak value, and peak-to-peak value. Fluctuation pattern characteristics are derived from fluctuation-normalized pressure differential time series data, including zero-crossing rate, waveform factor, and impulse factor. Spectral energy characteristics are derived from fluctuation-normalized pressure differential time series data, including dominant frequency, band energy percentage, and power spectral density distribution.

[0048] After the current data window and the preceding data window complete the calculation of corresponding features, they correspond one by one to form the direction and amount of change. Pressure difference level features, fluctuation pattern features, spectral energy features, and adjacent window evolution features together form a candidate time series feature set.

[0049] During the historical training phase, a unified method for calibrating blockage levels was used to create blockage level calibration samples. Within the same model training cycle, labels with different physical meanings, such as the geometric blockage area ratio and the flow resistance growth ratio, were not mixed.

[0050] When using Pearson correlation analysis, candidate time-series features with an absolute correlation coefficient of 0.7 are considered key time-series features. The online application does not rely solely on Pearson correlation analysis for selection.

[0051] The support vector regression model is trained using historically calibrated samples. Regularization parameters, kernel parameters, and insensitive loss parameters are selected through cross-validation. After training, key temporal features, scale parameters, support vectors, support vector coefficients, biases, and model version are saved.

[0052] After inputting key time-series characteristics into the model, a quantitative indicator of congestion level is generated and written into the database and distributed control system.

[0053] In a field configuration, three levels of early warning are triggered when the quantitative indicators of congestion reach 50%, 70%, and 85%, respectively. For units with large load spans, the quantitative indicators of congestion corresponding to confirmed maintenance states are statistically analyzed within each operating state group. Without changing the quantitative model of congestion, the early warning threshold is determined based on the historical indicator range corresponding to the maintenance state boundary. During the online phase, the corresponding early warning threshold group is selected based on the current operating state.

[0054] Soot blowing feedback events are determined by the start and end information of soot blowing. No model calibration samples are formed during soot blowing. After soot blowing ends and the differential pressure signal stabilizes, a new continuous data window is formed, and new key timing features are generated using the same dual-path and adjacent window feature flow. The new key timing features and the unified blockage degree calibration value constitute the model calibration sample.

[0055] After the model is updated, validation is performed. If the validation error reaches the model accuracy condition, the update ends and a new model version is released; if the validation does not meet the requirements, the original online model is retained and model maintenance information is generated.

[0056] To facilitate a better understanding of the present invention by those skilled in the art, the present invention will be further described below with reference to specific embodiments.

[0057] Example 1: In this example, the inlet and outlet pressure measuring device of the air preheater of a thermal power unit continuously outputs pressure signals at a sampling frequency of 1Hz. The current data window and the previous data window are each 10 minutes.

[0058] The raw differential pressure time series data is averaged over 5 sampling points to form absolute differential pressure time series data. The mean, variance, peak value, and peak-to-peak value of the current data window and the previous data window are extracted from the absolute differential pressure time series data.

[0059] The absolute pressure difference time series data are separately converted into fluctuation-normalized pressure difference time series data, and the zero-crossing rate, waveform factor, impulse factor, main frequency, low frequency band energy ratio, mid frequency band energy ratio, high frequency band energy ratio and power spectral density characteristics are extracted from two windows respectively.

[0060] When the average pressure difference in the current data window is 1.84 kPa and the average pressure difference in the preceding data window is 1.76 kPa, the evolution of the average pressure difference is 0.08 kPa. When the low-frequency energy proportion in the current data window is 0.68 and the proportion in the preceding data window is 0.61, the evolution of the low-frequency energy is 0.07.

[0061] Candidate time-series features are selected from historically calibrated samples to form key time-series features, which are then input into a support vector regression model to form a quantitative indicator of the degree of congestion.

[0062] After a soot blowing feedback event occurs, no calibration samples are generated during the soot blowing disturbance phase. After soot blowing ends and the system stabilizes, model calibration samples are generated using the same process. Once the calibration samples are processed and the verification error meets the requirements, updates cease, and the updated model is used for the next online window.

[0063] Example 2: This example is a further optimization based on Example 1. In this example, data quality detection, model version management and distributed control system communication are added.

[0064] The system monitors differential pressure range, constant pressure status, rate of change, and pressure measurement device status. Data quality flags are generated when anomalies occur. The system switches to the backup channel when it is available; windows filled using historical data are not included in model calibration samples.

[0065] Each successful model calibration generates a new model version, saving the training time, calibration event, and validation error. The industrial processor sends quantitative indicators of congestion level, data quality markers, and model version number to the distributed control system.

[0066] Example 3: This example is based on the same principle as Example 1, except that it is applied to a three-compartment rotary air preheater equipped with an online soot blowing device.

[0067] The differential pressure signal comes from the pressure measuring devices at the inlet and outlet of the air preheater, without adding any internal visual or displacement scanning equipment to the air preheater.

[0068] Online monitoring employs a 1Hz sampling frequency, a 10-minute data window, and a moving average of 5 sampling points. Quantitative indicators of congestion severity are input into the distributed control system, and tiered early warning information is generated based on on-site warning thresholds.

[0069] After soot blowing is performed by operators or existing control logic, the soot blowing event and quantitative indicators of blockage degree are recorded in relation to time. The stability window after soot blowing forms a model calibration sample.

[0070] In a preferred embodiment, an advanced technical solution is disclosed that adaptively extracts spectral energy features based on first-principles fluid dynamics and implements model penalty boundary constraints based on physical quantification indicators of soot blowing efficiency. This embodiment solves the problem that fixed frequency band division cannot accurately capture the turbulent shedding frequency shift caused by flow velocity changes under variable load conditions of thermal power units, and at the same time eliminates the negative perturbation of the regression hyperplane of the support vector regression model by low-quality model correction samples introduced by invalid soot blowing operations.

[0071] The flow resistance of the flue gas inside the air preheater is generated by the interaction between the static ash accumulation morphology on the surface of the heat exchange elements and the dynamic boundary layer of the fluid. The fluctuation of the differential pressure signal on the time axis is not purely random white noise, but contains information on the karman vortex street shedding frequency generated when the airflow bypasses the ash accumulation characteristic structure. This vortex street shedding frequency is directly related to the flue gas velocity and the characteristic length of the ash accumulation structure. Therefore, by adaptively tracking the shift of this frequency, the changes in flow resistance caused by ash accumulation and simple load fluctuations can be distinguished more accurately.

[0072] Fixed frequency band division has limitations in actual industrial environments. When the unit is under high load, the flue gas velocity increases, and the vortex shedding frequency shifts to higher frequencies; when the unit is under low load, the shedding frequency falls back to the low-frequency range. Using static thresholds to divide the spectrum would cut off the frequency continuum corresponding to the physical phenomena. In this embodiment, the industrial processor dynamically calculates the flow field parameters based on the current operating state.

[0073] The system reads the current data window. It extracts the mean absolute pressure difference from the time-series data within the current data window. The mean absolute pressure difference directly maps to the macroscopic flue gas kinetic energy flowing through the air preheater. The industrial processor establishes a mapping between the mean pressure difference and flow velocity based on a variant of Bernoulli's equation, introduces the dimensionless Strouhal number, and calculates the adaptive cutoff frequency. The specific physical parameter conversion formulas are as follows:

[0074]

[0075] in, Indicates the adaptive cutoff frequency. This indicates the preset Strouhal number of the air preheater heat exchanger elements. Indicates the characteristic length of the heat exchange element. This represents the average absolute pressure difference in the time series data within the current data window. This indicates the density of the flue gas inside the air preheater. In this industrial setting, the characteristic length of the corrugated heat exchange element... The configuration is 0.008m, with a Strauhall number. The calibration value is 0.21. The corrugated geometry of the heat exchange element was obtained through wind tunnel testing. Take the corrugated plate height. Flue gas density. The data is obtained by interpolation from the air preheater inlet temperature and static pressure data synchronously issued by the distributed control system.

[0076] The spectral energy calculation module is triggered. The industrial processor performs a Fast Fourier Transform on the fluctuating normalized pressure difference time-series data within the current data window, generating the corresponding power spectral density distribution sequence. The energy integral on the discrete frequency axis is strictly defined by an adaptive truncation frequency. The integral formula for low-frequency band energy is:

[0077]

[0078] in, Indicates low-frequency energy. Indicates the adaptive cutoff frequency. The power spectral density function represents the fluctuation normalized pressure difference time series data. This represents the frequency variable. The high-frequency band energy is the difference between the full-band energy and the low-frequency band energy. After integration, the proportions of low-frequency band energy and high-frequency band energy are compressed into the candidate time-series feature set. The dynamic partitioning mechanism ensures that the spectral energy characteristics remain sensitive to changes in the micro-topological structure of ash accumulation within a wide load adjustment range of 300MW to 600MW. The low-frequency band energy mainly corresponds to the quasi-static pressure difference shift caused by large-scale ash accumulation structures, while the high-frequency band energy is related to micro-scale sediment disturbances and turbulent boundary layer instability. The ratio of the two constitutes a dimensionless feature that is highly sensitive to the ash accumulation state.

[0079] Support vector regression models face the risk of data poisoning during long-term operation. The soot blowing feedback events in thermal power units vary significantly in effectiveness due to steam pressure fluctuations or mechanical jamming of the soot blowing gun. If all post-soot blowing pressure differential states are indiscriminately labeled as "clean" and used as input samples for model calibration, the structural risk of the support vector regression model will diverge drastically.

[0080] Establish a quantitative soot blowing efficiency evaluation mechanism. The industrial processor retrieves the pressure difference data within the 30 minutes prior to the triggering of the soot blowing feedback event to form the average absolute pressure difference before soot blowing. After the soot blowing operation ends and the variance of the pressure difference signal converges to the steady-state threshold, the average absolute pressure difference after soot blowing is calculated for the data window immediately adjacent to the target value. The formula for calculating the soot blowing efficiency evaluation coefficient is as follows:

[0081]

[0082] in, This represents the soot blowing efficiency evaluation coefficient. This represents the average absolute pressure difference before soot blowing. This represents the average absolute pressure difference after soot blowing. This represents the baseline reference pressure differential of the air preheater under completely clean conditions. Using an ideal clean state as a reference, this coefficient quantifies the proportion of flow resistance actually removed by this soot blowing operation to the maximum possible removal volume, thus objectively reflecting the physical effectiveness of the soot blowing action. Baseline Reference Pressure Differential The data was obtained by calling the baseline data array under the same load when the unit was restarted without ash after maintenance.

[0083] The sample weighting factor is generated based on the evaluation coefficient. The closer the evaluation coefficient is to 1, the closer the soot blowing operation has physically restored the heat exchange element to an ideal clean state. The corresponding model calibration samples should dominate in subsequent model updates. The nonlinear mapping formula for the sample weighting factor is:

[0084]

[0085] in, Indicates the first The sample weighting factor for each model correction sample. This indicates the operation of the natural exponent. This represents the preset attenuation control constant. The value range is typically 1.5 to 3.5, determined based on the characteristics of the sootblowing equipment and historical sootblowing effect statistics. For this site environment, a value of 2.5 is preferred. Indicates extracting the first The corresponding soot blowing efficiency evaluation coefficient when calibrating samples of a model. When smaller, Rapid decay significantly suppresses the contribution of inefficient dust samples to model updates.

[0086] The update algorithm reconstructs parameters in memory. Introducing a sample weighting factor alters the quadratic programming boundary of the dual problem in the support vector regression model. The optimization objective function is subject to the following constraints:

[0087]

[0088] in, and This represents the Lagrange multiplier in the dual problem of the support vector regression model. This represents the global regularization parameter. Indicates the first The sample weighting factor for each model calibration sample. The constraints directly prevent the poor-quality soot blowing samples from pulling on the regression hyperplane. When a mechanical failure of the soot blowing gun causes... When the value is as low as 0.1, the corresponding upper limit of the Lagrange multiplier is... Approaching zero. At the mathematical level, the algorithm automatically removes the "pseudo-state" labels caused by invalid data cleaning, ensuring the purity of the support vector distribution in the feature space.

[0089] Continuous industrial operation data verified the feasibility and quantitative output accuracy of this scheme. The core logic contained in this implementation method was deployed on a 600MW supercritical unit, and a 90-day grid connection test was conducted. During this period, the unit load frequently participated in deep grid peak shaving between 300MW and 600MW, and the air preheater performed 142 online soot blowing operations.

[0090] Table 1. Comparison of Adaptive Feature Tracking under Variable Load Conditions;

[0091] Unit load (MW) Mean absolute pressure difference (kPa) Theoretical eddy current frequency (Hz) This implementation features an adaptive cutoff frequency (Hz). Traditional fixed cutoff frequency (Hz) High-frequency energy ratio error 315 0.82 4.12 4.08 6.50 38.5% 450 1.35 5.34 5.25 6.50 22.1% 600 1.88 6.27 6.18 6.50 8.4%

[0092] Table 1 records the evolution trajectory of the dynamic parameters of the flow field. As the average absolute pressure difference of the load increases, the adaptive cutoff frequency generated in this embodiment closely follows the drift trajectory of the theoretical eddy frequency. The fixed cutoff frequency incorrectly assigns a large number of mid-frequency turbulence components to the low-frequency band in the low-load section, resulting in a distortion of energy proportion characteristics as high as 38.5%. The adaptive method converges the characteristic calculation error of all load sections to within 3%, ensuring that the spectral energy characteristics of the input support vector regression model reflect the true ash accumulation morphology rather than flow velocity interference.

[0093] The model's disturbance rejection performance was also quantitatively demonstrated. Three soot blowing events with different efficiencies were extracted and analyzed in isolation during the test period. Event A was standard soot blowing with normal steam parameters. Event B was accompanied by low steam pressure. Event C was a simulated jamming of the soot blowing gun's propulsion mechanism.

[0094] Table 2. Data table of soot blowing efficiency evaluation and model weight allocation;

[0095] Event Number Pressure difference before soot blowing (kPa) Pressure difference after soot blowing (kPa) Cleaning baseline pressure difference (kPa) Soot blowing efficiency evaluation coefficient Sample weighting factor Event A 1.65 1.12 1.05 0.883 0.746 Event B 1.70 1.45 1.05 0.384 0.214 Event C 1.68 1.61 1.05 0.111 0.108

[0096] Table 2 reflects the calculation results of the Lagrange multiplier constraint boundary. Event C produced almost no physical cleaning effect, and its sample weighting factor dropped to 0.108. Conventional methods without performance constraints forcibly fitted it as a clean sample, resulting in a 15% overall downward drift in the quantitative index of online congestion level in the following 24 hours, creating a serious risk of delayed reporting. This implementation directly marginalizes the training data corresponding to event C in the quadratic programming solution through a weighting mechanism, and no abnormal diffusion of variance fluctuations in the model output index is observed.

[0097] Table 3. Statistical table of the mean absolute error of quantitative monitoring of air preheater blockage in continuous cycle;

[0098] Statistical period Absolute value of error in Example 1 The absolute value of the error in this implementation scheme Error reduction First operating month 4.12% 2.85% 1.27% Second operating month 5.35% 3.01% 2.34% Third operating month 7.82% 3.12% 4.70%

[0099] Long-term data validates the core capability of this scheme in controlling model degradation. Table 3 shows that in the third month of operation, the error defense line of Basic Implementation Example 1 gradually collapsed to 7.82% due to the continuous accumulation of low-quality historical dust samples. In this implementation, by strictly eliminating invalid calibration samples in the state phase space and deeply binding spectral energy characteristics with first principles of fluid mechanics, the mean absolute error of the output was stably clamped at 3.12% after three months. Data cleaning based on physical constraints and the adaptive evolution mechanism jointly maintained the long-term alignment between the digital and physical spaces. Quantitative monitoring indicators were output stably. The distributed control system smoothly executed multi-level early warnings based on accurate indicator curves, guiding the on-site maintenance team to complete precise equipment maintenance.

[0100] Although preferred embodiments of the invention have been described, those skilled in the art, upon learning the basic inventive concept, can make other changes and modifications to these embodiments. Therefore, the appended claims are intended to be interpreted as including both the preferred embodiments and all changes and modifications falling within the scope of the invention.

[0101] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. It should be noted that any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A quantitative monitoring method for air preheater blockage based on differential pressure time-series characteristics, characterized in that, include: Collect the differential pressure signals at the inlet and outlet of the air preheater to form raw differential pressure time-series data; The original differential pressure time series data is denoised by moving average, and the denoised absolute differential pressure time series data is retained. The fluctuation-normalized differential pressure time series data is then formed from the absolute differential pressure time series data. In the current data window and the preceding data window, pressure difference level features are extracted from the absolute pressure difference time series data, and fluctuation morphology features and spectral energy features are extracted from the fluctuation normalized pressure difference time series data. Based on the changes of the corresponding features between the current data window and the preceding data window, adjacent window evolution features are formed. The characteristics of differential pressure level, fluctuation pattern, spectral energy, and adjacent window evolution are combined into a candidate time series feature set, and key time series features are selected from the candidate time series feature set. Key time-series features are input into a support vector regression model trained based on congestion level calibration samples to obtain a quantitative indicator of congestion level. After the dust removal feedback event, model calibration samples are formed according to the same feature construction process. The model calibration samples are used to update the support vector regression model, and the updated support vector regression model is used for subsequent data window congestion calculation.

2. The method for quantitative monitoring of air preheater blockage based on differential pressure time-series characteristics according to claim 1, characterized in that: The air preheater inlet and outlet pressure difference signals are sampled at a frequency of 1Hz to form the original pressure difference time series data. The window length of the current data window and the previous data window is 10 minutes. The current data window and the previous data window are continuous in time.

3. The method for quantitative monitoring of air preheater blockage based on differential pressure time-series characteristics according to claim 1, characterized in that: Moving average denoising uses a sliding window of 5 sampling points; the absolute pressure difference time series data directly retains the filtered pressure difference magnitude; the fluctuation-normalized pressure difference time series data is obtained by subtracting the mean of the corresponding data window from the absolute pressure difference time series data and then scaling it according to the standard deviation of the corresponding data window.

4. The method for quantitative monitoring of air preheater blockage based on differential pressure time-series characteristics according to claim 1, characterized in that: Pressure differential level characteristics include at least two of the following: mean, variance, peak value, and peak-to-peak value; The fluctuation morphology features include at least one of zero-crossing rate, waveform factor, and impulse factor; the spectral energy features include at least one of dominant frequency, preset frequency band energy proportion, and power spectral density distribution, where the dominant frequency is the frequency corresponding to the peak power spectral density; the process of extracting spectral energy features includes: obtaining the mean absolute pressure difference of the absolute pressure difference time series data within the current data window; calculating the adaptive cutoff frequency based on the mean absolute pressure difference, the preset Strouhal number of the air preheater heat exchange element, the characteristic length of the heat exchange element, and the flue gas density inside the air preheater, wherein the adaptive cutoff frequency is the square root of the quotient of the Strouhal number divided by the characteristic length and then multiplied by twice the mean absolute pressure difference divided by the flue gas density; dividing the integral interval of the power spectral density function into a low-frequency band and a high-frequency band based on the adaptive cutoff frequency; and extracting the discrete spectral energy proportion of the fluctuation normalized pressure difference time series data in the low-frequency band and the high-frequency band as the spectral energy features.

5. The method for quantitative monitoring of air preheater blockage based on differential pressure time-series characteristics according to claim 1, characterized in that: The evolution characteristics of adjacent windows are formed by comparing the pressure difference level characteristics, fluctuation pattern characteristics, or spectral energy characteristics of the current data window with those of the previous data window that have the same name and physical meaning. The evolution characteristics of adjacent windows include at least the direction and amount of change of the corresponding characteristics.

6. The method for quantitative monitoring of air preheater blockage based on differential pressure time-series characteristics according to claim 1, characterized in that: The selection of key time series features from the candidate time series feature set includes: calculating the Pearson correlation coefficient between the candidate time series features and the blockage degree calibration value based on historical blockage degree calibration samples, and identifying candidate time series features with an absolute value of correlation coefficient greater than 0.7 as key time series features.

7. The method for quantitative monitoring of air preheater blockage based on differential pressure time-series characteristics according to claim 1, characterized in that: The support vector regression model uses a radial basis function kernel. The regularization parameter, kernel parameter, and insensitive loss parameter are determined by the cross-validation results of the congestion level calibration samples. The congestion level calibration samples adopt a uniform congestion level calibration method within the same model training period.

8. The method for quantitative monitoring of air preheater blockage based on differential pressure time-series characteristics according to claim 1, characterized in that: Before forming key timing features, the differential pressure signals at the inlet and outlet of the air preheater are tested for data validity. When data that exceeds the range, remains constant for a long time, or has a rate of change that exceeds the allowable range of the equipment is detected, a data quality mark is generated, and continuous monitoring is completed using backup differential pressure sensor data or historical valid data. Data windows containing abnormal substitute data are not used as model calibration samples.

9. The method for quantitative monitoring of air preheater blockage based on differential pressure time-series characteristics according to claim 1, characterized in that: After the soot blowing feedback event ends and the system enters a stable operating state, the time series data of the pressure difference after soot blowing is collected. The model calibration sample is formed by following the same process as online blockage monitoring, including noise reduction, dual-path data construction, window feature extraction and key time series feature screening. The process of updating the support vector regression model using model correction samples includes: extracting the mean absolute pressure difference before soot blowing during the stable operation phase before the soot blowing feedback event is triggered; Extract the mean absolute pressure difference after soot blowing in the stable operating state after the soot blowing feedback event; calculate the soot blowing efficiency evaluation coefficient, which is equal to the difference between the mean absolute pressure difference before soot blowing and the mean absolute pressure difference after soot blowing divided by the difference between the mean absolute pressure difference before soot blowing and the benchmark reference pressure difference in the completely clean state; calculate the sample weighting factor of the corresponding model correction sample based on the soot blowing efficiency evaluation coefficient, and use the constraint condition with the sample weighting factor as the upper limit of the penalty coefficient to solve the dual problem of the support vector regression model to complete the model update; the update of the support vector regression model ends when the correction sample processing is completed and the model validation error meets the preset accuracy condition.

10. The method for quantitative monitoring of air preheater blockage based on differential pressure time-series characteristics according to claim 1, characterized in that: The raw differential pressure time series data, key time series characteristics, quantitative indicators of blockage degree, data quality markers, soot blowing feedback events and early warning information are written into the air preheater blockage monitoring database according to the timestamp.

Images