Dynamic Control Method for Electrostatic Precipitator Variable Frequency Power Supply Based on AI Recognition of Flue Gas Dust Features

By using an AI-based identification method based on flue gas dust characteristics, combined with a multi-parameter coupled logic decision matrix and dual hysteresis dead zone control, the frequency converter power supply control of the electrostatic precipitator was optimized, solving the problems of delay lag and false alarms, and achieving stable and efficient power supply control.

CN123102502APending Publication Date: 2026-11-17HUANENG CHAOHU POWER GENERATION CO LTD
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202611247101.8
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2026-08-17
Publication Date
2026-11-17

AI Technical Summary

Technical Problem

Existing frequency converter power supply control systems for electrostatic precipitators rely on a single electrical feedback parameter, leading to time delays, false alarms, and hardware damage, and lacking an effective transition protection mechanism.

Method used

An AI recognition method based on flue gas dust characteristics is adopted. By acquiring two-dimensional microscopic image frames, the dust morphology irregularity index is calculated. Combined with a multi-parameter coupled logic decision matrix, the synchronous matching of visual feedforward and electrical detection is achieved. Furthermore, a dual hysteresis dead zone control to prevent tripping oscillation is introduced to optimize the power control strategy.

Benefits of technology

It solves the problem of time delay, improves the physical rigor of dust status recognition, reduces false alarms, prevents hardware damage, and achieves stable dynamic control of the power supply.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN123102502A_ABST
    Figure CN123102502A_ABST
Patent Text Reader

Abstract

This invention relates to the field of electrostatic precipitator (ESP) control technology and discloses a dynamic control method for ESP variable frequency power supplies based on AI-based identification of flue gas dust characteristics. The method includes: using an AI edge detection model to read two-dimensional original microscopic image frames and calculate the dust morphology irregularity index; selecting a time window for the generation probability of the first-order electric field to calculate the equivalent morphology features; injecting detection pulses to calculate the transient differential impedance fluctuation rate; inputting the equivalent morphology features and the transient differential impedance fluctuation rate into a multi-parameter coupled logic decision matrix; and combining the high-limit threshold of the carbon-rich low-resistivity state of dust, the low-limit threshold of the pure ash high-resistivity state, and the differential impedance fluctuation rate threshold to output mode switching commands, thereby changing the output electrical topology mode of the high-frequency variable frequency power supply unit and the pulse superposition unit. This invention achieves synchronous matching of visual feedforward and electrical detection through cross-comparison to eliminate abnormal false alarms, and introduces dual hysteresis dead-zone control constraints to prevent hardware damage caused by frequent switching.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This invention relates to the field of electrostatic precipitator control technology, specifically to a dynamic control method for electrostatic precipitator variable frequency power supply based on AI recognition of flue gas dust characteristics. Background Technology

[0002] Electrostatic precipitators (ESPs) are core equipment for industrial flue gas purification. The control strategy of their variable frequency power supply directly affects dust removal efficiency and system operational stability. Current ESP ESP variable frequency power supply control systems rely on passive adjustments using post-processed unidirectional electrical feedback parameters such as spark frequency or secondary voltage and current. This feedback mechanism suffers from time lag; the control system cannot acquire the morphological characteristics of the dust-laden airflow before it enters the electric field. This results in the variable frequency power supply only responding and adjusting after back corona discharge or frequent flashovers have already occurred within the electric field, making it difficult to achieve advance prediction and dynamic control.

[0003] Existing control logic relies on a single dimension for judgment, generally depending solely on the exceeding of limits by local electrical parameters to determine the high or low resistance properties of dust. In practical industrial applications, sudden changes in local concentration or airflow disturbances in the flue cross-section flow field can easily cause instantaneous fluctuations in a single electrical parameter. Due to the lack of cross-validation between spatial and visual dimensions, feedback from a single parameter is prone to triggering false alarms or issuing incorrect control commands, resulting in insufficient physical rigor in the system's identification of the actual dust phase characteristics.

[0004] Conventional variable frequency power supply control strategies lack effective transition protection mechanisms when setting state switching logic. When the electric field operating parameters fluctuate slightly around the set alarm threshold, the system's judgment logic will frequently jump between different operating states. This repeated switching output command at the critical state boundary will cause the high-frequency variable frequency power supply unit and pulse superposition unit to generate high-frequency disordered switching actions. Long-term control oscillations not only affect the continuity of power supply but also accelerate the fatigue aging of core electrical components, increasing the risk of hardware damage. Summary of the Invention

[0005] To address the shortcomings of existing technologies, this invention provides a dynamic control method for electrostatic precipitator frequency converter power supplies based on AI recognition of flue gas dust characteristics. This method solves the problems of time lag caused by the reliance on post-redirect unidirectional electrical feedback in existing frequency converter power supply control strategies, susceptibility to false alarms caused by sudden changes in local concentration in the flow field due to single-dimensional judgment, and hardware damage caused by frequent switching between the high-frequency frequency converter power supply unit and the pulse superposition unit at the critical boundary of the operating condition due to the lack of a transition protection mechanism.

[0006] To achieve the above objectives, the present invention provides the following technical solution: This invention provides a dynamic control method for electrostatic precipitator variable frequency power supply based on AI recognition of flue gas dust characteristics, comprising the following steps: Historical parameters are obtained for benchmark calibration, and the generated upper limit threshold of dust rich carbon low resistivity state, lower limit threshold of pure ash high resistivity state, and differential impedance fluctuation rate threshold are written into the storage medium. Two-dimensional raw microscopic image frames are acquired, and the controller uses the deployed AI edge detection model to read the two-dimensional raw microscopic image frames and calculate the dust morphology irregularity index; Selected A level electric field is generated to form a probability arrival time window. When the probability arrival time windows of adjacent batches are independent of each other, the dust morphology irregularity index is used as the equivalent morphology feature. When the probability arrival time windows of adjacent batches overlap in the time domain, the equivalent morphology feature is calculated based on the dust morphology irregularity index. Based on the equivalent morphological features, reconstruction parameters are generated. When the time progresses to the probability arrival time window, the parameters are then sent to the first... A detection pulse is injected into the electric field to simultaneously acquire transient voltage and current signals and calculate the transient differential impedance fluctuation rate. The equivalent morphological features and the transient differential impedance fluctuation rate are input into a multi-parameter coupled logic decision matrix. Combined with the dust carbon-rich low-resistivity high limit threshold, the pure ash high-resistivity low limit threshold and the differential impedance fluctuation rate threshold output mode switching command, the output electrical topology mode of the high-frequency variable power supply unit and the pulse superposition unit is changed.

[0007] Preferably, the step of acquiring historical parameters for benchmark calibration, and writing the generated upper limit threshold of the dust-rich carbon-rich low-resistivity state, the lower limit threshold of the pure ash high-resistivity state, and the differential impedance fluctuation rate threshold into the storage medium includes: Historical dust collector ash hopper sampling and testing data were obtained to classify and label dust phases, and a multiple linear regression model was constructed to perform control parameter benchmark calibration and output aspect ratio term weights and edge gradient term weights. The trained multiple linear regression model is used to perform inference calculations on carbon-rich low-resistivity samples and pure ash high-resistivity samples, and the values ​​corresponding to the confidence intervals are extracted to adjust the upper limit threshold of the carbon-rich low-resistivity dust and the lower limit threshold of the pure ash high-resistivity dust. Historical voltage and current characteristic data are read to extract the critical impedance reference parameters of the anti-corona, the basic impedance fluctuation rate under the critical state of anti-corona is calculated, and a safety margin coefficient is introduced for derating correction to generate the differential impedance fluctuation rate threshold.

[0008] Preferably, the step of obtaining historical dust collector ash hopper sampling and testing data for dust phase classification and labeling includes: Extract the resistivity and loss on ignition values ​​for the same time period to construct a dust phase feature vector; When the loss on ignition value in the dust phase feature vector is greater than the loss on ignition threshold and the resistivity value is lower than the resistivity critical value, the corresponding historical dust collector ash hopper sampling and testing data is marked as the carbon-rich low-resistivity sample and the phase fitting target label value is set to 1. When the loss on ignition value in the dust phase characteristic vector is lower than the loss on ignition threshold and the resistivity value is higher than the resistivity critical value, the corresponding historical dust collector ash hopper sampling and testing data is marked as the pure ash high resistivity state sample and the phase fitting target label value is set to -1.

[0009] Preferably, the step of acquiring the two-dimensional raw microscopic image frame, wherein the controller uses a deployed AI edge detection model to read the two-dimensional raw microscopic image frame and calculate the dust morphology irregularity index, includes: The heat-traced sampling module introduces the dust airflow in the main flue into the optical measurement cavity at a constant speed, and the stroboscopic light source and the microscopic imaging component acquire the original two-dimensional microscopic image frame; The controller uses the AI ​​edge detection model to read the two-dimensional original microscopic image frame, converts the two-dimensional original microscopic image frame into a single-channel feature tensor matrix, inputs it into the AI ​​edge detection model, and outputs a pixel-level edge probability map. The adaptive truncation threshold of the pixel-level edge probability map is determined and converted into a binary contour map. All closed regions are extracted as candidate particles. Abnormal particles that are out of focus or overlapped are removed based on the ratio of the mean pixel gradient of the edge coordinate set to the convex hull area. The effective individual particles are retained and the step of calculating the dust morphology irregularity index is performed.

[0010] Preferably, calculating the dust morphology irregularity index using the effective monomer particles includes: The temporal statistical variance of the real-time aspect ratio of all effective individual particles within the visual sampling time window is calculated. Calculate the gradient magnitude of all effective single-unit particle edge pixels in the original grayscale image within the visual sampling time window and obtain the overall edge pixel gradient mean; The dust morphology irregularity index is calculated and output by mathematically fusing the temporal statistical variance and the mean gradient of the overall edge pixels with the aspect ratio term weight, the edge gradient term weight and the fitting bias constant using a linear model.

[0011] Preferably, the selected first The electric field is leveled and a probability arrival time window is generated, including: Based on the average flow velocity of the main flue section, the dust airflow reaches the first... The theoretical flight time at the entrance of the electric field; The time difference between the completion of recording the exposure of the microscopic image frame and the output of the dust morphology irregularity index is set as the inherent processing delay of the system. Subtracting the theoretical flight time from the inherent processing delay of the system generates a value for the first... The control time margin of the electric field is determined by selecting the minimum electric field level index corresponding to a control time margin greater than or equal to the lower threshold of the response time margin, and the corresponding independent electric field is selected as the first electric field. The electric field is used to generate the probability arrival time window based on the theoretical flight time.

[0012] Preferably, the step of calculating the equivalent morphological features based on the dust morphology irregularity index when the probability arrival time windows of adjacent batches overlap in the temporal domain includes: The product of the fluid diffusion coefficient and the theoretical flight time is used to generate the fluid diffusion tolerance and establish the boundary of the probability arrival time window; If the start time of the probability arrival time window of the current batch is less than the end time of the probability arrival time window of the previous batch, it is determined that the two probability arrival time windows overlap in the time domain. During the overlapping period, the dust morphology irregularity index of all independent batches involved in the overlap is aggregated and the average state is extracted as the equivalent morphology feature. The statistical fluctuation degree of deviation from the equivalent morphology feature is calculated to generate the equivalent local variance.

[0013] Preferably, reconstruction parameters are generated based on the equivalent morphological features and applied to the first... A level electric field is injected with a probe pulse to simultaneously acquire transient voltage and current signals and calculate the transient differential impedance fluctuation rate, including: The reconstruction parameters are generated by packaging the probe pulse width and probe pulse injection frequency together based on the equivalent morphological features; The pulse superposition unit directs the pulse to the first according to the reconstruction parameters. The detection pulse is injected into the electric field, and the isolation sampling module simultaneously acquires the transient voltage signal and the transient current signal; Extract the transient voltage signal change and the transient current signal change, calculate the current change rate between two adjacent sampling points, and when the current change rate is less than the current change rate safety threshold, introduce a perturbation constant to prevent the denominator from approaching zero. Combine the transient current signal change and the perturbation constant to prevent the denominator from approaching zero as the denominator, and perform a division operation between the transient voltage signal change and the denominator to obtain the transient differential impedance to prevent the denominator from approaching zero. Extract the maximum and minimum absolute values ​​of impedance within the duration of the probe pulse, and perform a normalization algorithm to obtain the transient differential impedance fluctuation rate constrained within the numerical range boundary.

[0014] Preferably, the equivalent morphological features and the transient differential impedance fluctuation rate are input into a multi-parameter coupled logic decision matrix, and combined with the dust carbon-rich low-resistivity state high limit threshold, the pure ash high-resistivity state low limit threshold, and the differential impedance fluctuation rate threshold output mode switching command, the output electrical topology mode of the high-frequency inverter power supply unit and the pulse superposition unit is changed, including: A two-dimensional state space is established by using the equivalent morphological features as the first input dimension and the transient differential impedance fluctuation rate as the second input dimension. In the two-dimensional state space, when the equivalent morphological feature is greater than or equal to the upper limit threshold of the dust-rich carbon low-resistivity state and the transient differential impedance fluctuation rate is greater than or equal to the differential impedance fluctuation rate threshold, it is determined to be in a low-resistivity risk state. A low-level blocking signal is sent to shield the pulse superposition unit from working and change the operating frequency of the high-frequency inverter power supply unit to maintain the safe upper limit operating frequency and switch to pure DC power supply topology mode. In the two-dimensional state space, when the equivalent morphological feature is less than or equal to the low threshold of the pure gray high-resistivity state and the transient differential impedance fluctuation rate is less than or equal to the preset fluctuation low threshold, it is determined to be in a high-resistivity state. The base DC voltage output by the high-frequency inverter power supply unit is reduced to the corona initiation voltage critical point, the shielding of the pulse superposition unit is removed, and the transient high-voltage pulse is superimposed according to the output pause time interval including the empirical charge discharge time constant and margin time of the high resistivity dust.

[0015] Preferably, a dual hysteresis dead-zone control constraint to prevent tripping oscillations is introduced into the multi-parameter coupled logic decision matrix, including: Based on the high limit threshold of the dust-rich carbon low-resistivity state, the low limit threshold of the pure ash high-resistivity state, the differential impedance fluctuation rate threshold, and the preset fluctuation low limit threshold, a morphology hysteresis recovery threshold and a fluctuation hysteresis recovery threshold corresponding to exiting the low-resistivity risk state or the high-resistivity state are generated by bias operation to construct a cross-state deviation difference. When crossing the cross-state deviation difference, the first hysteresis dead zone condition is satisfied. The duration for which the equivalent topographic feature and the transient differential impedance fluctuation rate are continuously maintained at the effective side of the topographic hysteresis recovery threshold and the fluctuation hysteresis recovery threshold by starting the anti-jitter timer is defined as the state over-limit duration. The duration of the state exceeding the limit is compared with the time-based stabilization threshold in real time. When the duration of the state exceeding the limit is greater than or equal to the time-based stabilization threshold, the second hysteresis dead zone condition is met.

[0016] Preferably, the execution of state recovery logic including a double hysteresis dead zone includes: When the original low-impedance risk state is met, the shielding of the pulse superposition unit is removed after the first hysteresis dead zone condition and the second hysteresis dead zone condition are met simultaneously, allowing the transient high voltage pulse to be output again, and the operating frequency of the high-frequency inverter power supply unit is restored to the preset operating frequency reference value. When the original state is in the high-impedance state, after the equivalent morphological feature is greater than the morphological hysteresis recovery threshold for exiting the high-impedance state, the transient differential impedance fluctuation rate is greater than the fluctuation hysteresis recovery threshold for exiting the high-impedance state, and the state over-limit duration reaches the time duration anti-jitter threshold, it is determined that the high-impedance state has been exited. The base DC voltage output by the high-frequency inverter power supply unit is restored to the preset base DC voltage normal value, and the pulse superposition unit is driven according to the reconstruction parameters.

[0017] This invention provides a dynamic control method for electrostatic precipitator variable frequency power supplies based on AI recognition of flue gas dust characteristics, which has the following beneficial effects: 1. This invention utilizes a deployed AI edge detection model to read two-dimensional raw microscopic image frames and generate a dust morphology irregularity index. It combines this index with flow field parameters to generate a probability arrival time window and calculates equivalent morphology features upon overlap. When the target airflow passes through, a detection pulse is injected according to the reconstructed parameters, achieving synchronous matching between visual feedforward prediction and electrical detection feedback on the time axis, thus solving the delay lag problem inherent in post-implemented unidirectional electrical feedback.

[0018] 2. This invention constructs a multi-parameter coupled logic decision matrix, using equivalent morphological features and transient differential impedance fluctuation rate as input dimensions of a two-dimensional state space. The dual-dimensional values ​​are cross-compared with the high-limit threshold for dust-rich carbon-rich low-resistivity states, the low-limit threshold for pure ash high-resistivity states, and the differential impedance fluctuation rate threshold to output switching commands. This eliminates false alarms caused by sudden changes in flow field concentration and improves the physical rigor of identifying low-resistivity risk states and high-resistivity states.

[0019] 3. This invention introduces a dual hysteresis dead-zone control constraint to prevent tripping oscillations into the multi-parameter coupled logic decision matrix. A hysteresis recovery threshold is generated through bias calculation, and an anti-jitter timer is introduced to compare the duration of the out-of-limit state with the time-based anti-jitter threshold. The state recovery logic is executed only after both the first and second hysteresis dead-zone conditions are met, preventing repeated output of switching commands at critical state boundaries and avoiding hardware damage to the high-frequency inverter power supply unit and pulse superposition unit due to frequent switching. Attached Figure Description

[0020] Figure 1 This is a schematic diagram of the structure of a dynamic control system for an electrostatic precipitator variable frequency power supply based on AI recognition of flue gas dust features according to the present invention. Figure 2The flowchart shows a dynamic control method for an electrostatic precipitator variable frequency power supply based on AI recognition of flue gas dust features according to the present invention. Figure 3 This is a flowchart illustrating the benchmark calibration process for obtaining historical parameters in this invention. Figure 4 This is a flowchart of the in-situ sampling and dust morphology feature extraction process of the present invention; Figure 5 This is a flowchart of the equivalent calculation of the target electric field dynamic redistribution and time window characteristics in this invention; Figure 6 This is a flowchart of the synchronous detection of probe pulse reconstruction and transient impedance in this invention; Figure 7 This is a flowchart of the multi-parameter logic decision-making and power topology mode switching of the present invention; Figure 8 The graph shows the dynamic adjustment relationship between the detection pulse width and the equivalent morphological features of this invention. Figure 9 This is a comparison chart of transient differential impedance fluctuation rates under different control methods of the present invention. Detailed Implementation

[0021] The technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0022] See attached document Figure 1 The present invention provides a dynamic control system for electrostatic precipitator variable frequency power supply based on AI recognition of flue gas dust characteristics, including a heat tracing sampling module, an optical measurement cavity, an independent electric field, a composite power supply module, and an isolated sampling module.

[0023] An isokinetic sampling probe is installed on the flue wall at the inlet of the electrostatic precipitator. A heated sampling module is connected between the isokinetic sampling probe and the bypass optical measurement cavity. The heated sampling module acquires the actual flue gas temperature at the measuring point in the main flue and simultaneously adjusts the wall temperature of the isokinetic sampling probe and the optical measurement cavity. The optical measurement cavity is equipped with a stroboscopic light source and a microscopic imaging component.

[0024] The electrostatic precipitator contains, sequentially distributed along the flue gas flow direction, the following components: Level independent electric field, For a positive integer greater than 1, the first... Level electric field representative Any level in an independent electric field, The value range is 1 to .

[0025] Each independent electric field stage has a high-voltage input terminal connected to a composite power supply module, which consists of a high-frequency inverter power supply unit and a pulse superposition unit connected in parallel. The high-frequency inverter power supply unit is used to supply power to the first stage... The electric field outputs the basic DC voltage; the pulse superposition unit is connected to the high-voltage output side of the high-frequency inverter power supply unit to superimpose transient high-voltage pulses on the basic DC voltage.

[0026] Each independent electric field stage has an isolation sampling module connected in series in its high-voltage power supply circuit. This isolation sampling module includes a resistive-capacitive voltage divider and a high-frequency transformer, used for synchronously acquiring data from the first stage. The transient voltage and transient current signals of the electric field during pulse injection.

[0027] See attached document Figure 2 This invention provides a dynamic control method for electrostatic precipitator frequency converter power supply based on AI recognition of flue gas dust characteristics. This method, implemented based on the aforementioned system, includes the following steps: S100, acquire historical parameters for benchmark calibration. The controller reads historical dust collector ash hopper sampling and testing data and historical voltage and current characteristic data from the unit control system, constructs a multiple linear regression model to perform control parameter benchmark calibration, and writes the generated aspect ratio term weight, edge gradient term weight, dust carbon-rich low-resistivity state upper limit threshold, pure ash high-resistivity state lower limit threshold, and differential impedance fluctuation rate threshold into the storage medium; S200 performs in-situ sampling and dust morphology feature extraction. The heat-traced sampling module introduces the dust airflow in the main flue into the optical measurement cavity at a constant velocity. The stroboscopic light source and the microscopic imaging component acquire two-dimensional raw microscopic image frames at a set frame rate. The controller uses an AI edge detection model to read the two-dimensional raw microscopic image frames and calculate the dust morphology irregularity index. S300 performs dynamic redistribution of the target electric field and equivalent calculation of time window characteristics. The controller, combining the average flow velocity of the main flue section and the inherent processing delay of the system, executes a target electric field traversal optimization rule to select the first [target electric field] that meets the lower threshold of the response time margin. The electric field generates a probability arrival time window, and when there is temporal overlap between the probability arrival time windows of adjacent batches, the equivalent morphological features and equivalent local variance are calculated. S400 performs synchronous detection of probe pulse reconstruction and transient impedance. The controller reconstructs the probe pulse width and probe pulse injection frequency based on equivalent topographic features, and the pulse superposition unit moves to the next pulse according to the reconstruction parameters. The first-stage electric field injection probe pulse is used to synchronously acquire the second-stage electric field. The controller receives the transient voltage and current signals of the electric field and calculates the transient differential impedance fluctuation rate. The S500 performs multi-parameter logic decision-making and power supply topology mode switching. The controller inputs the equivalent topology characteristics and transient differential impedance fluctuation rate into the multi-parameter coupled logic decision matrix, outputs mode switching commands, and changes the output electrical topology mode of the high-frequency inverter power supply unit and the pulse superposition unit according to the mode switching commands, thereby realizing dynamic frequency conversion adjustment of the control target.

[0028] To enable those skilled in the art to better understand the present invention, the following will elaborate on the technical details and specific implementation logic of the above-mentioned system operation and method execution process.

[0029] See attached document Figure 3 The controller performs preprocessing of historical operating data and dust phase classification and labeling, including the following steps: S101, Read historical dust collector ash hopper sampling and testing data to classify and label dust phases. The controller reads historical dust collector ash hopper sampling and testing data from the unit control system. This data mainly includes the resistivity and loss on ignition values ​​of historical dust samples. The resistivity value characterizes the conductivity of the dust layer on the electrode surface, and the loss on ignition value characterizes the mass fraction of unburned carbon particles in the dust.

[0030] The controller extracts resistivity and loss on ignition values ​​from the same time period to construct a dust phase characteristic vector. The controller internally presets a loss on ignition threshold and a resistivity critical value. Based on the typical operating characteristics of electrostatic precipitators in coal-fired power plants, the loss on ignition threshold is set to a range of 5% to 8%, and the resistivity critical value is set to a range of 10%. 10 Ω·cm to 10 11 Ω·cm.

[0031] When the loss on ignition value in the dust phase characteristic vector is greater than the loss on ignition threshold and the resistivity value is lower than the resistivity critical value, the controller marks the historical dust collector ash hopper sampling and testing data as a carbon-rich, low-resistivity sample and sets the phase fitting target label value of this carbon-rich, low-resistivity sample to 1. Because the dust contains a high amount of carbon particles, the carbon-rich, low-resistivity sample is prone to rapid charge dissipation on the electrode plates, causing the dust to lose its electrostatic adsorption and return to the flue gas flow field, triggering secondary dust re-entrainment.

[0032] When the loss on ignition value in the dust phase characteristic vector is lower than the loss on ignition threshold and the resistivity value is higher than the resistivity critical value, the controller marks the historical dust collector ash hopper sampling and testing data as pure ash high-resistivity samples and sets the phase fitting target label value of the pure ash high-resistivity sample to -1. Pure ash high-resistivity samples release charge slowly and easily form a dense insulating layer on the electrode surface. When the electric field strength increases, it will induce micro-discharge oscillations on the electrode.

[0033] When the data in the dust phase feature vector does not meet the above two types of boundary conditions, the controller marks the corresponding data as a normal dust phase sample and filters it as zero-weight data in subsequent model training. The controller classifies and labels the historical dust collector ash hopper sampling and testing data based on preset loss on ignition threshold and resistivity critical value, establishes the physical mapping relationship between the unit's testing parameters and the bottom phase of dust, and provides labeled sample data for subsequent AI model training.

[0034] S102, read historical voltage and current characteristic data to extract the critical impedance reference parameter of the back corona. The controller reads historical voltage and current characteristic data from the unit control system, extracts the inflection point of the voltage and current characteristics, and then retrieves the electrical operation log recorded by the unit control system under the boost test state to extract the output voltage and output current sequence corresponding to the high-frequency inverter power supply unit.

[0035] The controller calculates the discrete derivative of the voltage change with respect to the current change. During the normal corona discharge phase, the output voltage increases with the increase of the output current, and the derivative calculation results show a positive correlation trend. When the electrode dust insulation layer reaches its electrical withstand limit and breaks down, the current flowing through the electrode increases, and the inter-electrode voltage drops. The controller monitors the derivative calculation results in a time series. When the derivative calculation results of three consecutive sampling cycles turn from positive to negative, the first data frame with a negative zero-crossing point is determined as the inflection point of the voltage and current characteristics.

[0036] The inflection point of the voltage and current characteristics represents the initial threshold of the back corona breakdown phenomenon at the electrophysical level. The controller extracts the transient voltage and current values ​​at the corresponding inflection point of the voltage and current characteristics, calculates the ratio of the transient voltage to the transient current, and marks it as the critical impedance reference parameter for back corona and stores it in the storage medium. For the voltage and current derivatives and static impedance calculation algorithms of conventional electrical equipment, those skilled in the art can directly perform the calculation and extraction using the backward difference method. The data processing method is a well-known technology in this field and will not be elaborated here.

[0037] The controller performs AI model calibration of control parameter thresholds and feature weights, including the following steps: S103, construct a multiple linear regression model to calculate the weights of dust morphology features. The controller extracts dust microscopic image data accumulated within historical operating cycles, performs grayscale conversion, binarization, and connected component analysis on the image data, and extracts the contours of independent particles in the microscopic images. The controller calculates the ratio of the number of pixels on the long side to the short side of the minimum bounding rectangle of each independent particle contour, and calculates the temporal statistical variance of all particle ratios within that historical cycle to obtain historical aspect ratio feature values. The Sobel operator is used to calculate the gradient of the pixels at the contour edges, and the mean of the overall pixel gradient is calculated to obtain historical edge gradient feature values. The controller uses the phase state fitting target label values ​​of the carbon-rich low-resistivity samples and the pure gray high-resistivity samples classified and labeled in step S101 as the output target values ​​for supervised learning.

[0038] The controller constructs a supervised learning-based multiple linear regression model. This model uses historical aspect ratio features and historical edge gradient features as input feature vectors to establish a quantitative mathematical relationship between microscopic morphology and macroscopic phase. The calculation formula for the multiple linear regression model is as follows: ; In the formula, This indicates that the multiple linear regression model applies to the first... The dust phase fit index of each sample output; For sample index sequence number; This indicates the aspect ratio item weights that need to be calibrated. Indicates the first Historical aspect ratio features of each sample; This indicates the weights of the marginal gradient terms that need to be calibrated; Indicates the first Historical edge gradient feature values ​​of each sample; This represents the fitting bias constant.

[0039] The controller employs a gradient descent algorithm for iterative training on historical data. The training process aims to minimize the mean squared error between the dust phase fit exponent output by the model and the target phase fit label value. The loss function is calculated as follows: ; In the formula, This represents the mean squared error loss function; This represents the total number of samples used in the training process. Indicates the first The phase fitting target label values ​​are corresponding to each sample. When the loss function converges to the preset lower error limit, the controller stops training and outputs the aspect ratio weights and marginal gradient weights in the converged state. For the gradient descent solution step size setting and parameter update process of the multiple linear regression model, those skilled in the art can use conventional machine learning frameworks to implement it; the underlying network training method is a well-known technique in the field and will not be elaborated here.

[0040] S104. The control parameter thresholds are tuned by combining training results with baseline parameters. The controller uses the trained multiple linear regression model to calculate all carbon-rich, low-resistivity state samples from the historical database, obtaining a set of carbon-rich phase indices. The controller calculates the statistical distribution of this set, extracts the upper limit feature value corresponding to the 95% confidence interval, and tunes it as the upper limit threshold for the carbon-rich, low-resistivity state of the dust. When the dust morphology irregularity index calculated in subsequent real-time exceeds this upper limit threshold, the system determines that there is a tendency for secondary dust re-entrainment.

[0041] The controller performs inference calculations on pure gray high-resistivity state samples in the historical database to obtain a set of pure gray phase state indices. It then extracts the lower limit feature value corresponding to the 95% confidence interval of this pure gray phase state index set under the statistical distribution and tunes it to the low limit threshold of the pure gray high-resistivity state. This low limit threshold of the pure gray high-resistivity state is used to define the physical boundary at which dust densifies the insulating layer on the electrode surface.

[0042] The controller retrieves the historical transient voltage and current sequences corresponding to the back corona critical test stage in step S102, and calculates and generates a historical transient impedance absolute value sequence. To eliminate the measurement bias caused by the different absolute impedance base values ​​under different basic operating conditions, the controller presets a fluctuation extreme value normalization extraction formula: ; In the formula, This represents the impedance volatility calculated using normalization. This represents the maximum absolute value of impedance in the time series of transient impedance absolute values; This represents the minimum absolute value of impedance in the time series of transient impedance absolute values; This represents the zero bias constant, whose value is set to 10. -3 Ω to 10 -1 Ω.

[0043] The controller extracts the maximum and minimum absolute impedance values ​​from the historical transient impedance absolute value sequence and substitutes them into the above formula to calculate the base impedance fluctuation rate under the back corona critical state. A safety margin factor is then introduced to derating this base impedance fluctuation rate. Based on the insulation withstand characteristics of high-voltage electrical equipment, the safety margin factor is set to a range of 0.8 to 0.9. The controller then tunes the product of the base impedance fluctuation rate and the safety margin factor to the differential impedance fluctuation rate threshold.

[0044] This differential impedance volatility threshold serves as the electrical baseline for determining the initiation state of the back corona micro-discharge of the electrode plate using the multi-parameter coupled logic matrix during subsequent real-time operation. After completing the above calculations, the controller writes the calibrated aspect ratio weights, edge gradient weights, high threshold for the dust-rich carbon-rich low-resistivity state, low threshold for the pure ash high-resistivity state, and differential impedance volatility threshold into the storage medium, thus completing the baseline parameter calibration work before system operation.

[0045] See attached document Figure 4 The controller executes the in-situ sampling environment dynamic tracking and image abnormal particle removal logic in the in-situ sampling and dust morphology feature extraction process, including the following steps: S201, the controller drives the heat tracing sampling module to perform closed-loop tracking adjustment of the sampling temperature. The heat tracing sampling module includes a main flue thermocouple, a heat tracing cable, and a wall temperature sensor. The main flue thermocouple measures the actual flue gas temperature at the measuring point in the main flue in real time, and the controller uses this actual flue gas temperature as the dynamic target setpoint for heat tracing temperature control. The heat tracing cable covers the outer wall of the isokinetic sampling probe and the optical measurement cavity, and the wall temperature sensor is used to monitor the actual heated temperature of the outer wall.

[0046] The controller employs a proportional-integral-derivative (PID) control algorithm to adjust the heating power of the heat tracing cable. The system controls the wall temperature of the isokinetic sampling probe and the optical measurement cavity to follow the actual flue gas temperature changes at the main flue gas measuring point, with the controller maintaining the temperature deviation between them within ±2℃. This temperature control mechanism prevents moisture condensation and precipitation in the high-temperature flue gas during bypass sampling and cooling, maintaining the in-situ physical consistency of the physicochemical properties of the liquid film on the dust surface and the spatial aggregation morphology, thus avoiding interference from sampling distortion in subsequent morphology identification. For the PID control algorithm of conventional heating equipment, those skilled in the art can directly program and implement it using positional or incremental digital algorithms; its underlying closed-loop regulation principle is well-known in the field and will not be elaborated upon here.

[0047] S202, the optical measurement cavity acquires microscopic images. The heated sampling module introduces dusty airflow from the main flue into the optical measurement cavity at a constant velocity. A stroboscopic light source inside the cavity provides instantaneous illumination with a pulsed flash frequency matching the flow velocity of the fluid cross-section, suppressing image trailing caused by high-speed dust particles during exposure. The microscopic imaging component continuously exposes at a set sampling frame rate, acquiring two-dimensional raw microscopic image frames containing the distribution of dust particles in the airflow. The controller reads and stores these two-dimensional raw microscopic image frames in its internal buffer.

[0048] S203, the controller utilizes a lightweight AI edge detection model to remove abnormal particles from the original two-dimensional microscopic image frames. The controller pre-deploys a pre-trained AI edge detection model. This AI edge detection model employs a fully convolutional neural network architecture that includes depthwise separable convolutional layers and a feature pyramid network.

[0049] The offline training samples of the AI ​​edge detection model come from a standard dust particle image dataset collected during the historical maintenance of electrostatic precipitators. Manually labeled physical boundaries are used as labels for supervised learning. The offline training process of the model uses a binary cross-entropy loss function to evaluate the error between the output edge probability and the true labeled boundary, and an adaptive moment estimation optimization algorithm is used to iteratively update the network weight parameters.

[0050] The controller extracts the original two-dimensional microscopic image frames from the buffer, performs grayscale and normalization preprocessing on them, and converts them into a single-channel feature tensor matrix that conforms to the model input dimension. The shape dimension of this single-channel feature tensor matrix is ​​image width × image height × 1. The controller inputs the single-channel feature tensor matrix into the AI ​​edge detection model. The depthwise separable convolutional layers within the model extract the texture features of the image level by level. After the feature pyramid network fuses multi-scale features, it outputs a pixel-level edge probability map with the same size as the input image. The values ​​in the edge probability map map map the spatial probability that the corresponding pixel belongs to the physical edge of the particle.

[0051] S204, the controller executes the logic for removing out-of-focus and overlapping contour particles based on the edge probability map. The controller uses Otsu's algorithm to determine the adaptive truncation threshold of the edge probability map, converts it into a binary contour map, and extracts all closed regions in the map as candidate particles. To remove out-of-focus and blurred particles, the controller locates the edge coordinates of the candidate particles on the original grayscale image and calculates the mean pixel gradient of this set of edge coordinates. The controller sets a low-sharpness threshold, which, based on the statistical characteristics of edge pixel gradients in an 8-bit grayscale image, ranges from 45 to 65. When the mean pixel gradient of a candidate particle is lower than the low-sharpness threshold, the controller determines that the particle is outside the focal plane of the imaging component and removes it from the candidate queue.

[0052] For candidate particles that pass the defocusing detection, the controller continues to execute the overlapping particle removal logic. When multiple particles physically obstruct or adhere to each other in the flow field, their two-dimensional projection will exhibit concave edge characteristics. The controller calculates the actual enclosed area of ​​the candidate particle and the minimum circumscribed convex hull area of ​​the particle. The controller calculates the ratio of these two to obtain the convex hull area ratio, and the calculation formula is as follows: ; In the formula, This indicates the ratio of the convex hull area of ​​the candidate particles; This represents the actual enclosed area occupied by the candidate particle in the image; This represents the area of ​​the smallest circumscribed convex polygon that contains all the edge pixels of the candidate particle.

[0053] The controller has a preset lower threshold for the convex hull area ratio. Based on the typical geometric characteristics of dust particles, the lower threshold is set to a range of 0.85 to 0.95. When the convex hull area ratio of a candidate particle... When the convex hull area ratio falls below the set lower threshold, the controller determines that the candidate particle has overlapping occlusion or adhesion and removes it from the candidate queue. When the convex hull area ratio of the candidate particle is below the set lower threshold, the controller determines that the candidate particle has overlapping occlusion or adhesion and removes it from the candidate queue. When the ratio of the convex hull area to the set lower threshold is greater than or equal to the threshold value, the controller determines that the candidate particle has not been occluded or adhered and retains it in the candidate queue. After the above-mentioned defocusing and overlap logic judgment and elimination, the particle objects that are finally retained are marked by the controller as valid single particles, which serve as a reliable data source for subsequent extraction and calculation of the dust morphology irregularity index.

[0054] The controller performs the feature extraction calculation operation of dust morphology irregularity index in in-situ sampling and dust morphology feature extraction, including the following steps: S205, the controller calculates the temporal statistical variance of the effective particle aspect ratio. The controller extracts the set of effective single particles that are marked and retained after anomaly removal, calculates the minimum bounding rectangle of each effective single particle in the set, and extracts the ratio of the number of pixels on the long side to the number of pixels on the short side of the minimum bounding rectangle, which is used as the real-time aspect ratio of the corresponding particle.

[0055] Within a defined visual sampling time window, the system continuously acquires the effective single-unit particle sequence from microscopic image frames. The length of this visual sampling time window is determined based on the average flow velocity of the main flue section and the physical volume of the optical measurement cavity. Based on the complete residence cycle of the fluid through the measurement cavity, the value range of the visual sampling time window is set from 1s to 3s.

[0056] The controller calculates the temporal statistical variance of the real-time aspect ratio of all valid individual particles within the visual sampling time window. The calculation formula is as follows: ; In the formula, The time-domain statistical variance represents the effective particle aspect ratio; This represents the total number of effective monomer particles within the visual sampling time window. The index number indicating the effective monomer particles; Indicates to From 1 to The calculation results of all effective monomer particles are summed together; Indicates the first Real-time aspect ratio of each effective monomer particle; This represents the arithmetic mean of the real-time aspect ratios of all valid individual particles within the visual sampling time window.

[0057] The temporal statistical variance of the effective particle aspect ratio characterizes the particle morphology distribution. A larger temporal statistical variance indicates a higher proportion of irregularly shaped particles in the dust airflow sample, and a greater dispersion in the overall particle morphology. A smaller value indicates that the particle shapes tend to be uniform, mostly exhibiting regular spherical or ellipsoidal shapes.

[0058] S206, the controller calculates the dust morphology irregularity index. Besides the macroscopic dispersion characteristics of particle morphology, the microscopic roughness of the particle surface is also an important physical basis for characterizing the dust phase. The controller retrieves the edge pixel coordinates of all valid individual particles within the visual sampling time window and calculates the gradient magnitude of these edge pixels in the original grayscale image. The controller accumulates the edge gradient magnitudes of all valid individual particles within the visual sampling time window and divides them by the total number of edge pixels involved to obtain the overall average edge pixel gradient.

[0059] For the calculation of the gradient magnitude of pixel edges in an image, those skilled in the art can directly call and execute the Sobel operator or conventional open-source machine vision functions. The spatial convolution calculation method based on pixel neighborhood difference is a well-known technology in this field and will not be described in detail here.

[0060] The controller reads the aspect ratio weights, edge gradient weights, and fitting bias constants pre-written in the reference calibration phase from the storage medium. Using a linear model, it mathematically fuses the temporal statistical variance of the effective particle aspect ratio with the mean of the overall edge pixel gradient to output the real-time dust morphology irregularity index. The mathematical calculation model for the dust morphology irregularity index is as follows: ; In the formula, Indicates the irregularity index of dust morphology; The time-domain statistical variance represents the effective particle aspect ratio; This indicates the weight of the aspect ratio item being read; This represents the average gradient of all edge pixels; This indicates the weights of the edge gradient terms being read. This represents the fitting bias constant.

[0061] The controller stores the calculated dust morphology irregularity index into its internal cache. This dust morphology irregularity index comprehensively quantifies the geometric irregularity and surface roughness characteristics of the dust particle set, serving as a key input parameter for subsequent assessment of the risk of dust layer densification on the electrode and for guiding the dynamic redistribution control of the target electric field.

[0062] See attached document Figure 5 The controller performs dynamic optimization of the target electric field based on system delay compensation in the dynamic redistribution of the target electric field and the equivalent calculation of time window features, including the following steps: S301, the controller constructs the control timing margin calculation rules. The acquisition of dust images, model inference, and mathematical feature extraction require hardware computing power and occupy a certain time span, which is the inherent processing delay of the system. During this delay, the dust airflow in the main flue continues to move downstream of the electrostatic precipitator at a certain speed. To ensure that control commands can act on the same batch of dust sampled and analyzed in situ, the controller needs to time-domain align the spatial fluid motion law of the dust with the system's hardware execution cycle.

[0063] The controller reads the average flow velocity of the main flue section transmitted back from the unit's control system. The controller retrieves the internally stored physical dimensions of the electrostatic precipitator and extracts the physical distance along the path from the isokinetic sampling probe position of the heat tracing sampling module to the inlet of each downstream independent electric field stage. Inside the electrostatic precipitator, along the flue gas flow direction, are sequentially distributed... Level independent electric field, For the electric field series index, and The value ranges from 1 to A continuous sequence of positive integers.

[0064] The controller calculates the dust airflow reaching the first The theoretical flight time at the entrance of the electric field is calculated using the following formula: ; In the formula, Indicates that the dust airflow has reached the first The theoretical flight time of the electric field; Indicates the isokinetic sampling probe up to the first The physical distance along the path from the entrance of the electric field; This represents the average flow velocity across the main flue section.

[0065] The controller records the time difference between the completion of exposure of the microscopic image frame of the optical measurement cavity and the output of the dust morphology irregularity index, and tunes this time difference to the system's inherent processing delay, which is set to a range of 0.1s to 0.5s. The controller subtracts the theoretical flight time from the system's inherent processing delay to calculate the time difference for the first... Control time margin of the electric field.

[0066] S302, the controller iterates through the selected target electric field. The controller has a preset response time margin lower threshold, the physical reference of which depends on the electrical step setup time of the composite power supply module and the data packet transmission delay of the communication bus. To ensure that the power supply voltage can be stably output before a new wave of dust reaches the plate, the value range of this response time margin lower threshold is set to 0.5s to 1.5s.

[0067] The controller initiates a traversal query program, indexed by electric field level. Starting from 1, the values ​​are incremented progressively, comparing the control time margin and the lower threshold of the response time margin for each independent electric field. When the... When the control time margin of the electric field stage is greater than or equal to the lower threshold of the response time margin, the controller determines that the system possesses the temporal physical conditions for implementing feedforward parameter adjustment for that independent electric field stage. The controller selects the smallest electric field stage index that satisfies the above conditions. Stop the traversal operation and turn the number of iterations to the next highest level. The primary electric field is selected as the target electric field, thereby applying electric field adsorption force to the dust as early as possible in the spatial sequence.

[0068] When the controller traverses to the last stage of independent electric field, i.e., the index number equal If the control time margin of the final-stage electric field is still less than the lower threshold of the response time margin, it indicates that the current flue gas velocity is too high or the computing module has experienced a short-term blockage. In response to this condition, the controller forcibly sets the final-stage electric field to the target electric field. This forced setting prevents the control logic from entering a dead zone under extreme conditions. The controller binds the previously calculated dust morphology irregularity index with the theoretical flight time of the corresponding target electric field and stores it in the action queue, providing data support for the composite power supply module of the target electric field to execute relevant control commands on time.

[0069] The controller performs the tolerance generation and time-domain overlap smoothing of the probabilistic time window in the dynamic redistribution of the target electric field and the equivalent calculation of time window characteristics, including the following steps: S303, the controller generates the probability arrival time window. In the actual physical flow field of an electrostatic precipitator, dust particles do not completely follow the ideal plug flow motion model. Affected by the combined effects of flue gas turbulence and friction along the wall, dust batches that complete in-situ sampling within the optical measurement cavity at the same time point will experience different motion patterns as they move towards the next... During the migration of the target electric field, flow diffusion occurs along the path. This axial diffusion causes the time for the batch of dust to reach the target electric field to broaden from a fixed point in time to a time interval.

[0070] The controller acquires the real-time kinematic viscosity of the flue gas inside the main flue and the hydraulic diameter of the main flue. Combined with the average flow velocity of the main flue cross-section obtained in the previous steps, the local Reynolds number of the flow field space is calculated. For the calculation process of the local Reynolds number in pipe flow, those skilled in the art can directly calculate it using conventional fluid dynamics flow pattern determination formulas, which are well-known techniques in the field and will not be elaborated here.

[0071] The controller internally stores pre-calibrated empirical mapping data between Reynolds number and fluid diffusion coefficient. It uses the local Reynolds number of the current flow field space to look up a table and calculates the fluid diffusion coefficient corresponding to the current operating condition using a linear interpolation algorithm. A higher Reynolds number indicates more severe turbulence in the flow field, and the corresponding fluid diffusion coefficient increases accordingly. Based on the flow field characteristics of a conventional coal-fired unit flue, the value range of this fluid diffusion coefficient is set to 0.05 to 0.15. The controller calculates the fluid diffusion tolerance by multiplying the fluid diffusion coefficient by the theoretical flight time of the batch of dust, and then generates the probability arrival time window for that batch of dust. The formula for calculating the time boundary of this probability arrival time window is: ; ; In the formula, This indicates the start time of the probability arrival time window; This indicates the end time of the probability arrival time window; This indicates the starting moment of the action corresponding to the completion of visual sampling of the microscopic image; Indicates that the dust airflow has reached the first The theoretical flight time of the electric field; This represents the fluid diffusion coefficient estimated based on the local Reynolds number.

[0072] S304, the controller executes a time-domain overlap determination and mathematical smoothing strategy. With the system's periodic continuous sampling, the controller sequentially generates multiple probability arrival time windows arranged in chronological order in the action queue. Due to the axial broadening effect caused by fluid diffusion, the leading edge of the later batch of dust arrives earlier, while the tail of the previous batch arrives later. The controller continuously compares the probability arrival time window boundaries of adjacent batches.

[0073] The controller determines the current number The probability of a batch reaching the start time of the time window and the previous batch The numerical relationship between the probability of a batch reaching the end of the time window. When the... The start time of the batch is greater than or equal to the first... At the end of the batch, the controller determines that the two batches of dust did not spatially mix when entering the target electric field, and that their time windows are independent. At this point, the controller directly uses the original dust morphology irregularity index of the corresponding batch as the equivalent morphology feature and sets the equivalent local variance to the preset basis bias constant. That is, the minimum value of the basic fluctuation when there is no overlap.

[0074] When the The start time of the batch is less than the first At the end of the batch, the controller determines that the two probability arrival time windows overlap in the time domain. At this point, within the overlapping time period, the first... Within the electric field, two batches of mixed dust with different morphological characteristics will coexist simultaneously. To avoid frequent jumps in the underlying control parameters during the overlapping period, the controller executes a data smoothing algorithm. During the overlapping period, the controller aggregates the morphological irregularity indices of all independent batches of dust involved in the overlap and extracts the average state of these morphological characteristics as the equivalent morphological feature. The formula for calculating the equivalent morphological feature is as follows: ; In the formula, Indicates the equivalent morphological features within the overlapping time period; This represents the total number of time windows at which the probability of temporal overlap occurs at the current moment; The index number indicating the time window where the overlap occurs; Indicates to From 1 to The data corresponding to all overlapping time windows are summed together. Indicates the first The original dust morphology irregularity index is bound to the overlapping time windows.

[0075] Simultaneously, the controller calculates the statistical fluctuations of these overlapping features deviating from the average state, generating an equivalent local variance. The mathematical smoothing formula for the equivalent local variance is: ; In the formula, This represents the equivalent local variance within the overlapping time period; This represents the basis bias constant introduced to prevent the calculation result from being zero. The basis bias constant is set to ensure that the divisor in the subsequent pulse reconstruction algorithm is not zero, and its value range is set to 0.01 to 0.05.

[0076] The controller synchronously writes the generated equivalent topographic features and equivalent local variance into its internal storage register. When the control timeline advances to the overlap interval of this domain, the system calls upon this smoothed equivalent data to prepare for subsequent reconstruction. The probe pulse parameters injected by the electric field provide a continuous and stable feedforward feature source.

[0077] See attached document Figure 6 The controller executes the probe pulse parameter modulation logic based on equivalent topography features in the probe pulse reconstruction and transient impedance synchronization detection operation, including the following steps: S401, the controller extracts equivalent topographic features and establishes a feedforward coupled control mechanism. When the control time axis advances to the time-domain overlap interval of the target electric field, the controller reads the equivalent topographic features corresponding to the current operating condition from the internal storage register. The equivalent topographic features quantify the entry into the... The microscopic manifestations of the mixed dust under an electric field in terms of geometry and surface roughness.

[0078] When the equivalent morphology feature value is large, it indicates that the dust particles in the current electric field space are rich in irregular carbon particles or have high conductivity. Dust particles of different sizes and surface roughness exhibit differences in charge rate and spatial divergent motion trajectory under the base DC voltage. This inconsistent distribution of physical characteristics causes uneven distribution of space charge between the plates, thereby increasing the probability of inducing local micro-discharge and back corona discharge in the dust layer. To cope with the above complex operating conditions, the controller uses the equivalent morphology feature output from the preceding vision system as a direct control input parameter, establishing a feedforward coupling control link from visual features to electrical control, and dynamically adjusting the injection characteristics of the detection pulse.

[0079] S402, the controller calculates the adaptively reconstructed probe pulse width. To improve the time-domain resolution of electrical condition detection when dust characteristics are complex, the controller establishes negative feedback parameter adjustment logic. The larger the value of the equivalent topographic feature, the narrower the probe pulse width of the reconstructed output. The adaptive modulation formula for the probe pulse width is: ; In the formula, Indicates the width of the reconstructed probe pulse; This indicates the reference pulse width set internally by the control system; This represents the set pulse width modulation coefficient; This represents the equivalent topographic features read. Reference pulse width. Based on the rated resonant period setting of the composite power supply module, its value range is set to 10μs to 50μs. Pulse width modulation coefficient. The numerical magnitude used to match equivalent morphological features is set to a range of 0.5 to 2.0.

[0080] Due to the inherent turn-on and turn-off delays of power electronic switching devices such as insulated-gate bipolar transistors (IGBTs), the controller presets a minimum time span for reliable operation of the underlying hardware, using this as the lower limit threshold for the pulse width. Considering the electrical response characteristics of conventional high-power switching transistors, the range of this lower limit threshold is set to 1 μs to 5 μs. When the detection pulse width is calculated according to the above modulation formula... When the pulse width is less than the lower threshold, the controller will force the final output probe pulse width to be clamped to that lower threshold to prevent the hardware from failing to operate due to insufficient drive pulse width. The probe pulse width calculated according to the above modulation formula... When the pulse width is greater than or equal to the lower limit threshold, the controller directly outputs the calculated probe pulse width. The probe pulse width is used as the final output, thus ensuring the completeness of the pulse width adjustment logic.

[0081] S403, the controller calculates the adaptive reconfiguration probe pulse injection frequency. As the dust morphology differences increase, the... The frequency of transient impedance fluctuations within the electric field increases accordingly. The controller synchronously increases the injection frequency of the probe pulse to increase the sampling density of the electrical state, preventing the control system from missing intermittent discharge precursor signals. The adaptive modulation formula for the probe pulse injection frequency is: ; In the formula, This indicates the frequency of the probe pulse injection for reconstruction; This indicates the reference injection frequency set internally by the control system; This indicates the set frequency modulation coefficient. Reference injection frequency. Based on the basic inspection cycle setting of the control system, its value range is set to 10Hz to 50Hz. Frequency modulation coefficient The value range is set to 5 to 15.

[0082] Excessively high injection frequencies can increase switching losses within the composite power supply module and lead to unfavorable heat accumulation on the heatsink surface. The controller sets a maximum frequency threshold to protect the power supply circuit. Considering the physical heat dissipation conditions and component thermal resistance of a typical high-frequency inverter power supply unit, this maximum frequency threshold is set to a range of 200Hz to 500Hz. When the probe pulse injection frequency calculated according to the above modulation formula... When the maximum frequency threshold is exceeded, the controller will limit the final output probe pulse injection frequency to that maximum frequency threshold. This applies when the probe pulse injection frequency calculated using the modulation formula described above... When the frequency is less than or equal to the upper limit threshold, the controller directly injects the calculated probe pulse at the frequency. The probe pulse injection frequency, which is the final output, forms a complete amplitude limiting control range.

[0083] S404, the controller drives the pulse superposition unit to inject the reconstructed probe pulse. The controller packages the calculated probe pulse width and probe pulse injection frequency into data, generates the underlying drive control command, and sends the command to the next level via the communication bus. Composite power supply module corresponding to the electric field level.

[0084] During this process, the high-frequency inverter power supply unit inside the composite power supply module continuously supplies power to the first... The primary electric field outputs a stable base DC voltage. The pulse superposition unit within the composite power supply module receives and responds to the controller's drive commands, periodically switching the switching transistors according to the reconstructed probe pulse width and injection frequency specified in the commands, superimposing a transient high-voltage pulse onto the base DC voltage waveform. This transient high-voltage pulse is injected as an active detection signal into the primary electric field. A high-level electric field is used to excite the transient electrical response in the electric field power supply circuit, providing an excitation source for subsequent impedance feature extraction.

[0085] The controller performs zero-zero transient differential impedance calculation and fluctuation characteristic extraction in the probe pulse reconstruction and transient impedance synchronous detection operation, including the following steps: S405, the isolation sampling module synchronously acquires signals and executes hardware interlocking protection logic. In the pulse superposition unit, to the... During the period of the first-stage electric field injection probe pulse, the isolation sampling module utilizes its internal resistor-capacitor voltage divider and high-frequency transformer to synchronously acquire the first-stage electric field probe pulse. The transient voltage and current signals in the electric field power supply circuit are collected. The isolation sampling module converts the acquired continuous analog signals into digital sequences after analog-to-digital conversion, and then discretizes them into digital sequences at a set sampling frequency, which are then transmitted to the controller in real time.

[0086] The controller receives discrete transient current signals and calculates the rate of change of current between two adjacent sampling points. The controller has a preset safe threshold for the rate of change of current. The physical reference for this safe threshold depends on the transient surge current withstand capability of the power electronic devices within the high-frequency inverter power supply unit. Considering the safe operating area characteristics of high-power switching transistors, the safe threshold for the rate of change of current is set to a range of 1000 A / s to 5000 A / s.

[0087] The controller performs branching decisions based on the calculated current change rate. When the current change rate is greater than or equal to a safe threshold, the decision is made in the first branch. If an inter-electrode short circuit or spark breakdown occurs within the electric field, the controller immediately triggers hardware latch-up logic, sending a blocking pulse command to the composite power supply module to forcibly shut down all drive signals of the high-frequency inverter power supply unit and the pulse superposition unit, thus protecting the underlying hardware from damage caused by high-current thermal shock. When the rate of change of current is less than the safe threshold, the controller determines that the current electric field is in a normal discharge detection state, and continues to perform subsequent transient impedance calculations.

[0088] S406, the controller calculates the transient differential impedance to prevent division by zero. In conventional differential impedance calculations, the voltage change is usually directly divided by the current change. However, during the peak phase of the probe pulse injection or due to sampling noise, the change in the transient current signal between adjacent sampling points tends to approach zero. Direct division can lead to a divisor of zero or a result approaching infinity, causing operational divergence and software overflow anomalies in the control system.

[0089] To address the aforementioned numerical calculation divergence problem, the controller introduces a perturbation constant to prevent the denominator from approaching zero, and performs zero-prevention transient differential impedance calculations. The controller extracts transient voltage and current signals to calculate the first... The formula for calculating the transient differential impedance for each sampling period is: ; In the formula, Indicates the first The transient differential impedance for each sampling period; Indicates the first The change in transient voltage signal corresponding to each sampling period; Indicates the first The change in transient current signal corresponding to each sampling period; Represents a sign function, when Its value is 1 when it is greater than or equal to zero, when Its value is -1 when it is less than zero; This represents the perturbation constant to prevent the denominator from approaching zero.

[0090] To prevent the denominator from approaching zero, the perturbation constant is used. The introduction of this feature ensures that the denominator remains non-zero regardless of the magnitude of the current change, thus maintaining the numerical stability of the impedance calculation process. Based on the quantization resolution and noise floor level of the isolated sampling module, this perturbation constant prevents the denominator from approaching zero. The value range is set to 10. -5 A to 10 -3 A.

[0091] S407: The controller extracts the extreme values ​​of the fluctuations and calculates the normalized transient differential impedance fluctuation rate. To prevent the negative resistance effect caused by electric field discharge from causing the calculation denominators to cancel each other out and resulting in computational overflow, the controller will detect all transient differential impedances within the pulse duration. The absolute values ​​are taken and compiled into a time series of transient impedance absolute values. The controller traverses this time series of transient impedance absolute values ​​and extracts the maximum and minimum impedance absolute value extrema.

[0092] The redistribution of space charge between electrodes caused by the probe pulse leads to transient fluctuations in impedance data. To eliminate measurement bias caused by differences in the absolute impedance baseline values ​​under different basic operating conditions, the controller calls the fluctuation extreme value normalization extraction formula disclosed in step S104, substitutes the currently extracted maximum and minimum absolute impedance extreme values ​​into the formula, and calculates the transient differential impedance fluctuation rate within the current probe cycle. .

[0093] Using the normalization algorithm described above, the controller constrains the numerical boundary of the transient differential impedance volatility within the range of 0 to 1. The calculated transient differential impedance volatility reflects the... The dynamic response intensity of the dust charge state inside the electric field is recorded and stored as a core electrical feedback parameter in the controller's shared memory for subsequent multi-parameter logic decision-making and power topology mode switching modules to call.

[0094] See attached document Figure 7 The design of multi-parameter two-dimensional state determination rules and variable frequency power supply state switching in the controller's execution of multi-parameter logic decisions and power topology mode switching operations includes the following steps: S501, the controller constructs a multi-parameter two-dimensional state determination rule. The controller calls the equivalent topography features and transient differential impedance fluctuation rate, which have been time-aligned and stored in the internal memory register. The equivalent topography features quantify the microscopic visual feedforward features of the dust morphology, while the transient differential impedance fluctuation rate reflects the macroscopic electrical feedback features of the inter-electrode discharge state. The controller establishes a two-dimensional state space with the equivalent topography features as the first input dimension and the transient differential impedance fluctuation rate as the second input dimension.

[0095] The controller sets operational characteristic boundaries in a two-dimensional state space. The controller invokes the high-limit threshold for the dust-rich, low-resistivity state and the differential impedance fluctuation rate threshold calibrated in step S104; these two thresholds are combined to define the low-resistivity risk state. To replace the original empirical static parameters, the conventional benchmark value range for the high-limit threshold for the dust-rich, low-resistivity state is set to 15 to 30. The controller simultaneously invokes the low-limit threshold for the pure ash high-resistivity state and the preset fluctuation rate low-limit threshold calibrated in step S104; these two thresholds are combined to define the high-resistivity state. The conventional benchmark value range for the low-limit threshold for the pure ash high-resistivity state is set to 2 to 8; considering the relatively gradual impedance change in the high-resistivity state, the fluctuation rate low-limit threshold is constantly set to 0.1 to 0.3. The controller uses these set thresholds to divide the two-dimensional state space into multiple independent decision regions.

[0096] S502, the controller determines and responds to a low-resistance risk state. The controller compares the equivalent topographic features of the current control cycle with the transient differential impedance fluctuation rate. When the equivalent topographic features are greater than or equal to the upper limit threshold of the dust-rich carbon low-resistance state, and the transient differential impedance fluctuation rate is greater than or equal to the differential impedance fluctuation rate threshold, the controller determines the current control cycle's risk state. The electric field is in a low-resistance risk state. This state indicates that the dust particles entering the target electric field have irregular shapes, the space charge distribution between the electrodes is highly uneven, local micro-spark discharges frequently occur inside the electric field, and the overall equivalent impedance of the electric field shows a significant downward trend.

[0097] Under low-impedance risk conditions, if transient high-voltage pulses continue to be superimposed on the power supply circuit, the micro-spark discharge channel can easily evolve into a destructive, continuous arc breakdown. The controller immediately sends a mode switching command to the corresponding composite power supply module and sends a low-level blocking signal to shield the drive output of the pulse superposition unit, causing it to stop working.

[0098] Simultaneously, the controller dynamically adjusts the operating frequency of the high-frequency inverter power supply unit. The controller directly switches and maintains the switching frequency of the inverter bridge within the high-frequency inverter power supply unit at a set safe upper limit to reduce the ripple coefficient of the output rectified voltage. The high-frequency inverter power supply unit then switches to a pure DC power supply topology mode, supplying power to the first... The electric field outputs a constant high-amplitude DC fundamental frequency. This control action maintains the basic dust removal efficiency of the electric field by relying on a stable high average voltage, while avoiding severe arc breakdown. The safe upper limit of the operating frequency is limited by the heat dissipation capability of the power switching transistor, and its value is set to 20kHz.

[0099] S503, the controller determines and responds to a high-resistivity state. When the equivalent morphological characteristics are less than or equal to the lower threshold of the high-resistivity state for pure gray dust, and the transient differential impedance fluctuation rate is less than or equal to the lower threshold of the fluctuation rate, the controller determines that the current electric field is in a high-resistivity state. This state indicates that the dust particles have a uniform morphology and fine particle size, and a high-resistivity dust layer is easily deposited on the electrode surface. The migration of internal charges in the dust layer towards the grounded electrode is hindered, and the inter-electrode impedance increases abnormally, possessing the physical conditions to induce back corona discharge in the dust layer.

[0100] The controller sends a pulse power supply switching command to the composite power supply module. The high-frequency inverter power supply unit reduces its duty cycle, lowering the output base DC voltage to the corona initiation voltage critical point. The corona initiation voltage critical point can be obtained directly by those skilled in the art by querying the electric field volt-ampere characteristic curve pre-stored in the control system; this is well-known technology in the field and will not be elaborated upon here.

[0101] The controller unshields and activates the pulse superposition unit, then calls the internal memory to retrieve the high resistivity dust empirical charge discharge time constant preset therein. The physical reference for this high resistivity dust empirical charge discharge time constant depends on the dielectric properties of conventional high-alumina-silicon ash, and its value is set to range from 10ms to 50ms.

[0102] The controller calculates the time interval between two adjacent transient high-voltage pulses and defines it as the output pause time. The controller adds a preset margin time to the empirical charge discharge time constant of the high-resistivity dust, setting the output pause time to the sum of the empirical charge discharge time constant and the margin time. This margin time ranges from 2ms to 5ms. By superimposing a low-duty-cycle transient high-voltage pulse onto a lower base DC voltage, the pulse width is determined by the adaptively reconstructed probe pulse width from step S402. The pulse superposition unit provides transient strong ionization energy to the electric field. The control action of matching the empirical charge discharge time constant of the high resistivity dust makes the long output pause time conducive to the release of the charge accumulated inside the dust layer to the electrode, thereby suppressing the backflow of positive ions from the back corona discharge.

[0103] S504, the controller determines and responds to the normal operating state. When the equivalent topographic features and transient differential impedance fluctuation rate do not simultaneously fall into the aforementioned low-resistance risk state region or high-resistance state region, the controller determines that the current electric field is in a normal operating state. The controller maintains the base DC voltage of the high-frequency inverter power supply unit at a normal value and drives the pulse superposition unit to output active detection pulses according to the pulse width and frequency calculated by the aforementioned adaptive reconstruction algorithm, continuously performing cyclic detection of the electric field electrical parameters.

[0104] The controller performs multi-parameter logic decision-making and power topology mode switching operations with dual hysteresis dead-time control constraints to prevent frequent state switching, including the following steps: S505, the controller assesses the electrical risks of frequent state switching induced by fluid disturbances. Under conditions where the airflow inside the main flue experiences continuous fluid turbulence or localized abrupt changes in dust concentration, the calculated equivalent topology characteristics and transient differential impedance fluctuations are prone to high-frequency fluctuations near the set boundary thresholds. If only a single critical threshold is relied upon for state determination, the control system will frequently issue topology switching commands. This frequent switching not only causes high-frequency thermal stress and voltage / current spikes to the power semiconductor devices inside the high-frequency inverter power supply unit, but also triggers transient overvoltages in the power supply circuit, thereby increasing the electrical risk of triggering hardware protection actions and causing oscillation deadlock. To suppress this electrical risk of frequent state switching, the controller introduces a dual hysteresis dead-zone control constraint in the decision matrix to prevent tripping oscillations.

[0105] S506, the controller constructs a first-level hysteresis dead zone condition based on the cross-state deviation difference. The controller sets corresponding bias values ​​based on the high-limit threshold of the dust-rich, low-resistivity state and the differential impedance fluctuation rate threshold to establish a buffer region for state exit. The controller performs addition and subtraction operations on the originally set thresholds and bias values ​​to generate the hysteresis recovery threshold for exiting the corresponding state. The formula for calculating the morphological hysteresis recovery threshold for exiting the low-resistivity risk state is: ; In the formula, The morphological hysteresis recovery threshold represents the threshold for exiting a low-resistance risk state. This indicates the upper limit threshold of the dust-rich, low-resistivity state mentioned above; This indicates the set shape hysteresis bias.

[0106] Similarly, the formula for calculating the fluctuation hysteresis recovery threshold for exiting a low-resistance risk state is: ; In the formula, This represents the volatility hysteresis recovery threshold for exiting a low-resistance risk state. This represents the aforementioned threshold for the volatility of the differential impedance; This indicates the set hysteresis bias.

[0107] Combining the hysteresis characteristics of conventional control systems, the morphological hysteresis bias The value range is set to 5% to 15% of the upper limit threshold value of the dust-rich, low-resistivity state; fluctuation hysteresis bias. The value range is set to 5% to 15% of the differential impedance volatility threshold. Only when the control parameter crosses this cross-state deviation difference, i.e., the equivalent morphological characteristics decrease to a certain value... Below, and the transient differential impedance variability decreases to The controller only determines that the first hysteresis dead zone condition is met when the following conditions are met. This threshold separation setting allows the system to filter out small high-frequency disturbances in the parameters near the boundary. For the exit logic of the high-impedance state, the controller uses the same principle, adding the corresponding morphological hysteresis bias to the pure gray high-impedance state low threshold and the corresponding fluctuation hysteresis bias to the fluctuation low threshold, to generate the various hysteresis recovery thresholds for exiting the high-impedance state.

[0108] S507, the controller constructs a second hysteresis dead zone condition based on time duration. Upon satisfying the first hysteresis dead zone condition, the controller activates its internal anti-jitter timer. The controller records the duration for which the equivalent topographic features and transient differential impedance fluctuations remain continuously within the effective range of the hysteresis recovery threshold, defining this as the state out-of-limit duration.

[0109] The controller has a preset time-based anti-jitter threshold, which is based on the transient discharge time constant of the energy storage element inside the high-frequency inverter power supply unit. Considering the discharge characteristics of the large-capacity filter capacitor, the threshold value is set to range from 0.5s to 2.0s. The controller compares the duration of the out-of-limit state with the time-based anti-jitter threshold in real time. When the out-of-limit state duration is greater than or equal to the time-based anti-jitter threshold, the controller determines that the second hysteresis dead zone condition is met. When the out-of-limit state duration is less than the time-based anti-jitter threshold but the parameter reverses and exceeds the limit again, the timer is reset.

[0110] S508: The controller executes state recovery logic with dual hysteresis dead zones. When the electric field was originally in a low-resistance risk state, the controller only issues a mode switching command after both the first and second hysteresis dead zone conditions are met simultaneously. The controller unshields the pulse superposition unit, allowing it to re-output transient high-voltage pulses, and restores the operating frequency of the high-frequency inverter power supply unit to the reference value of the normal detection range.

[0111] When the electric field is initially in a high-resistance state, the controller determines that the system has exited the high-resistance state only after the equivalent morphological characteristics exceed the morphological hysteresis recovery threshold for exiting the high-resistance state, the transient differential impedance fluctuation rate exceeds the fluctuation hysteresis recovery threshold for exiting the high-resistance state, and the duration of the state exceeding the limit reaches the time-duration anti-jitter threshold. The controller then restores the base DC voltage output by the high-frequency inverter power supply unit to its normal value and drives the pulse superposition unit according to the reconstruction parameters. The aforementioned hysteresis dead-zone limiting algorithm, which introduces both cross-state deviation difference and time duration as dual judgment conditions, establishes an asymmetric state entry and exit mechanism, thereby maintaining the continuous power supply capability of the inverter power supply under disturbance conditions.

[0112] Herein, the reference value of the operating frequency is the optimal resonance frequency of the high-frequency variable-frequency power supply unit that matches the transformer leakage inductance under the conventional dust removal working condition, with its value range set to 10kHz to 50kHz, which is specifically determined by the control system through frequency sweep optimization during the initial power-on self-test. The conventional value of the basic DC voltage is the maximum stable operating average secondary voltage of the corresponding electric field under the current flue gas dust load and without spark breakdown, with its value range set to 40kV to 80kV, and this value is dynamically recorded and established by the electrostatic precipitator through volt-ampere characteristic curve scanning and corona critical tracking algorithm under the conventional operating state.

[0113] To further understand the present invention, the dynamic control method for the electrostatic precipitation variable-frequency power supply based on AI recognition of flue gas dust characteristics of the present invention will be further described below in combination with specific application embodiments.

[0114] Taking the electrostatic precipitator supporting a 300MW unit of a coal-fired power plant as an example, the specific execution logic and calculation process of the present invention are as follows: The controller reads the historical dust hopper sampling and test data of the precipitator in the unit control system to classify and mark the dust phase state, and extracts the inflection point of the voltage and current characteristic reverse fold. The gradient descent algorithm is used for iterative training of historical data to construct a multiple linear regression model. The finally calibrated generated aspect ratio term weight , the edge gradient term weight , and the fitting bias constant . Combined with the statistical distribution of historical data, the high limit threshold of the dust carbon-rich low-resistance state is set to 15, and the differential impedance fluctuation rate threshold is set to 0.4.

[0115] The heat tracing sampling module introduces the dust airflow in the main flue into the optical measurement chamber at a constant speed. Within the visual sampling time window, the system continuously acquires 2D original microscopic image frames, and the controller uses the AI edge detection model to perform abnormal particle rejection. The ratio of the number of pixels of the long side to the short side of the minimum enclosing rectangle of the retained effective single particle is extracted as the real-time aspect ratio, according to the formula , the system accumulates and sums the squared differences between the real-time aspect ratios of all effective single particles within the visual sampling time window and their arithmetic mean , and then divides by the total number of effective single particles . At the same time, the overall edge pixel gradient mean is calculated . Substitute the above values into the formula to obtain the dust morphology irregularity index .

[0116] ​The controller, combining the average flow velocity of the main flue section, calculates the theoretical flight time of the dust gas flow to each electric field level, and subtracts the inherent processing delay of the system to obtain the control time margin. The second-level electric field, satisfying the lower threshold of the response time margin, is selected as the target electric field. The fluid diffusion coefficient is estimated by looking up a table using the local Reynolds number, and the fluid diffusion tolerance is calculated to generate probability arrival time windows. After overlapping and smoothing of adjacent probability arrival time windows, the controller generates the equivalent topographic features. .

[0117] The reference pulse width set by the control system Pulse width modulation coefficient Through the formula Calculate the reconstructed probe pulse width The set reference injection frequency Frequency modulation coefficient Through the formula Calculate the reconstructed probe pulse injection frequency The pulse superposition unit injects probe pulses into the second-stage electric field according to the reconstruction parameters. The isolation sampling module synchronously acquires signals and extracts the maximum absolute value of impedance from the transient impedance absolute value time series. Extreme value of minimum impedance Substitute the zero bias constant into the equation. The transient differential impedance ripple rate was calculated. .

[0118] The controller inputs the equivalent topographic feature 17.7 and the transient differential impedance fluctuation rate 0.500 into the multi-parameter coupled logic decision matrix. Since 17.7 is greater than the upper limit threshold of 15 for the dust-rich carbon low-resistivity state, and 0.500 is greater than the differential impedance fluctuation rate threshold of 0.4, the controller determines that the current second-level electric field is in a low-resistivity risk state. Under the condition of satisfying the dual hysteresis dead-zone control constraints, the system sends a low-level blocking signal to the corresponding composite power supply module to shield the drive output of the pulse superposition unit; at the same time, the operating frequency of the high-frequency inverter power supply unit is switched and maintained at the safe upper limit operating frequency of 20kHz, switching to pure DC power supply topology mode.

[0119] To verify the reliability and dynamic adjustment effect of the control system of the present invention in actual operation, a continuous operation period of 72 hours was selected as the experimental period. The first 36 hours adopted the traditional fixed parameter control method, and the latter 36 hours adopted the dynamic control method of the present invention based on AI recognition of flue gas dust characteristics. The key control parameters during the operation were extracted and compared.

[0120] See attached document Figure 8The graph shows that the horizontal axis represents the equivalent topographic features, with values ​​ranging from 0 to 25; the vertical axis represents the detection pulse width, in microseconds, with values ​​ranging from 10 to 30 microseconds. The solid line in the graph illustrates the dynamic mapping relationship between the two. As the value of the equivalent topographic features on the horizontal axis increases, the value of the detection pulse width on the vertical axis shows a linear decreasing trend.

[0121] The control system internally establishes an adaptive negative feedback modulation logic for detecting the pulse width, and the corresponding calculation formula is as follows: When the equivalent morphology characteristic value is too large, it indicates that the dust particles in the current electric field space are rich in irregular carbon particles or have high conductivity, and the space charge distribution between the plates is uneven, which increases the probability of inducing local micro-discharge. In order to improve the time domain resolution of electrical condition detection under such conditions and prevent the omission of intermittent discharge precursor signals, the controller calculates and reconstructs a narrower detection pulse width according to the above formula, forming a linear negative correlation adjustment dynamic.

[0122] See attached document Figure 9 The horizontal axis of the graph represents operating time in hours, ranging from 0 to 80; the vertical axis represents transient differential impedance fluctuation rate, ranging from 0.1 to 0.6. The dashed lines represent operating data using the traditional control method over the first 36 hours, while the solid lines represent operating data using the control method of this invention over the period from 36 to 72 hours.

[0123] During the dashed line phase (0-36 hours), the transient differential impedance fluctuation rate remained at a relatively high level and exhibited oscillations, frequently fluctuating between 0.2 and 0.55. During the solid line phase (36-72 hours), the transient differential impedance fluctuation rate decreased and remained within a low range of 0.15 to 0.25. Traditional control methods lack the introduction of feedforward visual parameters, relying solely on post-event electrical parameter feedback. This results in a delayed response to changes in the dust phase, leading to frequent localized micro-spark discharges between electrodes and causing severe impedance fluctuations.

[0124] This invention acquires equivalent topographic features in real time and inputs them, along with transient differential impedance fluctuation rate, into a multi-parameter coupled logic decision matrix to determine a low-resistance risk state or a high-resistance state in the early stages of fluctuation. In the face of dust disturbances, the controller issues a mode-switching command, dynamically changing the output electrical topology mode of the high-frequency inverter power supply unit and the pulse superposition unit. Simultaneously, the system's dual hysteresis dead-zone control constraint, designed to prevent frequent state switching, filters out minor high-frequency disturbances near the boundary, avoiding false triggering and hardware oscillation deadlock. This effectively smooths the transient differential impedance of the electric field within a lower stable range, preventing continuous arc breakdown.

[0125] While specific embodiments of the present invention have been described above, those skilled in the art should understand that these specific embodiments are merely illustrative. Those skilled in the art can omit, substitute, and modify the details of the above methods and systems in various ways without departing from the principles and essence of the present invention. For example, combining the above method steps to perform substantially the same function and achieve substantially the same result using substantially the same method falls within the scope of the present invention.

Claims

1. A dynamic control method for electrostatic precipitator variable frequency power supply based on AI recognition of flue gas dust characteristics, characterized in that, Includes the following steps: Historical parameters are obtained for benchmark calibration, and the generated upper limit threshold of dust rich carbon low resistivity state, lower limit threshold of pure ash high resistivity state, and differential impedance fluctuation rate threshold are written into the storage medium. Two-dimensional raw microscopic image frames are acquired, and the controller uses the deployed AI edge detection model to read the two-dimensional raw microscopic image frames and calculate the dust morphology irregularity index; Selected A level electric field is generated to form a probability arrival time window. When the probability arrival time windows of adjacent batches are independent of each other, the dust morphology irregularity index is used as the equivalent morphology feature. When the probability arrival time windows of adjacent batches overlap in the time domain, the equivalent morphology feature is calculated based on the dust morphology irregularity index. Based on the equivalent morphological features, reconstruction parameters are generated. When the time progresses to the probability arrival time window, the parameters are then sent to the first... A detection pulse is injected into the electric field to simultaneously acquire transient voltage and current signals and calculate the transient differential impedance fluctuation rate. The equivalent morphological features and the transient differential impedance fluctuation rate are input into a multi-parameter coupled logic decision matrix. Combined with the dust carbon-rich low-resistivity high limit threshold, the pure ash high-resistivity low limit threshold and the differential impedance fluctuation rate threshold output mode switching command, the output electrical topology mode of the high-frequency variable power supply unit and the pulse superposition unit is changed.

2. The dynamic control method for electrostatic precipitator variable frequency power supply based on AI recognition of flue gas dust features according to claim 1, characterized in that, The process of acquiring historical parameters for benchmark calibration, and writing the generated high-limit threshold for the carbon-rich low-resistivity state of dust, the low-limit threshold for the high-resistivity state of pure ash, and the differential impedance fluctuation rate threshold into the storage medium includes: Historical dust collector ash hopper sampling and testing data were obtained to classify and label dust phases, and a multiple linear regression model was constructed to perform control parameter benchmark calibration and output aspect ratio term weights and edge gradient term weights. The trained multiple linear regression model is used to perform inference calculations on carbon-rich low-resistivity samples and pure ash high-resistivity samples, and the values ​​corresponding to the confidence intervals are extracted to adjust the upper limit threshold of the carbon-rich low-resistivity dust and the lower limit threshold of the pure ash high-resistivity dust. Historical voltage and current characteristic data are read to extract the critical impedance reference parameters of the anti-corona, the basic impedance fluctuation rate under the critical state of anti-corona is calculated, and a safety margin coefficient is introduced for derating correction to generate the differential impedance fluctuation rate threshold.

3. The dynamic control method for electrostatic precipitator variable frequency power supply based on AI recognition of flue gas dust features according to claim 2, characterized in that, The process of obtaining historical dust collector ash hopper sampling and testing data for dust phase classification and labeling includes: Extract the resistivity and loss on ignition values ​​for the same time period to construct a dust phase feature vector; When the loss on ignition value in the dust phase feature vector is greater than the loss on ignition threshold and the resistivity value is lower than the resistivity critical value, the corresponding historical dust collector ash hopper sampling and testing data is marked as the carbon-rich low-resistivity sample and the phase fitting target label value is set to 1. When the loss on ignition value in the dust phase characteristic vector is lower than the loss on ignition threshold and the resistivity value is higher than the resistivity critical value, the corresponding historical dust collector ash hopper sampling and testing data is marked as the pure ash high resistivity state sample and the phase fitting target label value is set to -1.

4. The dynamic control method for electrostatic precipitator variable frequency power supply based on AI recognition of flue gas dust features according to claim 2, characterized in that, The process of acquiring the original two-dimensional microscopic image frame, and the controller using the deployed AI edge detection model to read the original two-dimensional microscopic image frame and calculate the dust morphology irregularity index, includes: The heat-tracing sampling module introduces the dust airflow in the main flue into the optical measurement cavity at a constant speed, and the stroboscopic light source and the microscopic imaging component acquire the original two-dimensional microscopic image frame. The controller uses the AI ​​edge detection model to read the two-dimensional original microscopic image frame, converts the two-dimensional original microscopic image frame into a single-channel feature tensor matrix, inputs it into the AI ​​edge detection model, and outputs a pixel-level edge probability map. The adaptive truncation threshold of the pixel-level edge probability map is determined and converted into a binary contour map. All closed regions are extracted as candidate particles. Abnormal particles that are out of focus or overlapped are removed based on the pixel gradient mean and convex hull area ratio of the edge coordinate set. The effective single particles are retained and the dust morphology irregularity index is calculated.

5. The dynamic control method for electrostatic precipitator variable frequency power supply based on AI recognition of flue gas dust features according to claim 4, characterized in that, The calculation of the dust morphology irregularity index using the effective monomer particles includes: The temporal statistical variance of the real-time aspect ratio of all effective individual particles within the visual sampling time window is calculated. Calculate the gradient magnitude of all effective single-unit particle edge pixels in the original grayscale image within the visual sampling time window and obtain the overall edge pixel gradient mean; The dust morphology irregularity index is calculated and output by mathematically fusing the temporal statistical variance and the mean gradient of the overall edge pixels with the aspect ratio term weight, the edge gradient term weight and the fitting bias constant using a linear model.

6. The dynamic control method for electrostatic precipitator variable frequency power supply based on AI recognition of flue gas dust features according to claim 1, characterized in that, The selected number The electric field is leveled and a probability arrival time window is generated, including: Based on the average flow velocity of the main flue section, the dust airflow reaches the first... The theoretical flight time at the entrance of the electric field; The time difference between the completion of recording the exposure of the microscopic image frame and the output of the dust morphology irregularity index is set as the inherent processing delay of the system. Subtracting the theoretical flight time from the inherent processing delay of the system generates a value for the first... The control time margin of the electric field is determined by selecting the minimum electric field level index corresponding to a control time margin greater than or equal to the lower threshold of the response time margin, and the corresponding independent electric field is selected as the first-order electric field. The electric field is used to generate the probability arrival time window based on the theoretical flight time.

7. The dynamic control method for electrostatic precipitator variable frequency power supply based on AI recognition of flue gas dust features according to claim 6, characterized in that, The step of calculating the equivalent morphology features based on the dust morphology irregularity index when there is temporal overlap in the probability arrival time windows of adjacent batches includes: The product of the fluid diffusion coefficient and the theoretical flight time is used to generate the fluid diffusion tolerance and establish the boundary of the probability arrival time window; If the start time of the probability arrival time window of the current batch is less than the end time of the probability arrival time window of the previous batch, it is determined that the two probability arrival time windows overlap in the time domain. During the overlapping period, the dust morphology irregularity index of all independent batches involved in the overlap is aggregated and the average state is extracted as the equivalent morphology feature. The statistical fluctuation degree of deviation from the equivalent morphology feature is calculated to generate the equivalent local variance.

8. The dynamic control method for electrostatic precipitator variable frequency power supply based on AI recognition of flue gas dust features according to claim 1, characterized in that, Based on the equivalent morphological features, reconstruction parameters are generated and directed to the first... A level electric field is injected with a probe pulse to simultaneously acquire transient voltage and current signals and calculate the transient differential impedance fluctuation rate, including: The reconstruction parameters are generated by packaging the probe pulse width and probe pulse injection frequency together based on the equivalent morphological features; The pulse superposition unit directs the pulse to the first according to the reconstruction parameters. The detection pulse is injected into the electric field, and the isolation sampling module simultaneously acquires the transient voltage signal and the transient current signal; Extract the transient voltage signal change and the transient current signal change, calculate the current change rate between two adjacent sampling points, and when the current change rate is less than the current change rate safety threshold, introduce a perturbation constant to prevent the denominator from approaching zero. Combine the transient current signal change and the perturbation constant to prevent the denominator from approaching zero as the denominator, and perform a division operation between the transient voltage signal change and the denominator to obtain the transient differential impedance to prevent the denominator from approaching zero. Extract the maximum and minimum absolute values ​​of impedance within the duration of the probe pulse, and perform a normalization algorithm to obtain the transient differential impedance fluctuation rate constrained within the numerical range boundary.

9. The dynamic control method for electrostatic precipitator variable frequency power supply based on AI recognition of flue gas dust features according to claim 1, characterized in that, The equivalent morphological features and the transient differential impedance fluctuation rate are input into a multi-parameter coupled logic decision matrix. Combined with the high-limit threshold of the dust-rich carbon-rich low-resistivity state, the low-limit threshold of the pure ash high-resistivity state, and the differential impedance fluctuation rate threshold, an output mode switching command is used to change the output electrical topology mode of the high-frequency inverter power supply unit and the pulse superposition unit, including: A two-dimensional state space is established by using the equivalent morphological features as the first input dimension and the transient differential impedance fluctuation rate as the second input dimension. In the two-dimensional state space, when the equivalent morphological feature is greater than or equal to the upper limit threshold of the dust-rich carbon low-resistivity state and the transient differential impedance fluctuation rate is greater than or equal to the differential impedance fluctuation rate threshold, it is determined to be in a low-resistivity risk state. A low-level blocking signal is sent to shield the pulse superposition unit from working and change the operating frequency of the high-frequency inverter power supply unit to maintain the safe upper limit operating frequency and switch to pure DC power supply topology mode. In the two-dimensional state space, when the equivalent morphological feature is less than or equal to the low threshold of the pure gray high-resistivity state and the transient differential impedance fluctuation rate is less than or equal to the preset fluctuation low threshold, it is determined to be in a high-resistivity state. The base DC voltage output by the high-frequency inverter power supply unit is reduced to the corona initiation voltage critical point, the shielding of the pulse superposition unit is removed, and the transient high-voltage pulse is superimposed according to the output pause time interval including the empirical charge discharge time constant and margin time of the high resistivity dust.

10. The dynamic control method for electrostatic precipitator variable frequency power supply based on AI recognition of flue gas dust features according to claim 9, characterized in that, A dual hysteresis dead-zone control constraint to prevent tripping oscillations is introduced into the multi-parameter coupled logic decision matrix, including: Based on the high limit threshold of the dust-rich carbon low-resistivity state, the low limit threshold of the pure ash high-resistivity state, the differential impedance fluctuation rate threshold, and the preset fluctuation low limit threshold, a morphology hysteresis recovery threshold and a fluctuation hysteresis recovery threshold corresponding to exiting the low-resistivity risk state or the high-resistivity state are generated by bias operation to construct a cross-state deviation difference. When crossing the cross-state deviation difference, the first hysteresis dead zone condition is satisfied. The duration for which the equivalent topographic feature and the transient differential impedance fluctuation rate are continuously maintained at the effective side of the topographic hysteresis recovery threshold and the fluctuation hysteresis recovery threshold by starting the anti-jitter timer is defined as the state over-limit duration. The duration of the state exceeding the limit is compared with the time-based stabilization threshold in real time. When the duration of the state exceeding the limit is greater than or equal to the time-based stabilization threshold, the second hysteresis dead zone condition is met.

11. The dynamic control method for electrostatic precipitator variable frequency power supply based on AI recognition of flue gas dust features according to claim 10, characterized in that, Execute state recovery logic that includes a double hysteresis dead zone, including: When the original low-impedance risk state is met, the shielding of the pulse superposition unit is removed after the first hysteresis dead zone condition and the second hysteresis dead zone condition are met simultaneously, allowing the transient high voltage pulse to be output again, and the operating frequency of the high-frequency inverter power supply unit is restored to the preset operating frequency reference value. When the original state is in the high-impedance state, after the equivalent morphological feature is greater than the morphological hysteresis recovery threshold for exiting the high-impedance state, the transient differential impedance fluctuation rate is greater than the fluctuation hysteresis recovery threshold for exiting the high-impedance state, and the state over-limit duration reaches the time duration anti-jitter threshold, it is determined that the high-impedance state has been exited. The base DC voltage output by the high-frequency inverter power supply unit is restored to the preset base DC voltage normal value, and the pulse superposition unit is driven according to the reconstruction parameters.