Electrostatic dust removal smoke emission dynamic modeling method based on multi-parameter coupling analysis

Abstract

The invention relates to the technical field of monitoring modeling, in particular to an electrostatic precipitation smoke emission dynamic modeling method based on multi-parameter coupling analysis, which comprises the following steps: calculating electric field rate and acceleration through sliding window difference, judging electric field abrupt change trigger parameter adjustment, normalizing particle size and charge density, and mapping disturbance intensity index. And calculating response delay and a gradient field to generate a disturbance coefficient, transferring a model weight when the disturbance exceeds a threshold value, performing fitting reconstruction, and performing regression weight reverse normalization to output a steady-state parameter group. According to the method, electric field dynamic characteristics are captured through equal-interval sampling in combination with a sliding window, an acceleration super-threshold model is constructed through second-order difference, a dynamic adjustment mechanism is triggered, particle size and charge density are normalized, disturbance intensity is mapped through linear interpolation, speed and density abrupt change are quantized through gradient weighting, and weight transfer and a least square method are introduced to reconstruct parameters. The transient analysis capability and robustness of the model are improved, and high-precision synchronous mapping of the emission trend and the equipment state is realized.

Description

Technical Field

[0001] This invention relates to the field of monitoring and modeling technology, and in particular to a dynamic modeling method for electrostatic precipitator flue gas emissions based on multi-parameter coupling analysis. Background Technology

[0002] The field of monitoring and modeling technology involves the real-time acquisition, analysis, and modeling of data from the environment, industrial processes, and related physicochemical processes. Core aspects include data acquisition technology, modeling methods, dynamic monitoring mechanisms, and parameter fitting techniques. The aim is to construct a mathematical model that accurately reflects the operating state of the target system, providing a basis for control and decision-making. In this technical field, the systematic technical path covers sensor deployment and data acquisition, modeling algorithm selection and parameter setting, model verification and optimization, and other aspects.

[0003] Among them, the traditional dynamic modeling method for flue gas emissions from electrostatic precipitators refers to the modeling problem of flue gas concentration changing over time during electrostatic precipitator operation. The method adopted is a curve fitting method based on the original operating data. By collecting data such as inlet and outlet flue gas concentration, voltage and current values, and airflow velocity, a first-order or second-order difference model is used to perform numerical regression to establish a predictive relationship, or a multinomial regression model or grey prediction model is used to describe the emission trend, so as to approximately characterize the dynamic behavior of flue gas emissions during the operation of electrostatic precipitator equipment.

[0004] Traditional modeling methods rely on static regression analysis of raw data, employing fixed-order difference models and pre-defined polynomial structures. These methods cannot analyze the instantaneous acceleration changes in electric field parameters. When equipment operating conditions fluctuate drastically, model parameter updates lag behind real-time physical processes, leading to systematic biases in flue gas concentration prediction. Curve fitting methods based on raw data lack dynamic response mechanisms, making it difficult to characterize transient processes where electric field acceleration continuously exceeds thresholds. They also fail to adequately analyze the dynamic coupling effect between particle size distribution and charge density. Furthermore, model weight allocation strategies are limited by initial settings, and long-term operation can easily lead to parameter drift, affecting the accuracy and timeliness of emission control decisions. Summary of the Invention

[0005] To address the shortcomings of traditional modeling methods that rely on static regression analysis based on raw data and employ fixed-order difference models and preset polynomial structures, which fail to analyze the instantaneous acceleration changes in electric field parameters and cause systematic biases in dust concentration predictions when equipment operating conditions fluctuate drastically, curve fitting methods based on raw data lack dynamic response mechanisms, making it difficult to characterize transient processes where electric field acceleration continuously exceeds thresholds. Furthermore, they fail to adequately analyze the dynamic coupling effect between particle size distribution and charge density. The model weight allocation strategy is also limited by initial settings, leading to parameter drift over long-term operation and impacting the accuracy and timeliness of emission control decisions. This invention provides a dynamic modeling method for electrostatic precipitator dust emissions based on multi-parameter coupling analysis. The technical solution is as follows:

[0006] On the one hand, a dynamic modeling method for electrostatic precipitator flue gas emissions based on multi-parameter coupled analysis is provided, which includes:

[0007] S1: The electric field intensity signal is obtained by sampling at equal intervals. The time window data is processed by a sliding window mechanism. The electric field change rate is calculated by first-order difference and the electric field acceleration is calculated by second-order difference. When the acceleration of three consecutive windows exceeds the acceleration judgment threshold and the rate direction is consistent, the electric field change judgment result is generated.

[0008] S2: Call the electric field sudden change judgment result to trigger the parameter adjustment command, perform normalization processing on the dust particle size distribution data and charge density peak value, use linear interpolation to map the disturbance intensity, perform exponential weighting calculation according to the current pulse frequency offset, and output dynamic response parameter group;

[0009] S3: Based on the response delay coefficient in the dynamic response parameter group, the Sobel operator is used to calculate the velocity gradient field and density gradient field of the electrode edge region, and a weighted superposition operation is performed to calculate the linear combination of velocity fluctuation amplitude, directional offset angle and density abrupt change rate to generate the disturbance intensity coefficient.

[0010] S4: When the disturbance intensity coefficient exceeds the weight transfer threshold, the model weights are transferred to the particle size offset angle and the local electric field distribution vector. The offset angle and the electric field distribution vector are fitted using the least squares method to generate a reconstructed weight allocation table.

[0011] As a further aspect of the present invention, the electric field abrupt change determination result is used to trigger the calculation of subsequent disturbance response parameters to ensure the model's dynamic self-matching adjustment capability;

[0012] The electric field abrupt change determination results include acceleration threshold, velocity direction unidirectionality, and window trigger count. The dynamic response parameter set includes disturbance mapping baseline value, frequency offset, and charge density peak value. The disturbance intensity coefficient includes velocity fluctuation amplitude, flow direction deviation angle, and density gradient change rate. The reconstructed weight allocation table includes particle size distribution weight, field strength vector weight, and least squares regression coefficient.

[0013] As a further aspect of the present invention, the specific steps of S1 include:

[0014] S101: The electric field intensity signal is obtained by sampling at equal intervals. The starting position and step size of the sliding window are set, the window is advanced to divide the intensity sequence, the intensity offset between adjacent times in each sequence is calculated, and the electric field change rate sequence is obtained by combining them.

[0015] S102: Call the electric field change rate sequence, perform a second difference operation on the rate values ​​within the time window, construct a second-order difference sequence within the window and calculate the average value to obtain the electric field acceleration change sequence;

[0016] S103: Call the electric field acceleration change sequence, compare the average electric field acceleration in three consecutive sliding windows with the acceleration judgment threshold, filter the time window group where the acceleration exceeds the acceleration judgment threshold and the rate direction is consistent in the electric field change rate sequence, and generate the electric field sudden change judgment result.

[0017] The acceleration judgment threshold is a numerical limit used to determine whether the electric field acceleration is abnormal. It is set by combining the average electric field acceleration of the original data and K times the standard deviation, with K taking the value of 2 or 3 to improve sensitivity.

[0018] The consistent rate direction refers to the consistent positive and negative signs of the electric field change rate within multiple consecutive time windows, indicating that the electric field change trend is continuous and has not reversed.

[0019] As a further aspect of the present invention, the specific steps of S2 include:

[0020] S201: Call the electric field abrupt change judgment result, obtain the dust particle size distribution data and charge density peak data in the corresponding time period, calculate the normalization factor based on the particle size distribution data, perform interval mapping processing on the charge density peak parameter, unify the scale range of the two sets of data, and obtain the normalization parameter set.

[0021] S202: Based on the component values ​​in the normalized parameter group, set a disturbance intensity index reference sequence, use linear interpolation to sequentially map the normalized vector components to the corresponding disturbance intensity intervals, calculate the disturbance response factor set for each corresponding component based on the interpolation results, and generate a disturbance mapping value group.

[0022] S203: Call the disturbance mapping value group, extract the components based on the current current pulse frequency and the data offset of the previous two time windows, perform item-by-item exponential weighting, set the frequency offset exponential coefficient vector as the weighting factor, and perform group-by-group combination operation on the disturbance value and offset data to obtain the dynamic response parameter group.

[0023] As a further aspect of the present invention, the specific steps of S3 include:

[0024] S301: Based on the response delay coefficient in the dynamic response parameter group, extract the velocity and density data of the electrode edge region, input them into the Sobel operator to perform horizontal and vertical gradient calculations, and uniformly map them to a two-dimensional coordinate system to generate velocity density gradient distribution values.

[0025] S302: Call the velocity density gradient distribution value, extract the velocity and density gradient values ​​according to the coordinates, perform weighted fusion according to the weight rules set by the response delay coefficient, traverse the edge pixels to perform superposition operation, and generate velocity density fusion fluctuation value.

[0026] S303: Call the velocity density fusion fluctuation value, identify the amplitude point, extract the direction change trend and gradient slope, calculate the weighted total of amplitude, offset angle and mutation rate, and generate the disturbance intensity coefficient.

[0027] As a further aspect of the present invention, the velocity density gradient distribution value represents the spatial gradient change of velocity and density in the edge region of the electrode plate, revealing local dynamic characteristics;

[0028] The weighted fluctuation intensity obtained by fusing the velocity and density gradients reflects the overall degree of local disturbance.

[0029] The disturbance intensity coefficient is based on the extraction of direction, amplitude, and abrupt change information from the fused fluctuation quantity to quantify the overall intensity of the edge disturbance;

[0030] The Sobel operator is used to extract the horizontal and vertical gradients of the velocity and density matrices. It combines the response delay coefficient to set edge processing and normalization parameters, and after unified mapping, generates gradient distribution values ​​that can be used for fusion calculation.

[0031] As a further aspect of the present invention, the specific steps of S4 include:

[0032] S401: Call the disturbance intensity coefficient. When the disturbance intensity coefficient exceeds the weight transfer threshold, extract all weight parameters in the model and assign each weight parameter to the particle size offset angle and the local electric field distribution vector to generate the weight parameter transfer amount.

[0033] S402: Call the weight parameter transfer amount, combine the weight with the particle size offset angle, pair the values ​​of the particle size offset angle and the local electric field distribution vector, calculate the function expression of the two using the least squares method, extract the regression parameters in the expression, calculate the residual absolute value index, analyze the goodness of fit between the residual vector and the parameters, and generate the particle size and electric field fitting coefficient.

[0034] S403: Based on the particle size and electric field fitting coefficients, organize the data corresponding to the particle size offset angle and the local electric field distribution vector, summarize the weight distribution on the distribution, and obtain the reconstructed weight allocation table.

[0035] As a further aspect of the present invention, the method includes step S5:

[0036] S5: Call the reconstructed weight allocation table to perform weight regression, perform reverse normalization on the airflow velocity, pressure difference, offset angle and electric field distribution vector, use the gradient descent method to calculate the cumulative deviation between the parameter weights and the initial configuration, and output the steady-state modeling parameter set when the deviation is less than the set tolerance.

[0037] The steady-state modeling parameter set includes a weight matrix, residual cumulative amount, and inverse normalization coefficient.

[0038] As a further aspect of the present invention, the specific steps of S5 include:

[0039] S501: Based on the reconstructed weight allocation table, the airflow velocity, pressure difference, offset angle and electric field distribution vector are decomposed into sub-items, the original normalized values ​​are extracted respectively, the reverse normalization operation is performed, the data range of each parameter is gradually restored, and the parameter restoration distribution value is generated.

[0040] S502: Call the parameter to restore the distribution value, use the gradient descent method to progressively correct the parameter weights and the initial configuration weights, monitor the numerical differences between weights in real time, accumulate the weight change range of each parameter, and obtain the weight deviation trend value.

[0041] S503: Based on the weight deviation trend value, the gradient descent method is used to continuously determine the relationship between the cumulative deviation and the set tolerance. When the deviation of the parameter is lower than the set tolerance, the current parameter values ​​are summarized to establish a steady-state modeling parameter group.

[0042] As a further aspect of the present invention, the parameter restoration distribution value is obtained by restoring the normalized parameters to their original physical range values ​​through inverse normalization;

[0043] The weight deviation trend value is a numerical indicator that measures the trend of parameter weight changes corresponding to the initial weight during iterative optimization.

[0044] The gradient descent method iteratively adjusts the parameter weights, gradually reducing the weight deviation and approaching the set tolerance, thereby optimizing the steady-state modeling parameters.

[0045] The beneficial effects of the technical solutions provided in the embodiments of the present invention include at least the following:

[0046] By employing equal-interval time-series sampling combined with a sliding window mechanism to capture the dynamic characteristics of electric field intensity in real time, a continuous acceleration exceeding threshold judgment model is constructed through second-order difference operations to trigger a dynamic parameter adjustment mechanism. Normalization processing of dust particle size distribution and charge density peak data is used to establish a linear interpolation mapping of disturbance intensity index. Based on gradient field weighted superposition operations, velocity fluctuation and density mutation characteristics are quantified. A weight transfer mechanism combined with the least squares method is introduced to realize dynamic reconstruction of model parameters, improving the ability of electrostatic precipitator modeling to analyze transient changes in electric field, enhancing the real-time performance of parameter adjustment and weight allocation, optimizing the robustness of the model under complex disturbance conditions, solving the prediction lag problem of traditional methods under fluctuating operating conditions, and achieving high-precision synchronous mapping between emission trends and equipment operating status. Attached Figure Description

[0047] Figure 1 This is a schematic diagram of the workflow of the present invention;

[0048] Figure 2 This is a detailed flowchart of S1 of the present invention;

[0049] Figure 3 This is a detailed flowchart of the S2 process of the present invention;

[0050] Figure 4 This is a detailed flowchart of the S3 process of the present invention;

[0051] Figure 5 This is a detailed flowchart of the S4 process of the present invention;

[0052] Figure 6 This is a detailed flowchart of S5 of the present invention. Detailed Implementation

[0053] The technical solution of the present invention will now be described with reference to the accompanying drawings.

[0054] In embodiments of the present invention, words such as "exemplarily," "for example," etc., are used to indicate that something is an example, illustration, or description. Any embodiment or design described as "exemplary" in the present invention should not be construed as being more preferred or advantageous than other embodiments or designs. Specifically, the use of the word "exemplary" is intended to present the concept in a concrete manner. Furthermore, in embodiments of the present invention, the meaning expressed by "and / or" can be both, or either one.

[0055] In the embodiments of this invention, the terms "image" and "picture" may sometimes be used interchangeably. It should be noted that, without emphasizing the distinction between them, their intended meanings are consistent. Similarly, the terms "of," "corresponding (relevant)," and "corresponding" may sometimes be used interchangeably. It should be noted that, without emphasizing the distinction between them, their intended meanings are consistent.

[0056] In this embodiment of the invention, sometimes a subscript such as W1 may be written in a non-subscript form such as W1. When the difference is not emphasized, the meaning they express is the same.

[0057] To make the technical problems, technical solutions and advantages of the present invention clearer, a detailed description will be given below in conjunction with the accompanying drawings and specific embodiments.

[0058] Please see Figure 1 This invention provides a method for dynamic modeling of electrostatic precipitator flue gas emissions based on multi-parameter coupling analysis. The processing flow of this method may include the following steps:

[0059] S1: The electric field intensity signal is obtained by sampling at equal intervals. The time window data is processed by a sliding window mechanism. The electric field change rate is calculated by first-order difference and the electric field acceleration is calculated by second-order difference. When the acceleration of three consecutive windows exceeds the acceleration judgment threshold and the rate direction is consistent, the electric field change judgment result is generated.

[0060] S2: Call the electric field sudden change judgment result to trigger the parameter adjustment command, perform normalization processing on the dust particle size distribution data and charge density peak value, use linear interpolation to map the disturbance intensity, perform exponential weighting calculation according to the current pulse frequency offset, and output dynamic response parameter group;

[0061] S3: Based on the response delay coefficient in the dynamic response parameter set, the Sobel operator is used to calculate the velocity gradient field and density gradient field of the electrode edge region, and a weighted superposition operation is performed to calculate the linear combination of velocity fluctuation amplitude, directional offset angle and density abrupt change rate to generate the disturbance intensity coefficient.

[0062] S4: When the perturbation intensity coefficient exceeds the weight transfer threshold, the model weights are transferred to the particle size offset angle and the local electric field distribution vector. The least squares method is used to fit the offset angle and the electric field distribution vector to generate a reconstructed weight allocation table.

[0063] S5: Call the reconstructed weight allocation table to perform weight regression, perform reverse normalization on the airflow velocity, pressure difference, offset angle and electric field distribution vector, use the gradient descent method to calculate the cumulative deviation between the parameter weights and the initial configuration, and output the steady-state modeling parameter set when the deviation is less than the set tolerance.

[0064] The electric field abrupt change determination results include acceleration threshold, velocity direction unidirectionality, and window trigger count. The dynamic response parameter set includes disturbance mapping baseline value, frequency offset, and charge density peak value. The disturbance intensity coefficient includes velocity fluctuation amplitude, flow direction deviation angle, and density gradient change rate. The reconstructed weight allocation table includes particle size distribution weight, field strength vector weight, and least squares regression coefficient. The steady-state modeling parameter set includes weight matrix, residual accumulation, and inverse normalization coefficient.

[0065] Specifically, such as Figure 2 As shown, the specific steps of S1 are as follows:

[0066] S101: The electric field intensity signal is obtained by sampling at equal intervals. The starting position and step size of the sliding window are set, the window is advanced to divide the intensity sequence, the intensity offset between adjacent times in each sequence is calculated, and the electric field change rate sequence is obtained by combining them.

[0067] In the process of acquiring electric field intensity signals through equally spaced time-series sampling, the sampling period Δt must first be preset. For example, an electric field monitoring probe is installed at the outlet of an electrostatic precipitator in a thermal power plant. It continuously samples the electric field intensity at 20ms intervals, forming an intensity time series {Et}. A 30s data segment can be collected to construct an initial dataset {E0, E1, ..., Et}. 1499 Then, the sliding window length L = 10 is set, corresponding to a time period of 200ms, and the sliding step size S = 1, indicating that each time the window slides, the next moment is taken as the starting point, generating {W1, W2, ..., W...}. n}, where W i ={E i E i+1 , ..., E i+9 For each sequence, the offset ΔE of the electric field intensity at adjacent time points is calculated sequentially, such as for window W. i ={35.0, 34.5, 34.7, ..., 33.8} (unit: kV / m), calculate the difference sequence as ΔE = {E i+1 -E i E i+2 -E i+1 , ..., E i+9 -E i+8}, i.e., {-0.5, 0.2, ..., -0.4}, this process generates the electric field change rate sequence {v} by successively accumulating adjacent data pairs within the sliding window through difference operations. i}, where each v i This represents the arithmetic mean or maximum offset value of ΔE within the corresponding sliding window. If set as the mean, then:

[0068]

[0069] If ΔE within a certain window is {-0.5, 0.2, 0.1, -0.3, 0.0, -0.2, -0.1, 0.3, -0.4}, then In monitoring, to more clearly characterize the dynamic features of changes in electric field intensity, the rate of change of the window can be recorded in an array {v1, v2, ..., v...}. n In}, as shown below:

[0070] Table 1: Sample Table of Electric Field Intensity Change Rate

[0071]

[0072]

[0073] Table 1 shows the electric field intensity change rate results for some real-time data windows. The change rate is calculated based on the sample electric field intensity data. For example, in window 3, the input sequence is {34.2, 34.4, 34.5, ..., 35.0}, corresponding to a rate of 0.03 kV / m. The rate is extracted from the entire sequence by sliding windows one by one, further establishing a time-series model of the electric field's dynamic behavior. This model can be used for subsequent acceleration calculations and electric field abrupt change identification. The parameters used in the above calculations, such as the sliding window length L = 10 and the step size S = 1, can be adjusted according to the equipment response time. Generally, in real-time scenarios of thermal power plant dust removal, a sliding window of no more than 200 ms can respond promptly to electric field abrupt changes. Furthermore, determining whether there is a drastic change in a certain window can be done by setting a rate change threshold, for example, setting the threshold to |v... i A value >0.15kV / m is considered a rate mutation. Combining this criterion with real-time scenarios, a dynamic response can be made to sudden changes in flue gas concentration, power fluctuations, or plate discharge. This method not only refines the change process of the electric field signal, but also lays an initial foundation for subsequent electric field acceleration calculation and mutation identification.

[0074] S102: Call the electric field change rate sequence, perform a second difference operation on the rate values ​​within the time window, construct a second-order difference sequence within the window and calculate the average value to obtain the electric field acceleration change sequence;

[0075] After calling the electric field change rate sequence, a quadratic difference operation needs to be performed between adjacent points based on the rate values ​​within each sliding window. In practice, the window length is initially set to 10 points, corresponding to the electric field change rate sequence {v1, v2, ..., v...}. 10 The calculation method involves taking the second-order difference of every three consecutive rate points within the window, i.e., calculating Δ. 2 v j =v j+2 -2v j+1 +vj To obtain the second-order difference sequence {Δ 2 v1, Δ 2 v2, ..., Δ 2 If the velocity values ​​within a certain window are {-0.1, -0.08, 0.03, 0.05, -0.02, -0.01, 0.04, 0.06, 0.02, -0.03} (unit: kV / m), then the calculated difference sequence is {0.15, 0.03, -0.13, -0.12, 0.05, 0.05, -0.10, -0.17}. This difference sequence is then averaged to obtain the estimated value of the change in electric field acceleration.

[0076]

[0077] The above calculation process needs to be repeated in each sliding window advance to obtain the electric field acceleration sequence {a1, a2, ..., a...} corresponding to the time period. n The acceleration sequence allows observation of the drastic changes in the electric field. During real-time dust emission monitoring, if multiple consecutive windows show sustained negative or positive acceleration values ​​with absolute values ​​exceeding a specific threshold, a preliminary assessment can be made regarding a sudden change in the electric field state, such as abnormal discharge or input voltage fluctuations. In real-time operation scenarios, to enhance model robustness, an extreme value removal operation can be introduced into the acceleration estimation. This involves performing median filtering on the second-order difference sequence in each window, retaining only the middle 60% of the samples for weighted averaging. If the difference sequence is {0.15, 0.03, -0.13, -0.12, 0.05, 0.05, -0.10, -0.17}, after removing the highest and lowest values, the following calculation is performed:

[0078]

[0079] In situations where the electric field signal is subject to significant interference, the stability of acceleration estimation can be improved. Each value in the acceleration change sequence is calculated from the difference set of three points in the velocity sequence. Therefore, the reliability of the result depends on the dense sampling and stability of the velocity sequence.

[0080] S103: Call the electric field acceleration change sequence, compare the average electric field acceleration in three consecutive sliding windows with the acceleration judgment threshold, filter the time window group where the acceleration exceeds the acceleration judgment threshold and the velocity direction is consistent in the electric field change rate sequence, and generate the electric field sudden change judgment result.

[0081] After calling the electric field acceleration change sequence, the operation to determine whether the electric field state changes abruptly needs to be performed. This requires analyzing the average electric field acceleration within every three consecutive sliding windows. First, an acceleration judgment threshold T is set. aThe mean of the acceleration sequence based on the initial normal state Its standard deviation σ a The upper and lower boundaries formed by them, that is Where K is 2 or 3 to adjust the sensitivity. If the initial sample acceleration values ​​are {-0.01, 0.00, 0.02, -0.03, 0.01, -0.01, 0.02, 0.03, -0.02, 0.00}, then we have σ a ≈0.0173kV / m, with K=2, the threshold is T. a =0.002 + 2 × 0.0173 = 0.0366 kV / m. Then, the absolute values ​​of the average values ​​for every three consecutive windows in the acceleration sequence are compared. For example, if the average acceleration values ​​for the 5th to 7th windows are {0.02, 0.04, 0.06}, the calculated average acceleration is (0.02 + 0.04 + 0.06) / 3 = 0.04 kV / m, satisfying |0.04| > T. a If the three windows are marked as an abnormal interval, then the velocity directions of the three windows in the corresponding electric field velocity sequence are checked to see if they are consistent, i.e., whether the velocity signs are consistent. For example, if the velocity is {0.12, 0.18, 0.14}, it is considered to be positive growth in the same direction; conversely, if it is {-0.15, -0.19, -0.10}, it is considered to be negative growth in the same direction. If there are different signs, such as {-0.15, 0.18, -0.10}, it is considered to be inconsistent in direction. The consistency criterion here can be completed by checking whether the product of the velocity signs is greater than 0, i.e., v i ·v i+1 >0 and V i+1 ·v i+2 If the velocity direction is consistent and the average acceleration exceeds the threshold, the electric field is determined to have a sudden change. The current three-window time period is recorded as the time period of the sudden change. If the time of the 5th to 7th window in the sample is {100ms, 120ms, 140ms}, the electric field sudden change is marked as "t = 100~140ms". This determination can be used to trigger subsequent dust removal intensity or voltage adjustment strategies. The threshold is set by the two-sided boundary under the assumption of normal distribution. The velocity consistency is determined by the sign consistency judgment. All of these can be accurately operated through the standard Boolean logic structure.

[0082] Specifically, such as Figure 3 As shown, the specific steps of S2 are as follows:

[0083] S201: Call the electric field abrupt change judgment result, obtain the dust particle size distribution data and charge density peak data in the corresponding time period, calculate the normalization factor based on the particle size distribution data, perform interval mapping processing on the charge density peak parameter, unify the scale range of the two sets of data, and obtain the normalization parameter set.

[0084] To retrieve the electric field abrupt change determination result, i.e., to extract the time period information previously marked as abrupt change, it is necessary to first obtain the dust particle size distribution data and charge density peak data recorded within the corresponding time period, and then call the particle size distribution function array {D} respectively. i} and the peak charge density group {ρ i The normalization factor is calculated based on the particle size distribution. During execution, the proportion of particle sizes within each size range is first statistically analyzed. For example, if the particle size distribution obtained during a certain abrupt change period is {0.3μm: 12%, 0.5μm: 25%, 1.0μm: 35%, 2.5μm: 18%, 5.0μm: 10%}, then the normalization factor N is set according to the distribution density weighting method. d Let N d =∑ i w i ·D i , where w i The distribution percentage of particle size points is calculated as follows:

[0085] N d =0.3×0.|2+0.5×0.25+1.0×0.35+2.5×0.18+5.0×0.|0=1.355μm;

[0086] Subsequently, the charge density peak data set {ρ i Perform interval mapping processing, unifying the mapping interval to [0, 1], and adopt the minimum-maximum normalization method, that is, for each charge density value ρ i Calculate ρ i ′=(ρ i -ρ min ) / (ρ max -ρ min The charge density values ​​monitored within the abrupt change range were {0.2, 0.6, 0.5, 0.9, 0.3} μC / cm. 3 For example, we can obtain ρ min =0.2, ρ max =0.9, then the normalized value set is {0.0, 0.571, 0.429, 1.0, 0.143}, combined with the particle size normalization factor N. d To standardize its scale, the normalized charge density value is multiplied again by N. d ,Right now:

[0087] ρ i "=ρ i ′×N d ;

[0088] The values ​​ρ″ = {0.000, 0.774, 0.582, 1.355, 0.194} are obtained, forming the final normalized parameter set. This normalization operation allows the particle size data and charge density peak values ​​to undergo subsequent perturbation mapping within a unified scale. The calculation process of the normalization parameters relies entirely on the statistical and linear scaling operations of the sampled data. Each transformation is performed within a defined interval, and the different data sources are sequentially mapped to the same scale. The normalization factor N... d The normalized charge density value needs to be recalculated after each mutation segment is identified to avoid the influence of the previous state on the current state. No fuzzy expression is set in the normalization operation. The data of the participating terms in the normalization process are shown in Table 2.

[0089] Table 2: Particle Size Distribution and Charge Density Data

[0090]

[0091] As shown in Table 2, the weighting of particle size distribution is first performed during the calculation to obtain the uniformity factor N. d Then, by mini-maximum normalization of charge density, scale unification is achieved, and finally, a perturbation participation parameter set {ρj″} with a unified scale is obtained. All participation parameters can be directly obtained or sampled in industrial dust field monitoring. The mapped normalized parameter set can be used for subsequent perturbation exponent sequence derivation and dynamic response evaluation.

[0092] S202: Based on the component values ​​in the normalized parameter set, set the disturbance intensity index reference sequence, use linear interpolation to map the normalized vector components sequentially to the corresponding disturbance intensity interval, calculate the disturbance response factor set for each corresponding component based on the interpolation results, and generate the disturbance mapping value set.

[0093] A reference sequence for the disturbance intensity index is established by first reading the component values ​​in the normalized parameter set, assigning the set as {p1 = 0.00, p2 = 0.774, p3 = 0.582, p4 = 1.355, p5 = 0.194}. Then, the values ​​are mapped to the corresponding disturbance intensity intervals using linear interpolation. The disturbance intensity intervals are initially divided as [0, 0.25] → L, (0.25, 0.5] → M, (0.5, 0.75] → H, and (0.75, 1.0] → VH.

[0094] (>1.0)→XH, process each pi one by one. For example, for p1=0.00, determine that it falls within [0, 0.25]. Select the lower bound a1=0 and the upper bound b1=0.25 of the interval, and perform linear interpolation to calculate its mapping intensity exponent:

[0095]

[0096] Normalizing this value to between 0 and 1 still yields 0.00. For p2 = 0.774, it falls within (0.75, 1.0], corresponding to the interval (0.75, 1.0]. Let a2 = 0.75 and b2 = 1.0. The linear interpolation value is:

[0097]

[0098] Similarly, for p3 = 0.582, falling within (0.5, 0.75], the linear interpolation value is:

[0099]

[0100] For p4 = 1.355, it falls within the interval > 1.0. The upper bound of this interval is set as b4 = 1.5 as the preset maximum index. The linear interpolation value is:

[0101]

[0102] For p5 = 0.194, falling within [0, 0.25], a5 = 0, b5 = 0.25, the interpolation value is:

[0103]

[0104] This generates a reference sequence for the perturbation intensity index:

[0105] {I j} = {0.00, 0.096, 0.328, 0.71, 0.776};

[0106] Next, the set of disturbance response factors for each component is calculated based on the interpolation results. The execution process is as follows: for each I... j Create response factor vector {α} j，1 α j，2 α j，3 The vector dimension is set to 3, corresponding to the before, middle, and after time periods. The initial weights are set to {0.5, 0.3, 0.2}, and the linear interpolation results are distributed according to these weights in sequence.

[0107] α 1，1 =0.00 × 0.5 = 0.000, α 1，2 =0.00 × 0.3 = 0.000, α 1，3 =0.00 × 0.2 = 0.000;

[0108] α 2，1 =0.096 × 0.5 = 0.048, α 2，2 =0.096 × 0.3 = 0.0288, α 2，3 =0.096 × 0.2 = 0.0192;

[0109] And so on for all I j The calculated set of disturbance response factors is as follows:

[0110] {{α j，k}}={{0.000, 0.000, 0.000}, {0.048, 0.0288, 0.0192}, {0.164, 0.0984, 0.0656}, {0.355, 0.213, 0.142}, {0.388, 0.2328, 0.1552}};

[0111] Finally, a perturbation mapping set is generated, which is the response factors that are flattened and combined into a sequence:

[0112] The sequence {0.000, 0.000, 0.000, 0.048, 0.0288, 0.0192, 0.164, 0.0984, 0.0656, 0.355, 0.213, 0.142, 0.388, 0.2328, 0.1552} is used as the output.

[0113] S203: Call the disturbance mapping value group, extract the components based on the current current pulse frequency and the data offset of the previous two time windows, perform item-by-item exponential weighting, set the frequency offset exponential coefficient vector as the weighting factor, and perform item-by-item combination operation on the disturbance value and offset data to obtain the dynamic response parameter group;

[0114] The disturbance mapping value group is invoked. Based on the current current pulse frequency and the data offset of the previous two time windows, the components are extracted and subjected to term-by-term exponential weighting. The frequency offset exponential coefficient vector is set as a weighting factor. The disturbance value and offset data are combined and calculated group by group to obtain the dynamic response parameter group. During execution, the current frequency f is read first. c =50kHz, and the set of offsets for the first two time windows {δ -2 =0.005, δ -1 =0.008}, construct the frequency offset exponential coefficient vector {β} from it. -2 ,β -1 ,β0}, where β0=e 0 =1, and the examples are as follows:

[0115] β -2 =e -0.005 / 50000 =e -0.0000001 ≈0.9999999;

[0116] β -1 =e -0.008 / 50000 =e -0.00000016 ≈0.99999984;

[0117] β0 = 1;

[0118] Then, extract the first three components {α1 = 0.000, α2 = 0.000, α3 = 0.000} from the perturbation mapping value set, and perform a term-by-term exponential weighted operation:

[0119] R1=α1×β -2 =0.000 × 0.9999999 = 0.000;

[0120] R² = α² × β -1 =0.000 × 0.99999984 = 0.000;

[0121] R3=α3×β0=0.000×1=0.000;

[0122] This operation is performed sequentially on every three components of the perturbation mapping sequence. The frequency offset values ​​corresponding to the fourth to sixth components are 2, 1, and the current value, respectively. After inserting the frequency offset, the process is repeated to calculate R for each group of three values. j For example, for {α4=0.048, α5=0.0288, α6=0.0192}, we get:

[0123] R4=0.048×0.9999999=0.0479999952, R5=0.0288×0.99999984=0.028799995, R6=0.0192×1=0.0192;

[0124] Based on this calculation, a dynamic response parameter set {R} is generated. k For example, if the first six terms are {0.000, 0.000, 0.000, 0.0479999952, 0.028799995, 0.0192, ...}, and the fourteenth and fifteenth groups are {0.355, 0.213, 0.142} and {0.388, 0.2328, 0.1552} respectively, the corresponding response values ​​will continue to be calculated according to the frequency offset exponential coefficient. Finally, a dynamic response parameter group with the same length as the disturbance mapping value group will be output as the input for further calculation of the feedback system.

[0125] Specifically, such as Figure 4 As shown, the specific steps of S3 are as follows:

[0126] S301: Based on the response delay coefficient in the dynamic response parameter set, extract the velocity and density data of the electrode edge region, input them into the Sobel operator to perform horizontal and vertical gradient calculations, and uniformly map them to a two-dimensional coordinate system to generate velocity density gradient distribution values.

[0127] Based on the response delay coefficient in the dynamic response parameter set, an initial set of velocity and density data is first collected from the edge region of the electrode plate. This region consists of pixels at the edges of the grayscale distribution values ​​in the two-dimensional image. Specifically, the edge line of the electrode plate in a frame of the image is selected, and a 2-pixel width is extended inward from this edge as the edge region. The displacement change rate (velocity) and the number of pixels per unit area (density) are extracted pixel by pixel. For example, in the 5th frame, the displacement change of pixels in a certain edge region is 3 pixels / frame, which is converted to a velocity of 0.6 mm / s. The number of pixels in the region is 25, and the converted area is 0.25 mm². 2 The density is 100 particles / mm 2 Subsequently, gradient operators are constructed in both row and column directions, and Sobel filter kernels are used to perform horizontal and vertical gradient calculations, i.e., kernel matrices are used respectively:

[0128] G x = [-1, 0, +1; -2, 0, +2; -1, 0, +1[;

[0129] G y = [-1, -2, -1; 0, 0, 0; +1, +2, +1];

[0130] The algorithm slides across the velocity and density images respectively, performing matrix multiplication and summation to extract the gradient value at each pixel. This extracts the trends in velocity and density at edge regions. For example, during the extraction process, the velocity gradient at pixel (30, 45) is calculated as G. x =1.2, G y =0.9, density gradient is G x =2.5, G v =3.1. Then, the extracted velocity and density gradient values ​​are uniformly mapped to a two-dimensional coordinate system according to their coordinate positions to construct a gradient distribution image matrix. The coordinate axes represent the rows and columns of pixels respectively. The elements in the matrix record the combined value of the velocity gradient and density gradient at the corresponding pixel position. For example, in the construction result, the pixel (30, 45) corresponds to the recorded value of {velocity gradient: 1.2, density gradient: 2.5}. In this way, all edge region pixels are traversed to complete the generation of velocity and density gradient distribution values.

[0131] S302: Call the velocity density gradient distribution value, extract the velocity and density gradient values ​​according to the coordinates, perform weighted fusion according to the weight rules set by the response delay coefficient, traverse the edge pixels to perform superposition operation, and generate velocity density fusion fluctuation value;

[0132] The velocity density gradient distribution value is retrieved, and the velocity gradient value and density gradient value at each position are read sequentially according to pixel coordinates. These are then mapped to the previously obtained response delay coefficient. Based on the position-time mapping relationship, the response delay coefficient that should match each pixel is determined. For example, if a pixel is located in the edge region between the 3rd frame and the current frame, with a corresponding delay of 2 frames, the corresponding response delay coefficient is 0.712. This coefficient is derived from the frequency offset exponential calculation, specifically based on a frequency of 50kHz and a current pulse offset of 0.007, using exponential decay. The formula is as follows: The velocity gradient and density gradient values ​​are then weighted and fused according to set weights. The weights are dynamically adjusted based on the response delay coefficient. The fusion method is to multiply the velocity gradient by the first term of the weights and the density gradient by the second term of the weights. For example, if the velocity gradient weight is set to 0.6 and the density gradient weight is set to 0.4, then the weighted fusion result of velocity gradient 1.5 and density gradient 2.1 is 1.5×0.6+2.1×0.4=0.9+0.84=1.74. Pixels are processed sequentially to complete the calculation of pixel fusion values. The fusion result is recorded as a single value in the fusion fluctuation matrix. Pixels in the edge region are superimposed, that is, the fusion values ​​are summed and averaged. For example, if the sum is 698 and the number of participating pixels is 400, then the fusion fluctuation value is 698 / 400=1.745. Finally, the velocity-density fusion fluctuation value array is obtained and used for subsequent disturbance amplitude judgment.

[0133] S303: Call the velocity density fusion fluctuation value, identify the amplitude point, extract the direction change trend and gradient slope, calculate the weighted total of amplitude, offset angle and mutation rate, and generate the disturbance intensity coefficient;

[0134] The velocity density fusion fluctuation value is called first by iterating through the values ​​in the fusion array and comparing their amplitude to determine the fluctuation amplitude points. The amplitude point judgment threshold is set to the mean plus the standard deviation of the fusion value array. If the mean of the fusion value is 1.63 and the standard deviation is 0.35, then the amplitude judgment threshold is 1.98. Any point exceeding this value is marked as an amplitude point. For example, in the sample, the fusion value of pixel (25, 34) is 2.13, so it is marked as a valid point. Then, the direction and trend of the fusion value change of the marked points in consecutive frames are calculated. The direction change is judged by the change in the size of the pixel value at the same position in adjacent frames. If the value in the later frame is greater than that in the current frame, the direction is judged as rising, and vice versa. For example, the value of (25, 34) in frame 3 is 2.13, and in frame 4 it is 2.52, so the direction is rising and the trend is continuously increasing. Then, for each The gradient slope of the amplitude point is calculated within its neighborhood. The values ​​of adjacent 3×3 regions are taken, and the corresponding slope is calculated using the first-order difference method. For example, the slope value is calculated by dividing the maximum difference between adjacent points by the distance. That is, if the point value difference is 0.6 and the pixel spacing is 0.5mm, the gradient slope is 1.2. The slope, direction trend, and amplitude of the amplitude point are summarized and multiplied by the set weight factors. The amplitude weight is set to 0.5, the direction change weight is 0.3, and the gradient slope weight is 0.2. For example, the total weight of the three is 2.52×0.5=1.26, the direction increase is 1×0.3=0.3, and the slope is 1.2×0.2=0.24, with a total of 1.8. The total weight of each point is summed and averaged to generate the final disturbance intensity coefficient. For example, if the total weight of 20 valid points is 36, the disturbance intensity coefficient is 36 / 20=1.8.

[0135] Table 3: Examples of Velocity and Density Sampling

[0136]

[0137] Table 3 lists the velocity and density parameters of some pixels in the edge region, the calculated gradient value and the fusion value, and determines whether it is an amplitude point and its corresponding perturbation weighted total amount.

[0138] Specifically, such as Figure 5 As shown, the specific steps of S4 are as follows:

[0139] S401: Call the perturbation intensity coefficient. When the perturbation intensity coefficient exceeds the weight transfer threshold, extract all weight parameters in the model and assign each weight parameter to the particle size offset angle and the local electric field distribution vector to generate the weight parameter transfer amount.

[0140] The perturbation strength coefficient can be invoked by sampling and analyzing the initial electric field response of particle size samples under a specific electrostatic environment. A standard deviation threshold is set; when the local electric field standard deviation exceeds this threshold, the perturbation strength is considered to exceed the limit. For example, if the standard deviation threshold is set to 2.5 V / m, the electric field response values ​​of three groups of samples with particle sizes of 15 μm, 20 μm, and 25 μm are collected at different offset angles, and the standard deviation is calculated. If the standard deviation at an offset angle of 30° for a 20 μm particle size reaches 2.9 V / m, the judgment condition is met. All weight parameters in the model are extracted, and the connection weights in the model network are iterated through item by item, numbered sequentially in array form as w1, w2, ..., w m If the model is a three-layer neural network with 25 hidden layer weights and 10 output layer weights, then the total weights are 35. After extracting the weights, they are mapped according to the particle size offset angle and the local electric field distribution vector. That is, through a structured mapping table, the particle size offset angle group (e.g., 30°, 45°, 60°) is paired with the three-dimensional vector response group of the local electric field (e.g., [18.3, 21.1, 22.7] V / m). Each weight value is assigned one-to-one with its corresponding angle and electric field pair according to its index, generating an initial weight transfer array. Each weight transfer can be represented as the correlation strength value between the original weight value and the angle and electric field combination. Through this process, a one-dimensional array δ is constructed. i，j Where i represents the weight index and j represents the offset angle-electric field combination number. For example, i=1 and j=3 correspond to w1 corresponding to the offset angle 45° and the local electric field [21.1, 20.5, 22.0] V / m transfer relationship value is 0.84, forming a matrix with the following structure for subsequent use.

[0141] Table 4: Relationship between weight, particle size offset angle, and local electric field combination

[0142] Weight number Offset angle (°) Electric field distribution vector (V / m) Transfer value 1 30 [18.3，19.5，20.1] 0.72 2 45 [21.1，20.5，22.0] 0.84 3 60 [23.8，24.2，25.1] 0.91

[0143] As shown in Table 4, the transfer amount of the initial weight parameters can be calculated by measuring the difference between the weight and the local electric field distribution trend, thus completing the transformation of the weight value and the particle size-electric field combination mapping relationship.

[0144] S402: Call the weight parameter transfer amount, combine the weight with the particle size offset angle, pair the values ​​of the particle size offset angle and the local electric field distribution vector, calculate the function expression of the two using the least squares method, extract the regression parameters in the expression, calculate the residual absolute value index, analyze the goodness of fit between the residual vector and the parameters, and generate the particle size and electric field fitting coefficients.

[0145] Call the weight transfer array δ generated above i，j According to the pairing logic, the weight w is... i With the standardized particle size offset angle θ j Multiply by ' to form the intermediate variable term w'i ·θ j For example, if w1 = 0.72 and θ1′ = 0.45, then the product is 0.324, and at the same time, each dimension E in the local electric field distribution vector is extracted. j，k Calculate its relationship with the mean μ of the sample space. j，k The absolute value of the difference divided by the standard deviation σ j，k This refers to the deviation measure after Z-score standard normalization, calculating the absolute value of the residual index using the formula:

[0146]

[0147] Where, ΔR i，j w represents the absolute value index of the residual of the combination of the i-th weight parameter and the j-th particle size offset angle-electric field vector. i θ represents the i-th weight parameter (dimensionless) in the model. j ′ represents the j-th particle size offset angle, which is processed into a dimensionless parameter by the Z-score standard, E j，k μ represents the numerical value (in volts per meter, V / m) of the local electric field distribution vector in the k-th dimension at the j-th particle size offset angle. j，k σ represents the average value (in volts per meter, V / m) of the k-th dimension electric field at the j-th particle size offset angle within the sample space. j，k The standard deviation of the electric field in the k-th dimension represents the electric field at the j-th particle size offset angle (in volts per meter, V / m). This represents the joint variance of the i-th weighting parameter in the k-th dimension of the electric field direction (in volts per meter squared (V / m)). 2 ), The average variance of the electric field direction in the k-th dimension represents the particle size offset angle (in volts per square meter (V / m)). 2 ), γ i，j ρ represents the ratio (dimensionless) of the local electric field fitting deviation between the i-th weight parameter and the j-th particle size offset angle. i，j represents the predicted value of the i-th weight parameter under the j-th combination in the least squares fitting (dimensionless), n represents the total dimension of the local electric field distribution vector (unitless), and k is the summation index variable (unitless).

[0148] For example, for the three-dimensional electric field distribution vector [21.1, 20.5, 22.0], let μ... j = [20.0, 20.0, 21.0], σ j = [0.8, 1.2, 0.9], then the dimension normalization result is [1.375, 0.417, 1.111].

[0149] Adding each item together yields 2.903, which is the normalized total bias; then, the joint variance is called... For example, let the joint variance of w1 in the three-dimensional directions be [1.1, 1.3, 1.0] (V / m). 2 The average variance under the combination is:

[0150] The component difference term is calculated as follows:

[0151]

[0152] Assuming deviation ratio γ i，j = [0.1, 0.15, 0.05], then the absolute value of the difference is [0.817, 1.027, 1.052], and the mean is approximately 0.965.

[0153] The denominator is 1 + 0.965 = 1.965;

[0154] Call the aforementioned item w i ·θ j The sum of the deviations, 2.903, gives the numerator 3.227. The final calculated absolute value index of the residuals is:

[0155] ΔR i，j =|3.227 / 1.965-ρ i，j |, set ρ i，j =1.55.

[0156] The final residual is |1.642-1.55|=0.092. After performing this operation on the combination, the regression parameters corresponding to the smallest residual term are extracted and denoted as β1, β2, and β3, respectively. The goodness-of-fit index R2 is calculated using the mean and standard deviation of the residual vector, and finally a table of fitting coefficients between particle size and local electric field response is generated for the result output.

[0157] The advantage of this formula is that by introducing the standardized processing of the particle size offset angle and the normalized deviation term of the local electric field vector component to construct a composite residual term, the weights of the multi-source structured parameters in the same regression model are effectively integrated, thereby more comprehensively expressing the nonlinear response relationship and controlling the prediction error through residual reduction, making the mapping between parameters more reliable.

[0158] S403: Based on the fitting coefficients of particle size and electric field, organize the data corresponding to the particle size offset angle and the local electric field distribution vector, summarize the weight distribution on the distribution, and obtain the reconstructed weight distribution table.

[0159] Based on the aforementioned particle size-electric field fitting coefficient regression parameter set β i The particle size offset angles and their corresponding local electric field distribution vectors are processed group by group. For example, if the particle size offset angles are set to 30°, 45°, and 60°, the corresponding electric field distribution vectors are:

[0160] [19.3, 20.2, 20.5], [21.5, 22.0, 22.3], [23.1, 24.2, 24.7];

[0161] Through normalization and residual screening, the optimal regression coefficient set for each group is [0.62, 0.75, 0.83]. Then, using the coefficients as the core weights, the aforementioned weight parameters w1, w2, ..., w... m The weights are allocated to form a reconstruction weight table for each particle size offset angle and electric field distribution. Each row in the table corresponds to a set of angle-electric field combinations and the percentage of regression weights they carry. For example, the weight corresponding to the angle 45° is 0.75, which means that 75% of the i-th weight in the original model can be attributed to the particle size offset and local electric field response interval. The allocation results are summarized to form the final reconstruction weight allocation table, which is used as input for subsequent modeling iterations.

[0162] Specifically, such as Figure 6 As shown, the specific steps of S5 are as follows:

[0163] S501: Based on the reconstructed weight allocation table, the airflow velocity, pressure difference, offset angle and electric field distribution vector are decomposed into components, the original normalized values ​​are extracted respectively, the reverse normalization operation is performed, the data range of each parameter is gradually restored, and the parameter restoration distribution value is generated.

[0164] Based on the reconstructed weight allocation table, the normalized data of the four dimensions—airflow velocity, pressure difference, offset angle, and electric field distribution vector—are first read item by item, and their corresponding values ​​in the reconstruction table are extracted. Then, inverse normalization is performed on each parameter to restore its real-time physical value range. During the restoration process, the original data ranges for the parameters are first set: airflow velocity range is [15, 60] m / s, pressure difference range is [1000, 3000] Pa, offset angle range is [-30°, 30°], and each component of the electric field distribution vector is set to [0.1, 2.0] kV / cm. The inverse normalization operation uses the linear transformation formula x = x n ·(x max -x min )+x min , where x n To obtain the normalized value, the normalized value array is read item by item from the weight allocation table, and then transformed item by item in combination with the above interval. For example, when the normalized value of the airflow velocity is x n =0.6, then its restored value is:

[0165] x=0.6·(60-15)+15=42m / s;

[0166] Similarly, for the normalized value of the pressure difference x n =0.4, the restored value is:

[0167] x=0.4·(3000-1000)+1000=1800Pa;

[0168] The normalized value of the offset angle is x n =0.75, the restored value is:

[0169] x=0.75·(30-(-30))-30=15°;

[0170] The electric field vector is set to a three-dimensional distribution. Normalized values ​​[0.2, 0.5, 0.9] are read, and the restored values ​​are:

[0171] [0.2·(2.0-0.1)+0.1, 0.5·(2.0-0.1)+0.1, 0.9·(2.0-0.1)+0.1]=[0.48, 1.15, 1.81]kV / cm;

[0172] The restored values ​​are summarized below:

[0173] Table 5: Parameter Back-Normalized Values

[0174]

[0175]

[0176] As shown in Table 5, the original physical quantity can be accurately deduced from the normalized value for each parameter, and the final parameter distribution value can be obtained.

[0177] S502: Call the parameter to restore the distribution value, use the gradient descent method to progressively correct the parameter weights and the initial configuration weights in round by round, monitor the numerical differences between weights in real time, accumulate the weight change range of each parameter, and obtain the weight deviation trend value.

[0178] The restored parameter value sequence is called, and an initial parameter vector group is constructed using three types of data: airflow velocity, pressure difference, offset angle, and electric field distribution. This initial parameter vector group is compared with the initial configured weight value group, and the numerical differences are calculated item by item. In this step, the initial weights for airflow velocity are set to w1 = 0.25, pressure difference to W2 = 0.30, offset angle to w3 = 0.20, and the weight of the three components of the electric field vector is set to w4 = 0.25 after merging. In each iteration, the current parameter value is multiplied by its corresponding initial weight to generate the current weighted value. For example, if the airflow velocity is 42... The current weighted value is 42 × 0.25 = 10.5. After calculating the parameter weights accordingly, the difference between this value and the target output value (assuming it is the model fitting value, specifically: airflow output fitting value is 11, pressure difference fitting value is 9.5, offset angle is 3.8, and electric field is 6.0) is calculated. For example, the airflow velocity deviation is 11 - 10.5 = 0.5. This deviation is included in the cumulative deviation term for this round. The deviation values ​​for each round are uniformly accumulated and added to the original deviation sequence group. The current weight value is then adjusted according to the gradient update formula. That is, for each difference δ... i The adjustment amount for the weights is set as Δw. i =η·x i ·δ i The learning rate η is set to 0.01. Taking airflow speed as an example:

[0179] Δw1=0.01·42·0.5=0.21;

[0180] The new weights are w1′=0.25+0.21=0.46. Similarly, the remaining parameters are calculated as follows:

[0181] Pressure difference = 1800 × 0.3 = 540, Difference = 10 - 9.5 = 0.5;

[0182] Δw2=0.01·1800·0.5=9.0→w2′=0.30+9.0=9.30;

[0183] Since this value deviates significantly from the normal range, a maximum weight threshold needs to be set to 1.0. If the updated value exceeds 1.0, it needs to be corrected back to 1.0. The current change magnitude is recorded and entered into the original deviation curve. The weight change record formed after multiple adjustments is represented by the following array:

[0184] ΔW=[0.21, 0.70, 0.13, 0.26],

[0185] This array serves as the weight change trend vector after each round of gradient adjustment, and is cumulatively added to the input of the next round of iteration to achieve the gradual extraction of the weight deviation trend value.

[0186] S503: Based on the weight deviation trend value, the gradient descent method is used to continuously determine the relationship between the cumulative deviation and the set tolerance. When the deviation of the parameter is lower than the set tolerance, the current parameter values ​​are summarized to establish a steady-state modeling parameter set.

[0187] Based on the accumulated weighted deviation trend value set [0.21, 0.70, 0.13, 0.26], the deviation changes of multiple weights are judged item by item, and the tolerance range is set to [0.05, 0.20]. If the deviation of any parameter is lower than the lower limit of the tolerance, it is judged that it has converged. In the example, the deviation value corresponding to the offset angle is 0.13, which is lower than 0.20, and can be regarded as converged. The deviation of the airflow velocity is 0.21, which is slightly higher than the upper limit of the tolerance, and iterative correction is required. The deviation of the electric field distribution is also considered. The difference is 0.26, which does not meet the convergence condition, so it is retained as an unstable parameter. For the pressure difference with a deviation of 0.70, an additional adjustment step size is required, and a jump correction strategy is implemented. Specifically, the current learning rate η is increased to 1.5 times its original value for subsequent iterations. The stable convergence term records the current parameter value in the steady-state parameter set, such as recording the offset angle as 15°, and marking the current convergence timestamp and round index, such as convergence at the 8th round and the 200th step. The final steady-state modeling parameter set is as follows:

[0188] Steady-state parameter set = {offset angle: 15°, number of recording rounds: 8, number of recording steps: 200};

[0189] The remaining non-converged terms will be retained in the next round of parameter training, and the steady-state group will be submitted as the final model configuration parameter group for archiving and subsequent module calls.

[0190] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A dynamic modeling method for electrostatic precipitator flue gas emissions based on multi-parameter coupling analysis, characterized in that, Includes the following steps: S1: The electric field intensity signal is obtained by sampling at equal intervals. The time window data is processed by a sliding window mechanism. The electric field change rate is calculated by first-order difference and the electric field acceleration is calculated by second-order difference. When the acceleration of three consecutive windows exceeds the acceleration judgment threshold and the rate direction is consistent, the electric field change judgment result is generated. S2: Call the electric field sudden change judgment result to trigger the parameter adjustment command, perform normalization processing on the dust particle size distribution data and charge density peak value, use linear interpolation to map the disturbance intensity, perform exponential weighting calculation according to the current pulse frequency offset, and output dynamic response parameter group; S3: Based on the response delay coefficient in the dynamic response parameter group, the Sobel operator is used to calculate the velocity gradient field and density gradient field of the electrode edge region, and a weighted superposition operation is performed to calculate the linear combination of velocity fluctuation amplitude, directional offset angle and density abrupt change rate to generate the disturbance intensity coefficient. S4: When the disturbance intensity coefficient exceeds the weight transfer threshold, the model weights are transferred to the particle size offset angle and the local electric field distribution vector. The offset angle and the electric field distribution vector are fitted using the least squares method to generate a reconstructed weight allocation table.

2. The method for dynamic modeling of electrostatic precipitator flue gas emissions based on multi-parameter coupling analysis according to claim 1, characterized in that, The electric field abrupt change determination results include acceleration threshold, velocity direction unidirectionality, and window trigger count. The dynamic response parameter set includes disturbance mapping baseline value, frequency offset, and charge density peak value. The disturbance intensity coefficient includes velocity fluctuation amplitude, flow direction deviation angle, and density gradient change rate. The reconstructed weight allocation table includes particle size distribution weight, field strength vector weight, and least squares regression coefficient.

3. The method for dynamic modeling of electrostatic precipitator flue gas emissions based on multi-parameter coupling analysis according to claim 1, characterized in that, The specific steps of S1 include: S101: The electric field intensity signal is obtained by sampling at equal intervals. The starting position and step size of the sliding window are set, the window is advanced to divide the intensity sequence, the intensity offset between adjacent times in each sequence is calculated, and the electric field change rate sequence is obtained by combining them. S102: Call the electric field change rate sequence, perform a second difference operation on the rate values ​​within the time window, construct a second-order difference sequence within the window and calculate the average value to obtain the electric field acceleration change sequence; S103: Call the electric field acceleration change sequence, compare the average electric field acceleration in three consecutive sliding windows with the acceleration judgment threshold, filter the time window group where the acceleration exceeds the acceleration judgment threshold and the rate direction is consistent in the electric field change rate sequence, and generate the electric field sudden change judgment result.

4. The method for dynamic modeling of electrostatic precipitator flue gas emissions based on multi-parameter coupling analysis according to claim 3, characterized in that, The specific steps of S2 include: S201: Call the electric field abrupt change judgment result, obtain the dust particle size distribution data and charge density peak data in the corresponding time period, calculate the normalization factor based on the particle size distribution data, perform interval mapping processing on the charge density peak parameter, unify the scale range of the two sets of data, and obtain the normalization parameter set. S202: Based on the component values ​​in the normalized parameter group, set a disturbance intensity index reference sequence, use linear interpolation to sequentially map the normalized vector components to the corresponding disturbance intensity intervals, calculate the disturbance response factor set for each corresponding component based on the interpolation results, and generate a disturbance mapping value group. S203: Call the disturbance mapping value group, extract the components based on the current current pulse frequency and the data offset of the previous two time windows, perform item-by-item exponential weighting, set the frequency offset exponential coefficient vector as the weighting factor, and perform group-by-group combination operation on the disturbance value and offset data to obtain the dynamic response parameter group.

5. The method for dynamic modeling of electrostatic precipitator flue gas emissions based on multi-parameter coupling analysis according to claim 4, characterized in that, The specific steps of S3 include: S301: Based on the response delay coefficient in the dynamic response parameter group, extract the velocity and density data of the electrode edge region, input them into the Sobel operator to perform horizontal and vertical gradient calculations, and uniformly map them to a two-dimensional coordinate system to generate velocity density gradient distribution values. S302: Call the velocity density gradient distribution value, extract the velocity and density gradient values ​​according to the coordinates, perform weighted fusion according to the weight rules set by the response delay coefficient, traverse the edge pixels to perform superposition operation, and generate velocity density fusion fluctuation value. S303: Call the velocity density fusion fluctuation value, identify the amplitude point, extract the direction change trend and gradient slope, calculate the weighted total of amplitude, offset angle and mutation rate, and generate the disturbance intensity coefficient.

6. The method for dynamic modeling of electrostatic precipitator flue gas emissions based on multi-parameter coupling analysis according to claim 5, characterized in that, The velocity density gradient distribution value represents the spatial gradient change of velocity and density in the edge region of the electrode, revealing local dynamic characteristics; The weighted fluctuation intensity obtained by fusing the velocity and density gradients reflects the overall degree of local disturbance. The disturbance intensity coefficient is based on the extraction of direction, amplitude, and abrupt change information from the fused fluctuation quantity to quantify the overall intensity of the edge disturbance; The Sobel operator is used to extract the horizontal and vertical gradients of the velocity and density matrices. It combines the response delay coefficient to set edge processing and normalization parameters, and after unified mapping, generates gradient distribution values ​​that can be used for fusion calculation.

7. The method for dynamic modeling of electrostatic precipitator flue gas emissions based on multi-parameter coupling analysis according to claim 5, characterized in that, The specific steps of S4 include: S401: Call the disturbance intensity coefficient. When the disturbance intensity coefficient exceeds the weight transfer threshold, extract all weight parameters in the model and assign each weight parameter to the particle size offset angle and the local electric field distribution vector to generate the weight parameter transfer amount. S402: Call the weight parameter transfer amount, combine the weight with the particle size offset angle, pair the values ​​of the particle size offset angle and the local electric field distribution vector, calculate the function expression of the two using the least squares method, extract the regression parameters in the expression, calculate the residual absolute value index, analyze the goodness of fit between the residual vector and the parameters, and generate the particle size and electric field fitting coefficient. S403: Based on the particle size and electric field fitting coefficients, organize the data corresponding to the particle size offset angle and the local electric field distribution vector, summarize the weight distribution on the distribution, and obtain the reconstructed weight allocation table.

8. The method for dynamic modeling of electrostatic precipitator flue gas emissions based on multi-parameter coupling analysis according to claim 1, characterized in that, The method includes step S5: S5: Call the reconstructed weight allocation table to perform weight regression, perform reverse normalization on the airflow velocity, pressure difference, offset angle and electric field distribution vector, use the gradient descent method to calculate the cumulative deviation between the parameter weights and the initial configuration, and output the steady-state modeling parameter set when the deviation is less than the set tolerance. The steady-state modeling parameter set includes a weight matrix, residual cumulative amount, and inverse normalization coefficient.

9. The method for dynamic modeling of electrostatic precipitator flue gas emissions based on multi-parameter coupling analysis according to claim 8, characterized in that, The specific steps of S5 include: S501: Based on the reconstructed weight allocation table, the airflow velocity, pressure difference, offset angle and electric field distribution vector are decomposed into sub-items, the original normalized values ​​are extracted respectively, the reverse normalization operation is performed, the data range of each parameter is gradually restored, and the parameter restoration distribution value is generated. S502: Call the parameter to restore the distribution value, use the gradient descent method to progressively correct the parameter weights and the initial configuration weights, monitor the numerical differences between weights in real time, accumulate the weight change range of each parameter, and obtain the weight deviation trend value. S503: Based on the weight deviation trend value, the gradient descent method is used to continuously determine the relationship between the cumulative deviation and the set tolerance. When the deviation of the parameter is lower than the set tolerance, the current parameter values ​​are summarized to establish a steady-state modeling parameter group.

10. The method for dynamic modeling of electrostatic precipitator flue gas emissions based on multi-parameter coupling analysis according to claim 9, characterized in that, The parameter restoration distribution value is obtained by inverse normalization of the normalized parameters to restore their values ​​within the original physical range. The weight deviation trend value is a numerical indicator that measures the trend of parameter weight changes corresponding to the initial weight during iterative optimization. The gradient descent method iteratively adjusts the parameter weights, gradually reducing the weight deviation and approaching the set tolerance, thereby optimizing the steady-state modeling parameters.

Images