Intelligent control method for desulfurization and denitrification integrated device based on multi-sensor data fusion

Abstract

The invention discloses an intelligent control method for a desulfurization and denitrification integrated device based on multi-sensor data fusion, and relates to the technical field of industrial flue gas treatment, and the method comprises the following steps: (1) collecting flue gas parameter data in real time through a gas concentration sensor, a temperature sensor, a pressure sensor and a flow velocity sensor of the desulfurization and denitrification integrated device; according to the intelligent control method for the desulfurization and denitrification integrated device based on multi-sensor data fusion, the control precision and the response speed of the desulfurization and denitrification device under complex working conditions are improved through the collaborative design of dynamic credibility evaluation and the differentiable weight decision tree. Sensor noise is separated in real time based on a multi-scale space-time convolutional network, an edge-cloud dual verification mechanism is combined, the system can complete abnormal data shielding in a short time when smoke components suddenly change, compared with a traditional method, the pollutant concentration prediction error is reduced, the generation delay of an ammonia spraying amount regulation and control instruction is shortened, and the method is suitable for large-scale popularization and application. And the system robustness when the sensor fails or the load suddenly changes is ensured.

Description

Technical Field

[0001] The present invention relates to the technical field of industrial flue gas treatment, and specifically to an intelligent control method for a desulfurization and denitrification integrated device with multi-sensor data fusion. Background Art

[0002] In the field of intelligent control of desulfurization and denitrification integrated devices, existing technologies generally rely on the fusion of multi-source sensor data to optimize process parameters. However, the reliability of data fusion under dynamic working conditions has become the core bottleneck restricting the system efficiency. Traditional control methods usually fuse sensor data based on static weight allocation strategies. Its essential defect lies in the assumption that the credibility of sensor data is independent of the working conditions, and it cannot cope with dynamic interferences in complex scenarios such as sudden changes in flue gas composition, equipment aging, or load fluctuations. For example, when the sensor drifts due to environmental temperature changes or is affected by instantaneous noise, the fixed weight mechanism is difficult to identify data anomalies in real time, resulting in the fusion result deviating from the true value. At the same time, the coupling relationship between sensor data in the desulfurization and denitrification process has strong non-linear characteristics. Parameters such as ammonia injection amount, reaction tower temperature, and pollutant concentration show time-varying interaction effects. However, existing methods mostly use linear fusion models or shallow learning algorithms, which are difficult to accurately model such complex associations, resulting in cumulative deviations between the predicted values and actual values of key process parameters.

[0003] Although deep learning technology can capture non-linear characteristics, the problem of excessive model complexity makes it difficult to meet the millisecond-level real-time control requirements, and simplifying the model faces the risk of accuracy loss. This contradiction directly leads to decision-making delays or misjudgments in the intelligent control system under dynamic working conditions, which may cause excessive ammonia injection to increase operating costs, or emission exceedance to trigger environmental protection risks. Therefore, how to achieve real-time and reliable fusion of multi-source heterogeneous sensor data in a dynamic non-linear system and adaptively allocate weights to balance computational efficiency and fusion accuracy has become the key challenge for improving the adaptive ability of the intelligent control of desulfurization and denitrification integrated devices. Summary of the Invention

[0004] (I) Technical Problems to be Solved

[0005] In view of the deficiencies of the prior art, the present invention provides an intelligent control method for a desulfurization and denitrification integrated device with multi-sensor data fusion, which solves the problems of real-time and reliable fusion of multi-source heterogeneous sensor data and adaptive weight allocation in a dynamic non-linear environment.

[0006] (II) Technical Solutions

[0007] To achieve the above objectives, the present invention is realized through the following technical solutions: An intelligent control method for a desulfurization and denitrification integrated device with multi-sensor data fusion, comprising the following steps:

[0008] (1)Collect flue gas parameter data in real time through multiple gas concentration sensors, temperature sensors, pressure sensors and flow rate sensors of the desulfurization and denitrification integrated device;

[0009] (2)Input the flue gas parameter data into the implicit credibility evaluation module, decompose the data into a working condition characterization layer and a noise disturbance layer through a multi-scale spatio-temporal convolutional network, and dynamically calculate the real-time credibility scores of each sensor based on the energy ratio of the noise disturbance layer and the correlation coefficient of the historical data sliding window;

[0010] (3)Input the credibility scores and the current reaction tower temperature and ammonia injection amount parameters into the differentiable weight decision tree model, generate non-linear fusion weights through the path activation function, where the differentiable weight decision tree is optimized by lightweight adversarial training to simulate sensor failure and load mutation scenarios;

[0011] (4)Perform non-linear weighted fusion on the multi-sensor data according to the fusion weights, and output the pollutant concentration prediction value and the ammonia injection amount regulation instruction;

[0012] (5)Deploy a compressed LSTM model at the edge to verify the time series of the pollutant concentration prediction value. If the prediction deviation exceeds the dynamic threshold three times in a row, trigger the reconstruction of the weight decision tree;

[0013] (6)Based on the mass conservation equation, the cloud digital twin system reversely deduces the residual distribution of the fusion data. If the residual exceeds the limit, feedback a correction instruction to the edge to update the credibility score calculation logic;

[0014] (7)Adjust the decomposition channel number of the implicit credibility evaluation module and the node depth of the weight decision tree according to the verification results of steps (5) and (6) to achieve adaptive control under dynamic working conditions.

[0015] Preferably, the implementation method of the implicit credibility evaluation module in step (2) is:

[0016] The multi-scale spatio-temporal convolutional network includes three dilated convolutional layers, which extract the common features of sensor data with time windows of 1 second, 5 seconds, and 10 seconds respectively;

[0017] The noise disturbance layer separates the high-frequency noise components through the adaptive wavelet threshold algorithm, and its energy ratio calculation formula is:

[0018] , where W k is the wavelet decomposition coefficient, and S i is the original signal amplitude;

[0019] The correlation coefficient of the historical data sliding window is obtained by calculating the Pearson correlation between the current data segment and the same working condition data segment in the previous 30 minutes.

[0020] Preferably, the construction method of the differentiable weighted decision tree in step (3) includes:

[0021] (3.1) Define the tree node activation function as:

[0022] , where σ is the Sigmoid function, and ⊙ represents the Hadamard product;

[0023] (3.2) Inject random pulse noise with an amplitude of 15% - 30% of the sensor range during the training phase, and constrain the weight change rate not to exceed 10% through gradient penalty;

[0024] (3.3) Dynamically adjust the depth of the tree according to the real-time working condition complexity. When a sudden change in flue gas composition is detected, enable a maximum depth of 8 layers, and reduce it to 4 layers under steady-state working conditions.

[0025] Preferably, the structure optimization method of the compressed LSTM model in step (5) is:

[0026] Compress the hidden layer dimension of the standard LSTM to 256 dimensions, and implement the gating unit using grouped convolution with 4 groups;

[0027] The input sequence length is fixed at 10 seconds, and the model parameters are updated every 5 seconds. Only the training data of the most recent 1 hour is retained during the update.

[0028] Preferably, the reverse deduction method of the digital twin system in step (6) includes:

[0029] (6.1) Establish a virtual model including the structural dimensions of the reaction tower, catalyst distribution, and hydrodynamic parameters;

[0030] (6.2) Calculate the theoretical pollutant concentration value based on the Arrhenius equation and perform residual analysis with the fusion data;

[0031] (6.3) When the absolute value of the residual continuously exceeds 20% of the theoretical value for 5 times, it is determined that the fusion is abnormal and a correction coefficient matrix is generated.

[0032] Preferably, the implementation method of the gradient penalty constraint is:

[0033] Add a weight change rate regularization term to the loss function:

[0034] , where λ is the penalty factor, and its value ranges from 0.5 to 1.2.

[0035] Preferably, the specific rules of the adjustment operation in step (7) include:

[0036] When the edge device triggers reconstruction three times consecutively or the cloud feedbacks two correction instructions, increase the decomposition channel number of the multi-scale spatio-temporal convolutional network by 50%;

[0037] When there is no verification anomaly for 1 hour continuously, gradually reduce the number of nodes of the weighted decision tree to 70% of the initial state.

[0038] Preferably, it further includes a sensor anomaly handling mechanism: when the credibility score of a certain sensor is continuously lower than 0.4 for 5 times, start the redundant sensor switching program; the data of the newly enabled sensor needs to go through a 10-minute calibration period, during which its fusion weight increases linearly to the normal value.

[0039] Preferably, the implementation process of the grouped convolution includes: evenly dividing the input feature map into 4 groups according to the number of channels, independently performing convolution operations on each group and then splicing and outputting; the convolution kernel size is set to 3×3, the stride is set to 2, and zero-padding is used to keep the feature map size unchanged.

[0040] Preferably, the method for generating the correction coefficient matrix is: ; where α is the learning rate factor and takes values from 0.05 to 0.15, β is the residual weight coefficient and takes values from 0.8 to 1.2, R is the residual matrix, and I is the identity matrix.

[0041] (III) Beneficial effects

[0042] The present invention provides an intelligent control method for a desulfurization and denitrification integrated device with multi-sensor data fusion. It has the following beneficial effects:

[0043] (I). This intelligent control method for the desulfurization and denitrification integrated device with multi-sensor data fusion improves the control accuracy and response speed of the desulfurization and denitrification device under complex working conditions through the collaborative design of dynamic credibility evaluation and differentiable weighted decision tree. Based on the multi-scale spatio-temporal convolutional network to real-time separate sensor noise, combined with the edge-cloud dual verification mechanism, the system can complete abnormal data shielding in a short time when the flue gas composition mutates, the pollutant concentration prediction error is reduced compared with the traditional method, and the generation delay of the ammonia injection amount regulation instruction is shortened. At the same time, the lightweight adversarial training and dynamic computational flow pruning technology solve the problem of excessive model complexity, reduce the computational resource occupation under steady-state working conditions, achieve millisecond-level real-time control at the hardware end, and ensure the system robustness when the sensor fails or the load mutates.

[0044] (2) The intelligent control method for the integrated desulfurization and denitrification device with multi-sensor data fusion reduces the consumption of ammonia water and the occurrence rate of the risk of excessive emissions through precise dynamic regulation of the ammonia injection volume. Its adaptive parameter adjustment mechanism can be long-term adapted to equipment aging and catalyst activity decay. It has prominent technical scalability, and the multi-scale feature extraction and edge-cloud collaborative architecture can be migrated to industrial scenarios such as boiler combustion optimization, providing an efficient and reliable technical path for the intelligent upgrading of high-energy-consuming industries. Description of the Drawings

[0045] Figure 1 It is a schematic flow chart of the whole invention;

[0046] Figure 2 It is a timing diagram of the control logic of the invention. Detailed Embodiments

[0047] Next, the technical solutions in the embodiments of the present invention will be clearly and completely described in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative work shall fall within the protection scope of the present invention.

[0048] Please refer to Figure 1 and Figure 2 , the present invention provides a technical solution: an intelligent control method for an integrated desulfurization and denitrification device with multi-sensor data fusion, characterized by including the following steps:

[0049] (1) Real-time collect flue gas parameter data through multiple gas concentration sensors, temperature sensors, pressure sensors and flow rate sensors of the integrated desulfurization and denitrification device;

[0050] (2) Input the flue gas parameter data into the implicit credibility evaluation module, decompose the data into a working condition characterization layer and a noise disturbance layer through a multi-scale spatio-temporal convolutional network, and dynamically calculate the real-time credibility scores of each sensor based on the energy ratio of the noise disturbance layer and the correlation coefficient of the historical data sliding window;

[0051] (3) Input the credibility scores and the current reactor temperature and ammonia injection volume parameters into the differentiable weight decision tree model, generate non-linear fusion weights through the path activation function, wherein the differentiable weight decision tree is optimized by lightweight adversarial training to simulate sensor failure and load mutation scenarios;

[0052] (4) Perform non-linear weighted fusion on the multi-sensor data according to the fusion weights, and output the pollutant concentration prediction value and the ammonia injection volume regulation instruction;

[0053] (5) Deploy a compressed LSTM model at the edge to perform temporal verification on the predicted pollutant concentration values. If the prediction deviation exceeds the dynamic threshold for three consecutive times, trigger the reconstruction of the weight decision tree;

[0054] (6) Through the cloud digital twin system, reverse-deduce the residual distribution of the fusion data based on the mass conservation equation. If the residual exceeds the limit, feedback a correction instruction to the edge to update the credibility score calculation logic;

[0055] (7) Adjust the decomposition channel number of the implicit credibility evaluation module and the node depth of the weight decision tree according to the verification results of steps (5) and (6) to achieve adaptive control under dynamic working conditions.

[0056] It should be further noted that in the specific implementation process, the specific implementation method of the implicit credibility evaluation module in step (2) is as follows:

[0057] The multi-scale spatio-temporal convolutional network includes three dilated convolutional layers, which extract the common features of sensor data with time windows of 1 second, 5 seconds, and 10 seconds respectively. The dilation rate of the first layer is 2 and the convolutional kernel size is 3×3, the dilation rate of the second layer is 4 and the convolutional kernel size is 5×5, and the dilation rate of the third layer is 8 and the convolutional kernel size is 7×7;

[0058] The noise perturbation layer separates the high-frequency noise components through the adaptive wavelet threshold algorithm, and its energy ratio calculation formula is:

[0059] , where W k is the decomposition coefficient of the db4 wavelet basis, S i is the amplitude of the original signal, N: the total number of high-frequency coefficients of wavelet decomposition, M: the total number of signal sampling points, E noise : the noise energy ratio, used to evaluate the credibility of sensor data. The larger the value, the more serious the noise interference. The threshold setting formula is and the dynamic scaling factor range is 0.8 - 1.5;

[0060] The correlation coefficient of the historical data sliding window is obtained by calculating the Pearson correlation between the current data segment and the same working condition data segment in the previous 30 minutes. When the correlation coefficient is lower than 0.6 and the noise energy ratio is higher than 0.3, the credibility score drops to 0.2.

[0061] The basis for configuring the hyperparameters of the dilated convolutional layer, that is, the dilation rate and convolutional kernel design rules include: the dilation rate increases exponentially (2, 4, 8), matching the time window (1s / 5s / 10s) to ensure that the receptive field covers the periodic characteristics of sensor data; the convolutional kernel size follows k = 2 n + 1, n is the layer number, that is, 3×3 for the first layer, 5×5 for the second layer, and 7×7 for the third layer, to gradually extract local to global features.

[0062] Parameter initialization and training configuration include: The weights of the convolutional layer are initialized using He normal distribution, and the bias terms are initialized to 0; The optimizer is AdamW, with an initial learning rate of 3e-4, combined with the cosine annealing strategy, for 50 epochs; The loss function is weighted mean squared error: L = α·MSE(E noise )+(1-α)·MSE(R), where α = 0.7.

[0063] The feature dimensionality reduction method includes: The concatenated features are compressed to 64 channels through a 1×1 convolution, followed by batch normalization and ReLU activation;

[0064] The compression ratio calculation formula:

[0065] .

[0066] It should be further noted that in the specific implementation process, the construction method of the differentiable weight decision tree described in step (3) includes:

[0067] (3.1) Define the tree node activation function as:

[0068] , where σ is the Sigmoid function, used to map the input to the interval [0,1], ⊙ represents the Hadamard product, the weight matrices W1 and W2 are weight matrices, initialized using He normal distribution; b1, b2: bias terms, initialized to 0; x: input vector, including sensor credibility scores and operating condition parameters; f(x): node activation output, used to dynamically generate fusion weights;

[0069] (3.2) Inject four levels of random pulse noise with amplitudes of 15%, 20%, 25%, and 30% of the sensor range during the training phase. The pulse duration is 5 - 20 seconds, and the weight change rate is constrained by gradient penalty not to exceed 10%. The penalty factor λ takes values from 0.5 - 1.2;

[0070] (3.3) When it is detected that the SO2 concentration change rate exceeds 5% / s for 3 consecutive seconds, it is determined that the flue gas composition has mutated and a decision tree with a maximum depth of 8 layers is enabled. Under steady-state operating conditions, it is reduced to 4 layers, and the depth adjustment transition time is 200 ms.

[0071] Among them, the spatio-temporal distribution rule of noise injection in the adversarial training in step (3), that is, the adversarial noise injection strategy includes:

[0072] Spatial distribution: Randomly select 30% of the sensor channels to inject noise;

[0073] Temporal distribution: The noise duration follows a Poisson distribution (λ = 5 seconds) to simulate sudden interference;

[0074] Noise type: Except for impulse noise, Gaussian white noise (σ = 0.1 range) is superimposed to simulate environmental disturbances.

[0075] Among them, the depth adjustment quantization conditions are as follows:

[0076] The condition for enabling up to 8 layers of maximum depth is that the SO2 concentration change rate > 5% / s and lasts for 3 seconds, or the temperature gradient > 2°C / s and lasts for 5 seconds;

[0077] The condition for reducing to 4 layers is that the SO2 concentration fluctuation < 1% / s and the temperature gradient < 0.5°C / s for 10 minutes.

[0078] It should be further noted that in the specific implementation process, the structural optimization method of the compressed LSTM model described in step (5) is as follows:

[0079] Compress the hidden layer dimension of the standard LSTM to 256 dimensions. The gating units of the input gate, forget gate, and output gate are implemented using grouped convolution, with 4 groups and each group independently processing 16-dimensional features;

[0080] The input sequence length is fixed at 10 seconds, and the parameters are updated every 5 seconds using the sliding window method. The data normalization formula within the window is:

[0081] , where x is the original input data; μ 10min is the mean of the data in the previous 10 minutes, and σ 10min is the standard deviation of the data in the previous 10 minutes; 1e−6 is a very small constant to prevent division by zero errors; x′ is the normalized data for input to the compressed LSTM model.

[0082] It should be further noted that in the specific implementation process, the reverse deduction method of the digital twin system described in step (6) includes:

[0083] (6.1) Establish a virtual model including the structural dimensions of the reaction tower, catalyst distribution, and hydrodynamic parameters, divide it into 10 sections along the flue gas flow direction, and set the catalyst filling rate and temperature gradient parameters for each section;

[0084] (6.2) Calculate the theoretical pollutant concentration values based on the Arrhenius equation and hydrodynamic equations where C is the pollutant concentration, such as SO2, NOx; t is the time; x is the spatial coordinate along the length of the reaction tower; u is the flue gas velocity; D is the diffusion coefficient, and k is the chemical reaction rate constant, which is related to the catalyst activity;

[0085] (6.3) When the absolute value of the residual exceeds 20% of the theoretical value continuously for 5 times and the Frobenius norm ||R|| of the residual matrix F > 0.2, generate the correction coefficient matrix Cadj = 0.1(I - 1.0R) and synchronize it to the edge side every 30 seconds.

[0086] The rules for constructing the virtual model of the reaction tower include:

[0087] Mesh division: Divide it into 10 sections along the flue gas flow direction and 5 layers radially (center / transition / boundary layer);

[0088] Parameter calibration: The catalyst filling rate is fitted through the measured data of X-ray diffraction (XRD), and the diffusion coefficient D is calibrated by computational fluid dynamics (CFD) simulation;

[0089] Dynamic update: Synchronize the actual reaction tower structure parameters, such as the catalyst loss rate, once every 24 hours.

[0090] It should be further noted that, in the specific implementation process, the implementation method of the gradient penalty constraint is:

[0091] Add a regularization term for the weight change rate to the loss function:

[0092] , where The change rate of the fusion weight w with respect to time t, ‖·‖2: L2 norm, measures the change amplitude; λ: penalty factor, with a value range of 0.5 - 1.2, controlling the regularization strength; Lreg: regularization loss term, constraining the weight mutation and triggering a penalty when it exceeds 10%;

[0093] Among them, the change rate of the weight is approximated by the second derivative calculation formula as:

[0094] , when λ = 1.0, apply a strong penalty.

[0095] It should be further noted that, in the specific implementation process, the specific rules of the adjustment operation in step (7) include:

[0096] When the edge side continuously triggers three reconstructions or the cloud side feeds back two correction instructions, increase the decomposition channel number of the multi-scale spatio-temporal convolutional network to 96 channels and expand the historical data window to 60 minutes;

[0097] When there is no verification anomaly for 1 hour continuously, reduce 5% of the nodes every 10 minutes until reaching the initial 70%, and the elimination criterion is the nodes with an activation frequency lower than 10% within the recent 1 hour.

[0098] It should be further noted that, in the specific implementation process, it also includes a sensor anomaly handling mechanism:

[0099] When the credibility score of a certain sensor is continuously lower than 0.4 for 5 times, start the redundant sensor switching program;

[0100] The data of newly enabled sensors need to go through a 10 - minute calibration period. During the calibration period, the fusion weight is adjusted according to the linear increasing formula w t =min(1.0, 0.1 + 0.15t), where t is the calibration cycle number, 1 minute per cycle, and the maximum t = 10; w t : the sensor fusion weight of the t - th calibration cycle; 0.1 is the initial weight; 0.15 is the weight increasing step, ensuring a linear transition to 1.0 within 10 minutes.

[0101] It should be further noted that in the specific implementation process, the specific implementation of the grouped convolution is as follows: the input feature map is evenly divided into 4 groups according to the number of channels, and each group uses a 3×3 convolution kernel to perform independent operations with a stride of 2. The output feature map keeps the size unchanged through zero padding; the convolutional kernel parameters are shared among time steps within the same group, and the number of parameters is reduced by 40%.

[0102] It should be further noted that in the specific implementation process, the generation method of the correction coefficient matrix is as follows: ; where α is determined by gradient descent optimization with a value range of 0.05 - 0.15, β has a value range of 0.8 - 1.2, and the correction coefficient covers the leaf node parameters of the weight decision tree every 30 seconds; R: the residual matrix, representing the difference between the fused data and the theoretical value; I: the identity matrix, used to maintain the stability of matrix operations; α: the learning rate factor, with a value range of 0.05 - 0.15, controlling the correction amplitude; β: the residual weight coefficient, with a value range of 0.8 - 1.2, adjusting the influence intensity of the residual; C adj : the correction coefficient matrix, used to dynamically adjust the fusion weight.

[0103] The update fusion mechanism of the correction coefficient matrix is as follows: after the edge device receives C adj , the weight decision tree is updated according to the following rules: ; where η = 0.05 is the fusion rate, controlling the introduction intensity of the mechanism verification result; the update trigger condition is: the Frobenius norm of the residual matrix ‖R‖ F >0.2 and lasts for 5 verifications.

[0104] It should be further noted that in the specific implementation process, during the operation of the desulfurization and denitrification integrated device, first, a plurality of gas concentration sensors installed at the inlet, catalyst layer, and outlet of the reaction tower, including SO2, NOx sensors, temperature sensors, pressure sensors, and flow rate sensors, are used to collect flue gas parameter data in real time. Taking a desulfurization tower of a coal-fired power plant as an example, the sensors upload data every 200 ms, including the inlet SO2 concentration (range 0 - 5000 ppm), the temperature in the middle of the reaction tower (range 50 - 200 °C), the pressure inside the tower (-500 - 500 Pa), and the flue gas flow rate (0 - 15 m / s). The data is processed by the edge computing unit for standardization, and after eliminating the dimension difference, it is input into the implicit credibility evaluation module.

[0105] The implicit credibility evaluation module uses a multi-scale spatio-temporal convolutional network. Its three dilated convolutional layers are configured as follows: the first layer has a time window of 1 second, a dilation rate of 2, and a 3×3 convolutional kernel to extract transient fluctuation features; the second layer has a time window of 5 seconds, a dilation rate of 4, and a 5×5 convolutional kernel to capture medium-term trends; the third layer has a time window of 10 seconds, a dilation rate of 8, and a 7×7 convolutional kernel to identify long-term operating condition changes. For example, when the data of a certain temperature sensor drifts due to environmental vibration, its noise perturbation layer is decomposed by the db4 wavelet basis for 3 layers, and the energy ratio E of the high-frequency component is calculated. noise Suppose the current frame E of this sensor noise = 0.35, and the Pearson correlation coefficient R with the same operating condition data segment in the previous 30 minutes is 0.55. Then, according to the scoring rule with a weight of 2, its credibility score drops to 0.2, triggering subsequent weight reduction processing.

[0106] The credibility score and real-time operating condition parameters, such as the reaction tower temperature of 180 °C and the current ammonia injection volume of 2.5 m³ / h, are input into the differentiable weighted decision tree. The decision tree has been injected with random pulse noise of 20% of the range during the training stage, that is, ±1000 ppm for the SO2 sensor, and the weight change rate is constrained by gradient penalty. For example, when a certain pressure sensor suddenly fails and causes abnormal data, the weight matrices W1 and W2 in the activation function of the decision tree node are initialized by He, and learn the robust response to such abnormal scenarios during adversarial training, automatically reducing the fusion weight of the failed sensor to less than 0.1. When it is detected that the change rate of the SO2 concentration exceeds 5% / s for 3 consecutive seconds, such as suddenly rising from 2000 ppm to 3500 ppm, the depth of the decision tree immediately expands from 4 layers to 8 layers, and it takes 200 ms to complete the structure adjustment to ensure the fusion accuracy under complex mutation operating conditions.

[0107] The predicted value of the combined pollutant concentration, such as the predicted NOx concentration of 85 mg / m³, and the ammonia injection rate control instruction, such as increasing to 3.2 m³ / h, are output to the edge-side compressed LSTM model for time-series verification. The LSTM hidden layer is compressed to 256 dimensions. The input gate uses grouped convolution, with 4 groups, each group processing 16-dimensional features. The input sequence is fixed at 10 seconds of data, with 50 sampling points. For example, when the prediction deviation exceeds the dynamic threshold three times in a row, such as the actual measured value is 92 mg / m³ while the predicted value is 85 mg / m³, and the deviation of 7.6% exceeds the set threshold of 5%, the system immediately triggers the reconstruction of the weighted decision tree and recalculates the fusion weights of each sensor.

[0108] Among them, the boundary conditions for calculating the pollutant concentration are set, that is, the boundary conditions include:

[0109] Inlet boundary: The flue gas concentration and flow rate are input in real time by the edge-side fused data;

[0110] Outlet boundary: Free outflow condition (Neumann boundary);

[0111] Wall boundary: No-slip condition, and the reaction rate k is dynamically calculated according to the Arrhenius equation:

[0112] ; A is the pre-exponential factor, E a is the activation energy, R is the gas constant, and T is the local temperature.

[0113] At the same time, the cloud digital twin system constructs a virtual model of the reaction tower, divides it into 10 sections along the flue gas flow direction, and sets the catalyst filling rate and temperature gradient parameters for each section, such as the filling rate of section 3 is 65%;

[0114] Based on the fluid dynamics equation:

[0115] , where the flow velocity u = 8 m / s, the diffusion coefficient D = 0.15 m² / s, and the reaction rate constant k = 0.08 s -1 , calculate that the theoretical NO x concentration should be 88 mg / m³. When the actual fusion value is 85 mg / m³ and the absolute value of the residual exceeds 20% for 5 consecutive times, a correction coefficient matrix C adj = 0.1(I - 1.0R) is generated and synchronized to the edge side every 30 seconds through the MQTT protocol to adjust the leaf node parameters of the weighted decision tree.

[0116] According to the verification results between the edge side and the cloud side, when the reconstruction is triggered three times cumulatively within 24 hours or two correction instructions are received, the number of decomposition channels of the implicit evaluation module increases from 64 to 96, and the historical data window is extended to 60 minutes to retrain the convolutional kernel; if there is no abnormality for 1 hour continuously, 5% of the decision tree nodes are reduced every 10 minutes, and the nodes with an activation frequency lower than 10% are preferentially eliminated, such as a node for processing low-temperature working conditions that has been activated only 8 times recently.

[0117] For the sensor abnormality scenario, when the credibility score of a certain SO2 sensor is continuously lower than 0.4 for 5 times, it automatically switches to the backup sensor. After the new sensor is enabled, it goes through a 10-minute calibration period, and its fusion weight increases linearly according to w t =min(1.0, 0.1 + 0.15t). For example, the weight is 0.55 at the 3rd minute and reaches the normal value of 1.0 at the 10th minute. During this period, the data of the old sensor is temporarily isolated to avoid contaminating the fusion result with incorrect data.

[0118] At the model optimization level, the gradient penalty term L reg =1.0·max(0, ||∂w / ∂t||2 - 0.1) restricts the weight mutation. When an abnormal fluctuation of 15% is detected in the weight change rate during a certain update, the system automatically applies the penalty term to increase the loss function value by 1.2 times, forcing the model to roll back to the stable state. The grouped convolution module divides the 256-dimensional feature map into 4 groups, each group with 64 dimensions, and processes it with a 3×3 convolutional kernel with a stride of 2. The number of parameters is reduced from 589,824 of the standard convolution to 147,456. At the same time, zero padding is used to keep the size of the feature map unchanged to ensure the continuity of the time series prediction.

[0119] Among them, the implementation details of the gradient penalty include: calculating the gradient change rate once every 10 training steps; when ‖∂w / ∂t‖2 > 0.15, triggering gradient clipping, clip value = 0.1;

[0120] Dynamic penalty factor adjustment:

[0121] ;

[0122] Among them: λ: gradient penalty factor, controlling the constraint strength of the weight change rate;

[0123] Under steady-state working conditions: the penalty is weak, λ = 0.5, allowing the model to be fine-tuned to adapt to slow changes;

[0124] Under mutation working conditions: the penalty is strengthened, λ = 1.2, suppressing the drastic fluctuation of the weight and ensuring the stability of the system;

[0125] represents the absolute value of the change rate of the SO2 concentration; Represents the absolute value of the temperature gradient.

[0126] Through core technological innovations such as decoupling noise and operating condition characteristics via multi-scale spatio-temporal convolution, generating dynamic weights of differentiable decision trees, and double verification closed-loop between the edge and the cloud, this method can improve the reliability of data fusion under complex operating conditions while ensuring a millisecond-level response speed. In specific implementation, the dilation rate, adversarial noise amplitude, and number of section divisions in the parameters of each module are repeatedly verified and optimized through a pilot platform to ensure the feasibility of the technical solution in real industrial scenarios. The anomaly handling mechanism and adaptive parameter adjustment form a closed-loop optimization, enabling the system to have the ability of continuous learning and effectively cope with long-term changes such as equipment aging and catalyst deactivation.

[0127] Through the collaborative design of dynamic credibility assessment and differentiable weight decision trees, the control accuracy and response speed of the desulfurization and denitrification device under complex operating conditions are improved. Based on the multi-scale spatio-temporal convolution network to separate sensor noise in real time and combined with the double verification mechanism between the edge and the cloud, the system can complete the shielding of abnormal data in a short time when the flue gas composition changes suddenly, the prediction error of pollutant concentration is reduced compared with traditional methods, and the generation delay of the ammonia injection amount regulation instruction is shortened. At the same time, the lightweight adversarial training and dynamic computational flow pruning technology solve the problem of excessive model complexity, reduce the computational resource occupancy under steady-state operating conditions, achieve millisecond-level real-time control at the hardware end, and ensure the system robustness when sensors fail or the load changes suddenly.

[0128] Through precise dynamic regulation of the ammonia injection amount, the consumption of ammonia water is reduced, and the incidence rate of the risk of excessive emissions is decreased. Its adaptive parameter adjustment mechanism can be adapted to equipment aging and catalyst activity decay in the long term. The technical scalability is prominent, and the multi-scale feature extraction and edge-cloud collaborative architecture can be migrated to industrial scenarios such as boiler combustion optimization, providing an efficient and reliable technical path for the intelligent upgrading of high-energy-consuming industries.

[0129] It should be noted that in this article, relational terms such as first and second are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any actual relationship or order between these entities or operations. Moreover, the terms "include", "comprise" or any other variant thereof are intended to cover non-exclusive inclusion, so that a process, method, article or device including a series of elements not only includes those elements, but also includes other elements not expressly listed, or also includes elements inherent to such process, method, article or device. Without further limitation, an element defined by the statement "including a..." does not exclude the existence of another identical element in the process, method, article or device including the said element.

[0130] Although embodiments of the present invention have been shown and described, those of ordinary skill in the art will appreciate that various changes, modifications, substitutions and variations can be made to these embodiments without departing from the principles and spirit of the present invention. The scope of the present invention is defined by the appended claims and their equivalents.

Claims

1. An intelligent control method for a desulfurization and denitrification integrated device with multi-sensor data fusion, characterized in that, It includes the following steps: (1) Real-time collect flue gas parameter data through the gas concentration sensor, temperature sensor, pressure sensor and flow rate sensor of the desulfurization and denitrification integrated device; (2) Input the flue gas parameter data into the implicit credibility evaluation module, decompose the data into a working condition characterization layer and a noise disturbance layer through a multi-scale spatio-temporal convolutional network, and dynamically calculate the real-time credibility scores of each sensor based on the energy ratio of the noise disturbance layer and the correlation coefficient of the historical data sliding window; (3) Input the credibility scores and the current reaction tower temperature and ammonia injection amount parameters into the differentiable weight decision tree model, and generate non-linear fusion weights through the path activation function, where the differentiable weight decision tree is optimized by lightweight adversarial training to simulate sensor failure and load mutation scenarios; (4) Perform non-linear weighted fusion on the multi-sensor data according to the fusion weights, and output the pollutant concentration prediction value and the ammonia injection amount control instruction; (5) Deploy a compressed LSTM model at the edge to verify the time series of the pollutant concentration prediction value. If the prediction deviation exceeds the dynamic threshold for three consecutive times, trigger the reconstruction of the weight decision tree; (6) Reverse deduce the residual distribution of the fusion data based on the mass conservation equation through the cloud digital twin system. If the residual exceeds the limit, feedback the correction instruction to the edge to update the credibility score calculation logic; (7) Adjust the decomposition channel number of the implicit credibility evaluation module and the node depth of the weight decision tree according to the verification results of steps (5) and (6) to achieve adaptive control under dynamic working conditions.

2. The intelligent control method of the desulfurization and denitrification integrated device for multi-sensor data fusion according to claim 1, characterized in that: The implementation method of the implicit credibility evaluation module in step (2) is: The multi-scale spatio-temporal convolutional network includes three dilated convolutional layers, which extract the common features of sensor data with time windows of 1 second, 5 seconds, and 10 seconds respectively; The noise disturbance layer separates the high-frequency noise components through the adaptive wavelet threshold algorithm; The correlation coefficient of the historical data sliding window is obtained by calculating the Pearson correlation between the current data segment and the same working condition data segment in the previous 30 minutes.

3. The intelligent control method for the integrated desulfurization and denitration device with multi-sensor data fusion according to claim 2, wherein: The construction method of the differentiable weight decision tree in step (3) includes: (3.1) Define the tree node activation function as: , where σ is the Sigmoid function and ⊙ represents the Hadamard product; (3.2) Inject random pulse noise with an amplitude of 15% - 30% of the sensor range during the training stage, and constrain the weight change rate not to exceed 10% through gradient penalty; (3.3) Dynamically adjust the depth of the tree according to the real-time working condition complexity. When a sudden change in flue gas composition is detected, enable a maximum depth of 8 layers, and reduce it to 4 layers under steady-state working conditions.

4. The intelligent control method for the desulfurization and denitrification integrated device with multi-sensor data fusion according to claim 3, characterized in that: The structure optimization method of the compressed LSTM model in step (5) is: Compress the hidden layer dimension of the standard LSTM to 256 dimensions, and the gating unit is implemented by grouped convolution with 4 groups; The input sequence length is fixed at 10 seconds, and the model parameters are updated every 5 seconds. Only the training data in the most recent 1 hour is retained during the update.

5. The intelligent control method for the desulfurization and denitrification integrated device with multi-sensor data fusion according to claim 4, characterized in that: The reverse deduction method of the digital twin system in step (6) includes: (6.1) Establish a virtual model including the reaction tower structure size, catalyst distribution and fluid dynamics parameters; (6.2) Calculate the theoretical pollutant concentration value based on the Arrhenius equation and perform residual analysis with the fusion data; When the absolute value of the residual exceeds 20% of the theoretical value for five consecutive times, it is determined that the fusion is abnormal and a correction coefficient matrix is generated.

6. The intelligent control method for the integrated desulfurization and denitrification device with multi-sensor data fusion according to claim 5, characterized in that: The implementation method of the gradient penalty constraint is as follows: Add a regularization term of the weight change rate to the loss function: , where λ is a penalty factor with a value ranging from 0.5 to 1.

2.

7. The intelligent control method for the integrated desulfurization and denitration device with multi-sensor data fusion according to claim 6, wherein: The specific rules of the adjustment operation described in step (7) include: When the edge side triggers reconstruction three times continuously or the cloud feedbacks two correction instructions, increase the decomposition channel number of the multi-scale spatio-temporal convolutional network by 50%; When there is no verification abnormality for 1 hour continuously, gradually reduce the number of nodes of the weight decision tree to 70% of the initial state.

8. The intelligent control method of the desulfurization and denitrification integrated device with multi-sensor data fusion according to claim 7, characterized in that: It also includes a sensor abnormality handling mechanism: when the credibility score of a certain sensor is lower than 0.4 for five consecutive times, start the redundant sensor switching program; the data of the newly enabled sensor needs to go through a 10-minute calibration period, during which its fusion weight increases linearly to the normal value.

9. The intelligent control method for the integrated desulfurization and denitrification device with multi-sensor data fusion according to claim 8, characterized in that: The implementation process of the grouped convolution includes: evenly dividing the input feature map into 4 groups according to the number of channels, independently performing convolution operations on each group and then splicing and outputting; the convolution kernel size is set to 3×3, the stride is set to 2, and zero padding is used to keep the feature map size unchanged.

10. The intelligent control method for the desulfurization and denitrification integrated device with multi-sensor data fusion according to claim 9, characterized in that: The generation method of the correction coefficient matrix is as follows: ; where α is the learning rate factor and takes a value of 0.05 - 0.15, β is the residual weight coefficient and takes a value of 0.8 - 1.2, R is the residual matrix, and I is the identity matrix.

Images