Fireproofing spray tank steam spray and water spray collaborative control method based on deep learning

Abstract

This invention discloses a deep learning-based method for coordinated control of steam and water spraying in a flame-arresting spray tank. The method includes: real-time acquisition and preprocessing of data on exhaust gas temperature, oil concentration, lint content, and exhaust gas pressure within the tank; dynamic assessment of feature complexity using an improved Informer model and outputting predicted exhaust gas operating conditions; real-time calculation of the dynamic error signal between the predicted and actual values; estimation of the real-time total disturbance using an active disturbance rejection algorithm; automatic generation of disturbance compensation control signals to determine valve adjustment schemes; dynamic control of spraying timing, intensity, and duration based on the adjustment schemes; and real-time feedback of spraying execution results to optimize model parameters and control strategies. This invention achieves adaptive and precise control of the flame-arresting spray tank system, improving the stability and accuracy of coordinated control.

Description

Technical Field

[0001] This invention relates to the field of exhaust gas treatment and fire safety control technology for stenters in the printing and dyeing industry, and particularly to a method for coordinated control of steam spraying and water spraying in flame-retardant spray tanks based on deep learning. Background Technology

[0002] Flame arrestor spray tanks are crucial equipment in industrial production for controlling the temperature and impurities of exhaust gases, and their operational safety and stability are paramount. Currently, flame arrestor spray tanks primarily rely on steam and water spray systems to cool and decontaminate the exhaust gases inside the tank, effectively controlling impurities such as oil and lint. Traditional control strategies typically employ pre-set fixed spray patterns, relying on experience or simple threshold control methods, and using a single parameter such as exhaust gas temperature or impurity concentration to control the start and stop of steam and water sprays. Their drawback lies in the lack of effective perception and real-time response to dynamic changes in exhaust gas conditions, making it difficult to handle dynamic changes in multi-dimensional operating data such as exhaust gas temperature, oil concentration, lint content, and exhaust gas pressure.

[0003] Existing spray control methods generally do not consider the correlation between multi-dimensional operating condition data and lack accurate assessment of real-time disturbance factors in the spray system. This results in poor adaptability and low spray control accuracy when facing variable operating conditions, easily leading to under-spraying or over-spraying, reducing operational reliability and economy, and even causing additional energy consumption or safety hazards. In recent years, although some studies have proposed dynamic control methods based on simple feedback control, they still cannot comprehensively and accurately describe the dynamic characteristics of exhaust gas operating conditions, the control strategy lacks adaptability, and the overall operating effect of the spray system is difficult to significantly improve.

[0004] Therefore, how to provide a deep learning-based method for the coordinated control of steam spraying and water spraying in flame-retardant spray tanks is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0005] One objective of this invention is to propose a deep learning-based method for the coordinated control of steam and water spraying in flame-retardant spray tanks. Addressing the problem of poor adaptability in the existing technology for coordinated control of steam and water spraying, this invention proposes a technical solution based on an improved Informer model to predict exhaust gas conditions in real time, and integrating an active disturbance rejection control algorithm to estimate dynamic disturbances in real time and adaptively adjust spraying parameters. This invention possesses the technical advantages of high adaptability, good control accuracy, and strong system operational stability.

[0006] The deep learning-based coordinated control method for steam and water spraying in a flame-arresting spray tank, according to an embodiment of the present invention, includes:

[0007] Multidimensional operating condition data inside the flame-arresting spray tank are collected in real time and preprocessed to obtain a multidimensional real-time operating condition feature dataset.

[0008] By using an improved Informer model, the feature complexity of the multidimensional real-time operating condition feature dataset is dynamically judged and the attention sparse sampling ratio is adjusted in real time, and the predicted values ​​of exhaust gas temperature, oil concentration and lint content at the next moment are output.

[0009] The system collects real-time data on the actual exhaust gas temperature, oil concentration, and lint content at the next moment, and calculates these data with the predicted values ​​to obtain the real-time dynamic error signal of the exhaust gas operating conditions.

[0010] Using the real-time dynamic error signal as input, the extended state observer in the active disturbance rejection control algorithm is used to estimate the disturbance state of the exhaust gas condition, and the real-time total disturbance estimate of the exhaust gas condition is obtained.

[0011] Based on the real-time total disturbance estimate, the adjustment schemes for the steam spray valve and the water spray valve are determined.

[0012] Based on the adjustment scheme of steam spray valve and water spray valve, the spray timing, spray intensity and spray duration of steam spray and water spray are dynamically adjusted to obtain the spray coordinated control execution results;

[0013] Monitor the execution results of the spray coordinated control and feed them back to the data acquisition stage to continuously update and improve the parameters of the Informer model and the control strategy of the active disturbance rejection control algorithm.

[0014] Optionally, obtaining the multidimensional real-time operating condition feature dataset specifically involves:

[0015] Dynamic operating condition data at different vertical heights inside the flame-arresting spray tank are collected using temperature sensors, oil concentration sensors, lint content sensors, and exhaust gas pressure sensors to obtain a multi-dimensional raw operating condition data sequence.

[0016] The moving average filtering method is used to perform preliminary filtering on the multidimensional original working condition data sequence, and the median filtering method is used for secondary filtering to obtain the working condition data sequence after two filtering and denoising.

[0017] Based on the allowable range of exhaust gas temperature, oil concentration, lint content and exhaust gas pressure in the process design specifications of the flame-retardant spray tank, the working condition data sequence after two rounds of filtering and noise reduction is linearly normalized to obtain a standardized working condition data sequence.

[0018] The standardized working condition data sequence is divided into continuous data segments of uniform length, and then the trigonometric function position encoding method is used to perform feature embedding processing on each data segment to obtain the working condition feature data segment sequence.

[0019] Based on the standard timestamp carried by each data segment in the operating condition feature data segment sequence, the data segment sequences of exhaust gas temperature, oil concentration, lint content and exhaust gas pressure are aligned and synchronized in the time dimension to obtain a multi-dimensional real-time operating condition feature dataset.

[0020] Optionally, the improved Informer model outputs predicted values ​​for exhaust gas temperature, oil concentration, and lint content at the next time step, specifically as follows:

[0021] Input the multidimensional real-time operating condition feature dataset into the improved Informer model and calculate the complexity index value of each feature data segment.

[0022] The feature data segments are divided into high-complexity data segments and low-complexity data segments based on the size of the complexity index, and the self-attention sparse sampling ratio of each data segment is dynamically set according to the complexity index value.

[0023] The sparse attention weight matrix of the high-complexity data segment is obtained by adaptively increasing the self-attention sparse sampling ratio.

[0024] The sparse sampling ratio of self-attention for low-complexity data segments is adaptively reduced to obtain the sparse attention weight matrix for low-complexity data segments.

[0025] The sparse attention weight matrices corresponding to the high-complexity data segment and the low-complexity data segment are respectively processed by weight normalization, and the weight normalization result is subjected to matrix weighted summation with the input data of the corresponding data segment to obtain the fused working condition feature vector sequence.

[0026] A linear mapping operation is directly performed on the fused operating condition feature vector sequence to output the predicted values ​​of exhaust gas temperature, oil concentration, and lint content at the next time step.

[0027] Optionally, the improved Informer model includes a feature embedding module, a feature complexity dynamic analysis module, an adaptive dynamic sparse self-attention module, a feature aggregation module, and a prediction output module:

[0028] The feature embedding module is used to receive a multi-dimensional real-time working condition feature dataset, perform position encoding embedding, and obtain a feature embedding sequence carrying temporal position information.

[0029] The feature complexity dynamic analysis module is used to divide the feature embedding sequence into feature data segments and calculate the complexity index value of each feature data segment.

[0030] The adaptive dynamic sparse self-attention module is used to dynamically divide the feature data segment into high-complexity data segment and low-complexity data segment according to the complexity index value, and calculate the self-attention sparse sampling ratio corresponding to each feature data segment.

[0031] The feature aggregation module is used to perform probability sampling operations in the high-complexity data segment and the low-complexity data segment according to the self-attention sparse sampling ratio, respectively, to obtain the corresponding sparse attention weight matrix, and to perform weight normalization processing on the sparse attention weight matrix respectively to obtain the fused working condition feature vector sequence.

[0032] The prediction output module is used to calculate and output the predicted values ​​of exhaust gas temperature, oil concentration, and lint content at the next moment.

[0033] Optionally, obtaining the real-time dynamic error signal of the exhaust gas operating condition specifically includes:

[0034] Using sensors inside the flame-arresting spray tank, real-time data on exhaust gas temperature, oil concentration, and lint content at different heights of the flame-arresting spray tank are collected to obtain the actual operating condition data sequence for the next moment.

[0035] Based on the predicted values ​​of exhaust gas temperature, oil concentration, and lint content for the next moment output by the prediction output module, the actual operating condition data sequence is aligned with the predicted values ​​one by one according to the timestamp.

[0036] For each timestamp, the difference between the actual operating condition data sequence and the predicted value is calculated dimension by dimension to obtain the error of exhaust gas temperature, oil concentration and lint content at the next time point.

[0037] Using the sliding window method, with the absolute value of the operating condition error of each dimension within the current and a predetermined number of time points in the past as input, the average sequence of operating condition errors of exhaust gas temperature, oil concentration and lint content is obtained.

[0038] The comprehensive error value of exhaust gas operating conditions is calculated by weighting the errors of three dimensions—exhaust gas temperature, oil concentration, and lint content—in the average sequence of operating conditions, and is used as the real-time dynamic error signal of exhaust gas operating conditions.

[0039] Optionally, the step of using the extended state observer in the active disturbance rejection control algorithm to estimate the exhaust gas operating condition disturbance state and obtain the real-time total disturbance estimate of the exhaust gas operating condition specifically involves:

[0040] The comprehensive error value of the exhaust gas operating condition is used as a real-time dynamic error signal input to the extended state observer;

[0041] Based on the extended state equations pre-established within the extended state observer, the state observation vector of the exhaust gas operating system is constructed in real time. The state observation vector includes the exhaust gas operating state variables and the total disturbance variables.

[0042] By expanding the observation gain matrix of the state observer, the observed values ​​of the exhaust gas condition state variables in the state observation vector are updated in real time, and the observed values ​​of the total disturbance variables are updated by using the error between the state observation vector and the real-time dynamic error signal.

[0043] Based on historical operating condition data sequences, the pole placement method is used to calculate and determine the observation gain matrix of the extended state observer offline, so that the observation error of the extended state observer can approach zero within a predetermined convergence time.

[0044] The observed values ​​of the total disturbance variables in the real-time updated state observation vector are used as the real-time total disturbance estimate for the exhaust gas condition.

[0045] Optionally, the step of determining the adjustment scheme for the steam spray valve and the water spray valve based on the real-time total disturbance estimate is as follows:

[0046] Using the real-time total disturbance estimate as input, the total control action required by the exhaust gas operating system is calculated in real time using the disturbance compensation control law;

[0047] The control action allocation ratio coefficients of steam spray valves and water spray valves are determined based on historical operating data.

[0048] The valve control action of the steam spray valve and the water spray valve are calculated based on the total control action and control action distribution ratio coefficient obtained in real time.

[0049] Using a pre-established linear relationship model between valve control action and valve opening, the control action of steam spray valve and water spray valve is converted into the corresponding valve opening adjustment amount in real time.

[0050] The target valve opening values ​​for the steam spray valve and water spray valve at the next moment are calculated in real time, and the calculation results are used as the adjustment scheme for the steam spray valve and water spray valve.

[0051] Optionally, the adjustment scheme for the steam spray valve and the water spray valve is implemented to obtain the spray coordinated control execution result, specifically as follows:

[0052] Based on the adjustment schemes for the steam spray valves and water spray valves, the actual target opening values ​​for the steam spray valves and water spray valves are determined respectively.

[0053] Based on the actual target opening values ​​of the steam spray valve and the water spray valve, the corresponding steam spray flow rate and water spray flow rate are calculated using a pre-established linear conversion model between the spray valve opening and the spray flow rate.

[0054] By using real-time monitoring data of the exhaust gas conditions of the flame arrestor spray tank, and based on the predetermined spray start-up threshold, it is determined whether the start-up conditions for steam spray and water spray are met at the current moment, so as to determine the specific start-up timing for steam spray and water spray.

[0055] Using steam spray flow rate and water spray flow rate as inputs, the spray intensity of steam spray and water spray is determined in real time.

[0056] Based on the preset termination thresholds for exhaust gas temperature, oil concentration, and lint content, the actual collected exhaust gas temperature, oil concentration, and lint content data are compared in real time to determine whether steam spraying and water spraying have reached the preset spraying termination conditions at each moment, so as to determine the spraying duration of steam spraying and water spraying in real time.

[0057] The system records control parameters such as the opening degree of steam and water spray valves, spray flow rate, spray start-up timing, spray intensity, and spray duration in real time, forming the spray coordinated control execution results.

[0058] Optionally, the parameters of the continuously updated and improved Informer model and the control strategy of the active disturbance rejection control algorithm are specifically as follows:

[0059] Real-time acquisition of sprinkler coordinated control execution data to form a sprinkler coordinated control execution data sequence;

[0060] Align the spray coordinated control execution data sequence with the real-time operating data to obtain the corresponding data sequence of spray execution and exhaust gas operating conditions.

[0061] Real-time calculation of the changing trends and magnitudes of exhaust gas temperature, oil concentration, lint content, and exhaust gas pressure under the action of spraying;

[0062] Based on the changing trend and magnitude of exhaust gas conditions, the prediction error of the improved Informer model is evaluated, and the basic sparse sampling ratio of the improved Informer model is updated in real time.

[0063] Based on the spray execution data sequence and the real-time dynamic error signal sequence of the exhaust gas condition, the deviation between the real-time total disturbance estimate and the actual disturbance of the exhaust gas condition is calculated, and the adjustment direction of the observation gain matrix of the extended state observer is determined.

[0064] Calculate the actual response level of the disturbance compensation control action to the exhaust gas condition, and update the control gain constant in the disturbance compensation control law in real time.

[0065] The beneficial effects of this invention are:

[0066] (1) This invention uses an improved Informer model to dynamically predict multidimensional exhaust gas operating data in real time, integrates an active disturbance rejection control algorithm to estimate the disturbance state of exhaust gas operating conditions in real time and performs spray coordinated control, adaptively adjusts the spray timing, intensity and duration of steam and water spray, effectively improves the accuracy and stability of spray control, and enhances the system's adaptability to complex exhaust gas operating conditions.

[0067] (2) This invention achieves the accuracy and generalization of exhaust gas condition prediction by dynamically adjusting the sampling ratio of the improved Informer model attention through an adaptive dynamic sparse self-attention mechanism, significantly improving the response performance and reliability of spray coordinated control, and showing better adaptability and control effect under complex dynamic conditions of industrial flame-retardant spray tanks.

[0068] (3) In terms of real-time coordinated control of exhaust gas temperature, oil concentration and lint content, this invention estimates disturbance variables in real time by expanding the state observer and adaptively generates disturbance compensation control signals, which effectively solves the problems of poor adaptability and insufficient disturbance compensation capability of spray control in the prior art. It breaks through the bottleneck of traditional control methods being unable to respond precisely to multi-dimensional working condition changes, realizes specific and significant improvements in spray coordinated control, and effectively improves the control effect and economy of exhaust gas treatment system in industrial production process. Attached Figure Description

[0069] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings:

[0070] Figure 1 This is a flowchart of the deep learning-based collaborative control method for steam spraying and water spraying in a flame-retardant spray tank proposed in this invention. Detailed Implementation

[0071] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0072] refer to Figure 1 A deep learning-based method for coordinated control of steam and water spraying in flame-arresting spray tanks includes:

[0073] Multidimensional dynamic operating condition data of exhaust gas temperature, oil concentration, lint content and exhaust gas pressure in the flame arrestor spray tank are collected in real time. The collected data is then denoised, normalized and feature-embedded to obtain a pre-processed multidimensional real-time operating condition feature dataset.

[0074] The improved Informer model inputs a multidimensional real-time operating condition feature dataset to dynamically determine the feature complexity and adjust the attention sparse sampling ratio in real time, and outputs the predicted values ​​of exhaust gas temperature, oil concentration, and lint content for the next moment.

[0075] The system collects real-time data on the actual exhaust gas temperature, oil concentration, and lint content at the next moment, and compares and calculates these data with the predicted values ​​one by one to obtain the real-time dynamic error signal of the exhaust gas operating conditions.

[0076] Using the real-time dynamic error signal as input, the extended state observer in the active disturbance rejection control algorithm is used to estimate the disturbance state of the exhaust gas condition, and the real-time total disturbance estimate of the exhaust gas condition is obtained.

[0077] Based on the real-time total disturbance estimate, the adjustment schemes for the steam spray valves and water spray valves are determined in real time.

[0078] Based on the adjustment scheme of steam spray valve and water spray valve, the spray timing, spray intensity and spray duration of steam spray and water spray are dynamically adjusted to obtain the spray coordinated control execution results;

[0079] The system monitors the results of the spray coordinated control in real time and feeds them back to the data acquisition stage, continuously updating and improving the parameters of the Informer model and the control strategy of the active disturbance rejection control algorithm.

[0080] In this embodiment, the real-time acquisition of multi-dimensional dynamic operating condition data, including exhaust gas temperature, oil concentration, lint content, and exhaust gas pressure inside the flame-retardant spray tank, and the denoising, normalization, and feature embedding of the acquired data, yields a pre-processed multi-dimensional real-time operating condition feature dataset, specifically:

[0081] By pre-arranging temperature sensors, oil concentration sensors, lint content sensors, and exhaust gas pressure sensors at vertical intervals along the inner surface of the flame-retardant spray tank wall, dynamic operating condition data of exhaust gas temperature, oil concentration, lint content, and exhaust gas pressure at different vertical heights inside the tank are collected in real time and continuously to obtain a multi-dimensional original operating condition data sequence.

[0082] The moving average filtering method is used to perform preliminary filtering on the multidimensional original working condition data sequence to remove random noise interference. The median filtering method is then used to perform secondary filtering on the data after the preliminary filtering to eliminate impulse noise caused by environmental vibration or sensor abnormalities, resulting in a working condition data sequence after two filtering and denoising processes.

[0083] Based on the allowable range of exhaust gas temperature, oil concentration, lint content and exhaust gas pressure in the process design specifications of the flame-retardant spray tank, the maximum and minimum thresholds of each working condition data parameter are extracted and used as the upper and lower limits of normalization. Linear normalization is performed on each data point of the working condition data sequence after two filtering and noise reduction operations to obtain a fully standardized working condition data sequence.

[0084] A fixed-length sliding window method is used to slice the standardized working condition data sequence, dividing it into continuous data segments of uniform length. Then, a trigonometric function position encoding method is used to embed features into each data segment. Specifically, based on the positional order of each data segment in the complete time series, sine and cosine position encoding values ​​are calculated for each data point in the data segment and appended to the end of the feature dimension of the corresponding data point to establish the positional relationship between data points, thereby obtaining the working condition feature data segment sequence after feature embedding.

[0085] Based on the standard timestamp carried by each data segment in the working condition feature data segment sequence, the time dimension of the data segment sequences of exhaust gas temperature, oil concentration, lint content and exhaust gas pressure are aligned and synchronized. Multiple data dimensions are merged into a unified multidimensional feature sequence according to the same timestamp to obtain a multidimensional real-time working condition feature dataset.

[0086] In this embodiment, the step of inputting the multi-dimensional real-time operating condition feature dataset into the improved Informer model, dynamically judging the feature complexity and adjusting the attention sparse sampling ratio in real time, and outputting the predicted values ​​of exhaust gas temperature, oil concentration, and lint content for the next moment is as follows:

[0087] The multidimensional real-time working condition feature dataset is input into the improved Informer model, and the complexity index value of each feature data segment is calculated. The complexity index value is obtained by calculating the Euclidean distance between the data points of each dimension in each feature data segment, obtaining the distance matrix corresponding to each feature data segment, and calculating the average value of all off-diagonal elements in each distance matrix. The average value of all off-diagonal elements is used as the complexity index value of the corresponding feature data segment.

[0088] Based on the magnitude of the complexity index, the feature data segment is divided into a high-complexity data segment and a low-complexity data segment. The specific division method is as follows:

[0089] A threshold for the division of complexity index values ​​is preset. When the complexity index value of a feature data segment is greater than or equal to the division threshold, it is judged as a high-complexity data segment.

[0090] When the complexity index of a feature data segment is less than the partitioning threshold, it is judged as a low-complexity data segment.

[0091] The self-attention sparse sampling ratio for each feature data segment is dynamically set based on the difference between the complexity index value and the partitioning threshold.

[0092] The self-attention sparse sampling ratio for high-complexity data segments is adaptively increased, specifically through the following method:

[0093] Divide the difference between the complexity index value corresponding to the high-complexity data segment and the partitioning threshold by the partitioning threshold to obtain the proportional adjustment coefficient of the high-complexity data segment.

[0094] The self-attention sparse sampling ratio of the high-complexity data segment is obtained by multiplying the sum of 1 and the proportional adjustment coefficient by the basic sparse sampling ratio.

[0095] Based on the self-attention sparse sampling ratio, a probability sampling operation is performed in the high-complexity data segment to obtain the sparse attention weight matrix corresponding to the high-complexity data segment.

[0096] The self-attention sparse sampling ratio for low-complexity data segments is adaptively reduced, specifically through the following method:

[0097] Divide the difference between the partitioning threshold and the complexity index value corresponding to the low-complexity data segment by the partitioning threshold to obtain the proportion adjustment coefficient of the low-complexity data segment.

[0098] The value obtained by subtracting the proportional adjustment coefficient from 1 and then multiplying it by the basic sparse sampling ratio is used as the self-attention sparse sampling ratio for low-complexity data segments.

[0099] Based on the self-attention sparse sampling ratio, a probability sampling operation is performed in the low-complexity data segment to obtain the sparse attention weight matrix corresponding to the low-complexity data segment.

[0100] The sparse attention weight matrices corresponding to the high-complexity data segment and the low-complexity data segment are respectively processed by weight normalization to obtain their respective attention normalization weight matrices. Then, matrix weighted summation is performed on the respective attention normalization weight matrices and the input data of the corresponding data segment to obtain the fused working condition feature vector sequence.

[0101] The fused operating condition feature vector sequence is directly subjected to linear mapping operation. The specific method is as follows: a linear mapping matrix is ​​pre-constructed, and matrix multiplication is performed between the linear mapping matrix and the fused operating condition feature vector sequence. The operation result is directly mapped to the three-dimensional prediction output space of exhaust gas temperature, oil concentration, and lint content, and the predicted values ​​of exhaust gas temperature, oil concentration, and lint content at the next moment are output.

[0102] In this embodiment, the improved Informer model includes a feature embedding module, a feature complexity dynamic analysis module, an adaptive dynamic sparse self-attention module, a feature aggregation module, and a prediction output module:

[0103] The feature embedding module is used to receive a multi-dimensional real-time working condition feature dataset, perform position encoding embedding, and obtain a feature embedding sequence carrying temporal position information.

[0104] The feature complexity dynamic analysis module is used to divide the feature embedding sequence into feature data segments and calculate the complexity index value of each feature data segment. Specifically, it calculates the Euclidean distance matrix between data points in each dimension within each feature data segment and calculates the average value of all off-diagonal elements in the matrix to obtain the complexity index value corresponding to the feature data segment.

[0105] The adaptive dynamic sparse self-attention module is used to dynamically divide feature data segments into high-complexity data segments and low-complexity data segments based on complexity index values. A pre-set threshold for the complexity index values ​​is used; when the complexity index value of a feature data segment is greater than or equal to the threshold, it is determined to be a high-complexity data segment, and when the complexity index value is less than the threshold, it is determined to be a low-complexity data segment. Based on the difference between the complexity index value and the threshold, the self-attention sparse sampling ratio corresponding to each feature data segment is dynamically calculated, specifically:

[0106] For high-complexity data segments, the difference between the complexity index value corresponding to the feature data segment and the partitioning threshold is divided by the partitioning threshold to obtain the proportional adjustment coefficient of the high-complexity data segment. Then, the sum of 1 and the proportional adjustment coefficient is multiplied by the basic sparse sampling ratio to obtain the self-attention sparse sampling ratio of the high-complexity data segment.

[0107] For low-complexity data segments, the difference between the partitioning threshold and the complexity index value corresponding to the feature data segment is divided by the partitioning threshold to obtain the proportional adjustment coefficient of the low-complexity data segment. The value obtained by subtracting the proportional adjustment coefficient from 1 is then multiplied by the basic sparse sampling ratio to obtain the self-attention sparse sampling ratio of the low-complexity data segment.

[0108] The feature aggregation module is used to perform probability sampling operations in the high-complexity data segment and the low-complexity data segment according to the self-attention sparse sampling ratio to obtain the corresponding sparse attention weight matrix. After performing weight normalization processing on the sparse attention weight matrix corresponding to the high-complexity data segment and the low-complexity data segment respectively, it performs matrix weighted summation operation with the input data of the corresponding data segment to obtain the fused working condition feature vector sequence.

[0109] The prediction output module pre-constructs a linear mapping matrix, performs matrix multiplication with the fused operating condition feature vector sequence using the linear mapping matrix, and directly maps the calculation result to the three-dimensional prediction output space of exhaust gas temperature, oil concentration, and lint content, outputting the predicted values ​​of exhaust gas temperature, oil concentration, and lint content at the next moment.

[0110] In this embodiment, the real-time acquisition of actual exhaust gas temperature, oil concentration, and lint content data at the next moment, and the comparison and calculation with the predicted values ​​one by one to obtain the real-time dynamic error signal of the exhaust gas operating conditions, specifically involves:

[0111] By using temperature sensors, oil concentration sensors, and lint content sensors that are pre-arranged vertically and uniformly on the inner wall surface of the flame-arresting spray tank, the actual working condition data of exhaust gas temperature, oil concentration, and lint content at different height positions of the flame-arresting spray tank at the next moment are collected in real time, and the actual working condition data sequence at the next moment is obtained.

[0112] Based on the predicted values ​​of exhaust gas temperature, oil concentration, and lint content for the next moment output by the prediction output module, the actual operating condition data sequence is aligned with the predicted values ​​one by one according to the timestamp.

[0113] For each timestamp, the difference between the actual operating condition data sequence and the predicted value is calculated dimension by dimension, specifically as follows:

[0114] The measured value of the exhaust gas temperature at the corresponding moment in the actual operating condition data sequence is subtracted from the predicted value of the exhaust gas temperature at the corresponding predicted moment to obtain the exhaust gas temperature error at the next moment.

[0115] The oil concentration error at the next moment is obtained by subtracting the predicted value of the oil concentration at the corresponding predicted moment from the measured value of the oil concentration at the corresponding moment in the actual working condition data sequence.

[0116] The error of the lint content at the next moment is obtained by subtracting the predicted value of the lint content at the corresponding predicted moment from the measured value of the lint content at the corresponding moment in the actual working condition data sequence.

[0117] Using the sliding window method, the absolute value of the operating condition error of each dimension within the current and previous predetermined number of time points is used as input, and the moving average value of the error is calculated dimension by dimension to obtain the average sequence of operating condition errors of exhaust gas temperature, oil concentration and lint content.

[0118] The comprehensive error value of the exhaust gas operating conditions is calculated by weighting the errors in three dimensions—exhaust gas temperature, oil concentration, and lint content—from the average series of operating condition errors. Specifically:

[0119] Based on the historical operating condition data sequence and the historical adjustment execution history of the spray valve, the Pearson correlation coefficient between the historical error data and the valve adjustment range in three dimensions—exhaust gas temperature, oil concentration, and lint content—was calculated.

[0120] The weighting coefficient corresponding to the average working condition error of each dimension is determined by the proportion of the absolute value of the Pearson correlation coefficient of each dimension to the sum of the absolute values ​​of the Pearson correlation coefficients of the three dimensions.

[0121] The average values ​​of the operating conditions error in the three dimensions of exhaust gas temperature, oil concentration and lint content at the current moment are weighted and summed using the determined weighting coefficients to obtain the comprehensive error value of exhaust gas operating conditions. The comprehensive error value of exhaust gas operating conditions is then used as the real-time dynamic error signal of exhaust gas operating conditions.

[0122] In this embodiment, the step of using the real-time dynamic error signal as input and employing the extended state observer in the active disturbance rejection control algorithm to estimate the exhaust gas operating condition disturbance state, thereby obtaining the real-time total disturbance estimate of the exhaust gas operating condition, specifically involves:

[0123] The comprehensive error value of the exhaust gas operating condition is used as a real-time dynamic error signal input to the extended state observer. The extended state observer is an observer that defines the unknown disturbance of the exhaust gas operating condition system as a new extended state variable and incorporates it into the extended state equation, thereby performing real-time synchronous estimation of the exhaust gas operating condition state variables and the unknown disturbance. The extended state equation extends the unknown disturbance in the mathematical model of the exhaust gas operating condition system into the system's additional state, and combines the system state variables and the additional state variables to form the extended state observation vector, so as to realize the real-time joint observation of the system state and the total disturbance variables.

[0124] Based on the extended state equations pre-established within the extended state observer, the state observation vector of the exhaust gas operating system is constructed in real time. The state observation vector includes the exhaust gas operating state variables and the total disturbance variables.

[0125] The observation gain matrix of the extended state observer is used to update the observed values ​​of the exhaust gas condition state variables in the state observation vector in real time, and the error between the state observation vector and the real-time dynamic error signal is used to update the observed values ​​of the total disturbance variables. The observation gain matrix of the extended state observer is a set of fixed coefficient matrices determined offline using the pole placement method based on the linear extended state space model of the exhaust gas condition system. It is used to adjust the observed values ​​of the exhaust gas condition state variables and the total disturbance variables in the state observation vector in real time, so that the error between the state observation vector and the real-time dynamic error signal converges to zero. The extended state space model includes a state equation composed of the exhaust gas condition state variables and the extended disturbance variables, and an observation equation with the real-time dynamic error signal as the output.

[0126] Based on historical operating condition data sequences, the observation gain matrix of the extended state observer is calculated and determined offline using the pole placement method. The specific method is as follows:

[0127] Based on the mathematical model of the exhaust gas operating system, a linear extended state space equation is constructed, and the characteristic polynomial corresponding to the extended state space equation is calculated.

[0128] The desired characteristic polynomial of the observer closed-loop system is pre-defined by the pole placement method. The eigenvalues ​​corresponding to each closed-loop pole in the desired characteristic polynomial are substituted into the characteristic polynomial. The observation gain matrix is ​​obtained by solving the algebraic equation system, so that the error between the state observation vector of the extended state observer and the real-time dynamic error signal can converge to zero within a predetermined convergence time.

[0129] The observed values ​​of the total disturbance variables in the real-time updated state observation vector are used as the real-time total disturbance estimate for the exhaust gas condition.

[0130] In this embodiment, the step of determining the adjustment scheme for the steam spray valve and the water spray valve in real time based on the real-time total disturbance estimate specifically involves:

[0131] Using the real-time total disturbance estimate as input, the total control action required by the exhaust gas operating system is calculated in real time using the disturbance compensation control law. The disturbance compensation control law is calculated by subtracting the ratio of the real-time total disturbance estimate to the control gain constant of the exhaust gas operating system from the basic control action determined by the error between the current desired operating state and the actual operating state.

[0132] The control action allocation ratio coefficients of steam spray valves and water spray valves are determined based on historical operating condition data. The control action allocation ratio coefficients are determined by offline analysis of the linear correlation between the valve adjustment amplitude of steam spray valves and water spray valves and the changes in exhaust gas temperature, oil concentration and lint content in the historical operating condition data sequence.

[0133] The valve control actions of the steam spray valve and the water spray valve are calculated based on the total control action and control action distribution ratio coefficient obtained in real time. Specifically:

[0134] The total control action is multiplied by the control action distribution ratio coefficient corresponding to the steam spray valve to obtain the valve control action of the steam spray valve.

[0135] The total control action is multiplied by the control action distribution ratio coefficient corresponding to the water spray valve to obtain the valve control action of the water spray valve.

[0136] The sum of the control action allocation ratio coefficients corresponding to the steam spray valve and the water spray valve is 1;

[0137] Using a pre-established linear relationship model between valve control action and valve opening, the control action of steam spray valve and water spray valve is converted into the corresponding valve opening adjustment amount in real time.

[0138] The target valve opening values ​​for the steam spray valve and water spray valve at the next moment are calculated in real time, and the calculation results are used as the adjustment scheme for the steam spray valve and water spray valve.

[0139] In this embodiment, the dynamic adjustment of the spraying timing, spraying intensity, and spraying duration of the steam spraying and water spraying valves according to the adjustment scheme of the steam spraying valves and water spraying valves to obtain the spraying coordinated control execution result is specifically as follows:

[0140] Based on the adjustment schemes for the steam spray valves and water spray valves, the actual target opening values ​​for the steam spray valves and water spray valves are determined respectively.

[0141] Based on the actual target opening values ​​of the steam spray valve and the water spray valve, linear conversion models obtained in advance through valve performance calibration are used to convert the actual target opening value of the steam spray valve into the corresponding steam spray flow rate and the actual target opening value of the water spray valve into the corresponding water spray flow rate in real time.

[0142] By using real-time monitoring data of the exhaust gas conditions of the flame arrestor spray tank, and based on the predetermined spray start-up threshold, it is determined whether the start-up conditions for steam spray and water spray are met at the current moment, so as to determine the start-up timing of steam spray and water spray.

[0143] Using steam spray flow rate and water spray flow rate as input, and based on the correspondence between spray flow rate and spray intensity calibrated in advance through spray intensity, the steam spray flow rate is converted into the steam spray intensity, and the water spray flow rate is converted into the water spray intensity in real time. The spray intensity is the amount of steam or water sprayed into the flame arrestor spray tank per unit time.

[0144] Based on the preset termination thresholds for exhaust gas temperature, oil concentration, and lint content, the actual collected exhaust gas temperature, oil concentration, and lint content data are compared in real time to determine whether steam spraying and water spraying have reached the preset spraying termination conditions at each moment, so as to determine the spraying duration of steam spraying and water spraying in real time.

[0145] The system records control parameters such as the opening degree of steam and water spray valves, spray flow rate, spray start-up timing, spray intensity, and spray duration in real time, forming the spray coordinated control execution results.

[0146] In this embodiment, the real-time monitoring of the spray coordinated control execution results and the feedback to the data acquisition stage, continuously updating and improving the parameters of the Informer model and the control strategy of the active disturbance rejection control algorithm, specifically includes:

[0147] The execution data, including the actual opening degree of steam spray and water spray valves, actual spray flow rate, spray intensity, spray duration, and spray start and stop times, are collected in real time from the execution results of the spray coordinated control to form a spray coordinated control execution data sequence.

[0148] The spray coordinated control execution data sequence is aligned item by item with the real-time operating data of exhaust gas temperature, oil concentration, lint content and exhaust gas pressure collected at the corresponding time to obtain a complete data sequence corresponding to the spray execution and exhaust gas operating status.

[0149] Based on the data sequence corresponding to the spray execution and the exhaust gas condition, the changing trends and magnitudes of exhaust gas temperature, oil concentration, lint content and exhaust gas pressure under the spray coordinated control effect in each execution cycle are calculated in real time.

[0150] Based on the changing trend and magnitude of the exhaust gas conditions under the coordinated control of spraying, the error level between the predicted and actual measured values ​​of exhaust gas temperature, oil concentration and lint content of the improved Informer model is evaluated periodically. The error level is used to determine the adjustment direction and magnitude of the basic ratio of attention sparse sampling in the improved Informer model, and the basic sparse sampling ratio of the improved Informer model is updated in real time.

[0151] Based on the spray coordinated control execution data sequence and the corresponding real-time dynamic error signal sequence of the exhaust gas condition, the deviation between the estimated real-time total disturbance of the exhaust gas condition and the actual disturbance is calculated cycle by cycle, and the adjustment direction of the observation gain matrix of the extended state observer is determined by using the deviation.

[0152] Based on the data sequence corresponding to the spray execution and the exhaust gas condition, the actual response level of the disturbance compensation control action to the exhaust gas condition in each execution cycle is calculated. The response level is then used to determine the real-time update direction and adjustment range of the control gain constant in the disturbance compensation control law, and the control gain constant in the disturbance compensation control law is continuously updated.

[0153] Example 1:

[0154] To verify the feasibility of this invention in practice, it was applied to the flame-retardant spray tank of the stenter exhaust gas treatment system in a dyeing and printing plant in an industrial park. A real-time dynamic collaborative control experiment was conducted on high-temperature exhaust gas to verify the adaptability and reliability of the proposed method under complex industrial conditions. In this application scenario, traditional methods typically employ a fixed-mode threshold control strategy, i.e., setting fixed spray parameters based on experience, using only simple temperature or concentration parameters as the control basis. This approach cannot comprehensively and effectively cope with the complex dynamic changes in actual industrial exhaust gas temperature, oil concentration, lint content, and exhaust gas pressure, resulting in slow control response, low spray accuracy, poor adaptability, and significant safety hazards and energy waste.

[0155] In practice, researchers first deployed high-precision temperature sensors, oil concentration sensors, lint content sensors, and exhaust gas pressure sensors at different vertical heights on the inner wall of the flame-arresting spray tank to continuously collect dynamic data on different operating parameters in real time. Subsequently, the collected multidimensional dynamic operating data was processed sequentially using moving average filtering and median filtering to remove random and impulse noise interference. The data was then linearly normalized according to the process specifications of the flame-arresting spray tank, forming a standardized multidimensional real-time operating feature dataset. Next, a fixed-length sliding window method was used to slice the data, and a trigonometric function positional encoding method was employed to perform feature embedding, constructing an operating feature sequence with clear temporal and positional characteristics.

[0156] For the operating condition feature sequence, the improved Informer model proposed in this invention is used to calculate the Euclidean distance matrix between data points of each dimension within each feature data segment in real time, and the average value of the off-diagonal elements in the matrix is ​​determined as the complexity index. Researchers dynamically divide the feature data segments into high-complexity and low-complexity categories by setting a threshold for the complexity index, and adaptively adjust the self-attention sparse sampling ratio. Probability sampling is performed on data segments of different complexities to obtain a sparse attention weight matrix. After weighted summation with the input data matrix, the accurate predicted values ​​of exhaust gas temperature, oil concentration, and lint content at the next time step are directly output through linear mapping.

[0157] Subsequently, researchers collected real-time exhaust gas operating data for the next moment and compared it with the predicted values ​​one by one. They calculated the moving average error dimensionally using the sliding window method and, based on historical operating data, calculated the Pearson correlation coefficients between the dimensional errors of exhaust gas temperature, oil concentration, and lint content and the valve adjustment amplitude. This determined the error weighting coefficients and yielded the comprehensive error value for the exhaust gas operating conditions. This comprehensive error value served as the real-time dynamic error signal input to the extended state observer. The state observation vector was constructed in real-time, and the observation gain matrix, pre-determined using the pole placement method, was used to update the state variables and the total disturbance estimate in the observation vector in real-time.

[0158] During the control execution phase, the disturbance compensation control signal is calculated in real time using the real-time total disturbance estimate and the system control gain constant. This determines the specific adjustment schemes for the steam and water spray valves, and the corresponding steam and water spray flow rates, spray intensity, and spray duration are calculated in real time. By recording the actual valve opening, spray flow rate, start and stop times, and spray intensity in real time, the spray coordinated control execution result is formed. This result is periodically fed back to dynamically update the attention sparse sampling base ratio of the Informer model, the observation gain matrix of the extended state observer, and the control gain constant in the control law.

[0159] The table below shows a comparison between the predicted and measured values ​​of some exhaust gas operating parameters during the actual operation of the flame-retardant spray tank:

[0160] Table 1 Comparison of Predicted and Measured Performance of Exhaust Gas Conditions for Flame-Arresting Spray Tanks

[0161] Exhaust gas sample number Actual temperature measurement (°C) Temperature forecast (°C) Actual oil concentration (mg / m³) Oil contamination concentration prediction (mg / m³) Actual measured lint content (g / m³) Prediction of lint content (g / m³) Actual measured exhaust gas pressure (kPa) Exhaust gas pressure prediction (kPa) 1 185 183.2 42.5 41.9 0.65 0.67 3.45 3.42 2 192 190.8 46.7 47.1 0.72 0.70 3.52 3.50 3 176 175.4 39.8 40.2 0.59 0.58 3.41 3.43 4 198 196.7 49.3 48.9 0.78 0.77 3.60 3.57 5 181 180.5 41.2 41.5 0.63 0.62 3.44 3.45

[0162] As shown in Table 1 above, comparing the predicted and measured performance of the flame-retardant spray tank under exhaust gas conditions, the method proposed in this invention demonstrates a significant advantage in predicting the accuracy of key parameters under exhaust gas conditions. Taking exhaust gas temperature as an example, the error between the predicted and measured values ​​for each sample remains within 2℃. For sample 1, the actual temperature was 185℃, while the predicted temperature was 183.2℃, with an error of 0.97%, demonstrating high prediction accuracy. Regarding oil concentration, the deviation between the predicted and measured values ​​is generally controlled within 0.6 mg / m³. The prediction accuracy for lint content and exhaust gas pressure also reaches a high level, with the error range between the predicted and measured values ​​for each indicator significantly lower than that of traditional fixed-mode control methods.

[0163] During implementation, by dynamically adjusting the spray parameters in real time, insufficient or excessive spraying was effectively avoided. The system demonstrated outstanding performance in terms of operational stability and energy efficiency, with overall energy consumption significantly reduced compared to traditional methods, resulting in a substantial improvement in safety and economy. This embodiment fully verifies the superior performance of the method of the present invention in industrial flame-retardant spraying scenarios, and it is particularly suitable for industrial application environments with complex exhaust gas conditions that change rapidly and with frequent fluctuations in temperature and impurity content, demonstrating significant application value and promising prospects for promotion.

[0164] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A deep learning-based method for coordinated control of steam and water spraying in a flame-arresting spray tank, characterized in that, include: Multidimensional operating condition data inside the flame-arresting spray tank are collected in real time and preprocessed to obtain a multidimensional real-time operating condition feature dataset. By using an improved Informer model, the feature complexity of the multidimensional real-time operating condition feature dataset is dynamically judged and the attention sparse sampling ratio is adjusted in real time, and the predicted values ​​of exhaust gas temperature, oil concentration and lint content at the next moment are output. The improved Informer model outputs predicted values ​​for exhaust gas temperature, oil concentration, and lint content at the next time step, specifically: Input the multidimensional real-time operating condition feature dataset into the improved Informer model and calculate the complexity index value of each feature data segment. The feature data segments are divided into high-complexity data segments and low-complexity data segments based on the size of the complexity index, and the self-attention sparse sampling ratio of each data segment is dynamically set according to the complexity index value. The sparse attention weight matrix of the high-complexity data segment is obtained by adaptively increasing the self-attention sparse sampling ratio. The sparse sampling ratio of self-attention for low-complexity data segments is adaptively reduced to obtain the sparse attention weight matrix for low-complexity data segments. The sparse attention weight matrices corresponding to the high-complexity data segment and the low-complexity data segment are respectively processed by weight normalization, and the weight normalization result is subjected to matrix weighted summation with the input data of the corresponding data segment to obtain the fused working condition feature vector sequence. The fused operating condition feature vector sequence is directly subjected to linear mapping operation to output the predicted values ​​of exhaust gas temperature, oil concentration and lint content at the next moment. The improved Informer model includes a feature embedding module, a feature complexity dynamic analysis module, an adaptive dynamic sparse self-attention module, a feature aggregation module, and a prediction output module. The feature embedding module is used to receive a multi-dimensional real-time working condition feature dataset, perform position encoding embedding, and obtain a feature embedding sequence carrying temporal position information. The feature complexity dynamic analysis module is used to divide the feature embedding sequence into feature data segments and calculate the complexity index value of each feature data segment. The adaptive dynamic sparse self-attention module is used to dynamically divide the feature data segment into high-complexity data segment and low-complexity data segment according to the complexity index value, and calculate the self-attention sparse sampling ratio corresponding to each feature data segment. The feature aggregation module is used to perform probability sampling operations in the high-complexity data segment and the low-complexity data segment according to the self-attention sparse sampling ratio, respectively, to obtain the corresponding sparse attention weight matrix, and to perform weight normalization processing on the sparse attention weight matrix respectively to obtain the fused working condition feature vector sequence. The prediction output module is used to calculate and output the predicted values ​​of exhaust gas temperature, oil concentration, and lint content at the next moment. The system collects real-time data on the actual exhaust gas temperature, oil concentration, and lint content at the next moment, and calculates these data with the predicted values ​​to obtain the real-time dynamic error signal of the exhaust gas operating conditions. Using the real-time dynamic error signal as input, the extended state observer in the active disturbance rejection control algorithm is used to estimate the disturbance state of the exhaust gas condition, and the real-time total disturbance estimate of the exhaust gas condition is obtained. Based on the real-time total disturbance estimate, the adjustment schemes for the steam spray valve and the water spray valve are determined. Based on the adjustment scheme of steam spray valve and water spray valve, the spraying timing, spraying intensity and spraying duration of steam spray and water spray are dynamically adjusted to obtain the spraying coordinated control execution results; Monitor the execution results of the spray coordinated control and feed them back to the data acquisition stage to continuously update and improve the parameters of the Informer model and the control strategy of the active disturbance rejection control algorithm.

2. The method for coordinated control of steam spraying and water spraying in a flame-retardant spray tank based on deep learning as described in claim 1, characterized in that, The obtained multidimensional real-time operating condition feature dataset is specifically as follows: Dynamic operating condition data at different vertical heights inside the flame-arresting spray tank are collected using temperature sensors, oil concentration sensors, lint content sensors, and exhaust gas pressure sensors to obtain a multi-dimensional raw operating condition data sequence. The moving average filtering method is used to perform preliminary filtering on the multidimensional original working condition data sequence, and the median filtering method is used for secondary filtering to obtain the working condition data sequence after two filtering and denoising. Based on the allowable range of exhaust gas temperature, oil concentration, lint content and exhaust gas pressure in the process design specifications of the flame-retardant spray tank, the working condition data sequence after two rounds of filtering and noise reduction is linearly normalized to obtain a standardized working condition data sequence. The standardized working condition data sequence is divided into continuous data segments of uniform length, and then the trigonometric function position encoding method is used to perform feature embedding processing on each data segment to obtain the working condition feature data segment sequence. Based on the standard timestamp carried by each data segment in the operating condition feature data segment sequence, the data segment sequences of exhaust gas temperature, oil concentration, lint content and exhaust gas pressure are aligned and synchronized in the time dimension to obtain a multi-dimensional real-time operating condition feature dataset.

3. The method for coordinated control of steam spraying and water spraying in a flame-retardant spray tank based on deep learning as described in claim 1, characterized in that, The real-time dynamic error signal for obtaining the exhaust gas operating condition is specifically as follows: Using sensors inside the flame-arresting spray tank, real-time data on exhaust gas temperature, oil concentration, and lint content at different heights of the flame-arresting spray tank are collected to obtain the actual operating condition data sequence for the next moment. Based on the predicted values ​​of exhaust gas temperature, oil concentration, and lint content for the next moment output by the prediction output module, the actual operating condition data sequence is aligned with the predicted values ​​one by one according to the timestamp. For each timestamp, the difference between the actual operating condition data sequence and the predicted value is calculated dimension by dimension to obtain the error of exhaust gas temperature, oil concentration and lint content at the next time point. Using the sliding window method, with the absolute value of the operating condition error of each dimension within the current and a predetermined number of time points in the past as input, the average sequence of operating condition errors of exhaust gas temperature, oil concentration and lint content is obtained. The comprehensive error value of exhaust gas operating conditions is calculated by weighting the errors of three dimensions—exhaust gas temperature, oil concentration, and lint content—in the average sequence of operating conditions, and is used as the real-time dynamic error signal of exhaust gas operating conditions.

4. The method for coordinated control of steam spraying and water spraying in a flame-arresting spray tank based on deep learning as described in claim 1, characterized in that, The method of using the extended state observer in the active disturbance rejection control algorithm to estimate the disturbance state of the exhaust gas condition and obtain the real-time total disturbance estimate of the exhaust gas condition is as follows: The comprehensive error value of the exhaust gas operating condition is used as a real-time dynamic error signal input to the extended state observer; Based on the extended state equations pre-established within the extended state observer, the state observation vector of the exhaust gas operating system is constructed in real time. The state observation vector includes the exhaust gas operating state variables and the total disturbance variables. By expanding the observation gain matrix of the state observer, the observed values ​​of the exhaust gas condition state variables in the state observation vector are updated in real time, and the observed values ​​of the total disturbance variables are updated using the error between the state observation vector and the real-time dynamic error signal. Based on historical operating condition data sequences, the pole placement method is used to calculate and determine the observation gain matrix of the extended state observer offline, so that the observation error of the extended state observer can approach zero within a predetermined convergence time. The observed values ​​of the total disturbance variables in the real-time updated state observation vector are used as the real-time total disturbance estimate for the exhaust gas condition.

5. The method for coordinated control of steam spraying and water spraying in a flame-retardant spray tank based on deep learning as described in claim 1, characterized in that, The adjustment scheme for the steam spray valve and the water spray valve, determined based on the real-time total disturbance estimate, is as follows: Using the real-time total disturbance estimate as input, the total control action required by the exhaust gas operating system is calculated in real time using the disturbance compensation control law; The control action allocation ratio coefficients of steam spray valves and water spray valves are determined based on historical operating data. The valve control action of the steam spray valve and the water spray valve are calculated based on the total control action and control action distribution ratio coefficient obtained in real time. Using a pre-established linear relationship model between valve control action and valve opening, the control action of steam spray valve and water spray valve is converted into the corresponding valve opening adjustment amount in real time. The target valve opening values ​​for the steam spray valve and water spray valve at the next moment are calculated in real time, and the calculation results are used as the adjustment scheme for the steam spray valve and water spray valve.

6. The method for coordinated control of steam spraying and water spraying in a flame-arresting spray tank based on deep learning according to claim 1, characterized in that, The adjustment scheme for the steam spray valves and water spray valves was implemented to obtain the results of the coordinated spray control, specifically: Based on the adjustment schemes for the steam spray valves and water spray valves, the actual target opening values ​​for the steam spray valves and water spray valves are determined respectively. Based on the actual target opening values ​​of the steam spray valve and the water spray valve, the corresponding steam spray flow rate and water spray flow rate are calculated using a pre-established linear conversion model between the spray valve opening and the spray flow rate. By using real-time monitoring data of the exhaust gas conditions of the flame arrestor spray tank, and based on the predetermined spray start-up threshold, it is determined whether the start-up conditions for steam spray and water spray are met at the current moment, so as to determine the specific start-up timing for steam spray and water spray. Using steam spray flow rate and water spray flow rate as inputs, the spray intensity of steam spray and water spray is determined in real time. Based on the preset termination thresholds for exhaust gas temperature, oil concentration, and lint content, the actual collected exhaust gas temperature, oil concentration, and lint content data are compared in real time to determine whether steam spraying and water spraying have reached the preset spraying termination conditions at each moment, so as to determine the spraying duration of steam spraying and water spraying in real time. The control parameters are recorded in real time to form the spray coordinated control execution results. The control parameters include the opening degree of steam spray and water spray valves, spray flow rate, spray start time, spray intensity and spray duration.

7. The method for coordinated control of steam spraying and water spraying in a flame-arresting spray tank based on deep learning as described in claim 1, characterized in that, The parameters of the continuously updated and improved Informer model and the control strategy of the active disturbance rejection control algorithm are as follows: Real-time acquisition of sprinkler coordinated control execution data to form a sprinkler coordinated control execution data sequence; Align the spray coordinated control execution data sequence with the real-time operating data to obtain the corresponding data sequence of spray execution and exhaust gas operating conditions. Real-time calculation of the changing trends and magnitudes of exhaust gas temperature, oil concentration, lint content, and exhaust gas pressure under the action of spraying; Based on the changing trend and magnitude of exhaust gas conditions, the prediction error of the improved Informer model is evaluated, and the basic sparse sampling ratio of the improved Informer model is updated in real time. Based on the spray execution data sequence and the real-time dynamic error signal sequence of the exhaust gas condition, the deviation between the estimated real-time total disturbance of the exhaust gas condition and the actual disturbance is calculated, and the adjustment direction of the observation gain matrix of the extended state observer is determined. Calculate the actual response level of the disturbance compensation control action to the exhaust gas condition, and update the control gain constant in the disturbance compensation control law in real time.

Images