Soft measurement method for outlet pressure of acid making fan based on LSTM (Long Short Term Memory)

Abstract

The invention belongs to the technical field of flue gas acid making, and particularly discloses an acid making fan outlet pressure soft measurement method based on LSTM (Long Short Term Memory), which comprises the following steps of: performing cleaning, down-sampling and normalization processing on fan inlet and outlet pressure original data; based on a time sequence analysis technology, extracting input features in the preprocessed data; an LSTM network is adopted to establish a soft measurement model, and long-term dependence characteristics in a time sequence are captured through a gating mechanism; performing training optimization on model hyper-parameters through grid search; and the model evaluates and predicts the performance by using the data of the test set, deploys the qualified model to an industrial control system, and predicts, regulates and controls the outlet pressure of the acid making fan in real time. According to the method, soft measurement and LSTM are combined, long-term dependency and dynamic change information in a time sequence are extracted, weighting processing is carried out on different scale features to improve prediction precision, real-time prediction of multi-step variables is achieved, and the method has the advantages of being high in prediction precision, high in timeliness and low in hardware cost.

Description

Technical Field

[0001] The invention relates to the technical field of flue gas acid production, and in particular to an LSTM-based soft measurement method for outlet pressure of an acid production fan with high prediction accuracy, strong timeliness and low hardware cost. Background Art

[0002] Sulfuric acid is an important industrial raw material. Its production process involves high concentrations of sulfur dioxide (SO 2 ) The acid-making fan plays a key role in transporting the purified flue gas to the non-equilibrium conversion process. The accurate measurement of its pressure parameters is crucial to optimizing process control and ensuring the safe operation of equipment. However, due to the high temperature, high pressure and corrosive environment of flue gas acid production, industrial sensors often have problems such as drift and noise interference, which makes it impossible to accurately measure the outlet pressure of the acid-making fan, which seriously restricts the precise control of the air volume of the acid-making fan, thereby affecting the energy consumption and quality of acid production.

[0003] Traditional pressure measurement methods mainly rely on high-precision sensors, but there are problems such as high hardware cost, difficult maintenance and delayed data collection. With the rapid development of artificial intelligence technology, data-driven soft measurement technology provides an efficient and low-cost alternative to solve such problems. For example, in the prior art, there is a soft measurement solution based on mechanism modeling that establishes a theoretical model of fan pressure through fluid mechanics and energy conservation equations, and calibrates key parameters (such as air volume coefficient) in combination with the fan characteristic curve; although its model has strong interpretability and is suitable for scenes with stable working conditions, and relies on a small amount of historical data, it is particularly suitable for initial modeling or scenes with insufficient data; but it also has problems such as poor nonlinear adaptability, difficulty in handling complex working conditions such as changes in gas composition and temperature fluctuations, and requires precise fan characteristic parameters. In actual applications, the model is prone to inaccuracy due to equipment aging. In addition, there are data-driven modeling schemes that use machine learning algorithms (such as support vector machines and RBF neural networks) to learn the nonlinear relationship between pressure and auxiliary variables (such as motor current and vibration signals) from historical data. Although it has strong nonlinear processing capabilities and is particularly suitable for complex industrial scenarios, and the model has good generalization and can adapt to operating condition drift through incremental learning, it also relies on high-quality data, data noise or missing data will lead to degraded model performance, and the model is weak in interpretability and difficult to diagnose faults. In addition, there are hybrid modeling schemes that integrate mechanism models and data-driven models, use mechanism equations to provide physical constraints, and correct model deviations through data-driven correction. This can balance accuracy and interpretability, physical constraints reduce the risk of model overfitting, and can adapt to multi-condition switching, such as processing pressure characteristics under different loads through staged modeling. However, it also has high development complexity, requires interdisciplinary knowledge (such as fluid mechanics and machine learning), and real-time performance is limited by the collaborative optimization efficiency of hybrid models.

[0004] Therefore, the current research on soft measurement methods for acid-making fan outlet pressure is still insufficient, especially the lack of in-depth exploration of time series dependence characteristics and the establishment of effective prediction models. There is an urgent need for a soft measurement method for acid-making fan outlet pressure with high prediction accuracy, strong timeliness and low hardware cost. Summary of the invention

[0005] In view of the deficiencies in the prior art, the present invention provides a LSTM-based soft measurement method for outlet pressure of an acid-making blower with high prediction accuracy, strong timeliness and low hardware cost.

[0006] The present invention is implemented in this way: including data preprocessing, feature extraction, model construction, hyperparameter optimization, model evaluation and deployment steps, the specific contents are: A. Data preprocessing: Collect the original data of the inlet and outlet pressure of the acid-making blower and perform cleaning, downsampling and normalization processing; B. Feature extraction: Based on time series analysis technology, extract the input features for prediction from the preprocessed data; C. Model construction: A soft measurement model based on the LSTM model is established using the LSTM network, and the long-term dependency characteristics in the time series are captured through the gating mechanism; D. Hyperparameter optimization: Optimize the hyperparameters of the constructed model through grid search; E. Model evaluation and deployment: The prediction performance of the trained and optimized model is evaluated using the test set data. The qualified model is deployed to the industrial control system to perform real-time prediction and regulation of the acid-making fan outlet pressure.

[0007] Furthermore, in step A, the specific process is as follows: A10. First, remove the abnormal values ​​in the raw data of the real-time pressure of the inlet and outlet of the acid-making fan and supplement the missing values; A20, then merge multiple records within two consecutive minutes into one record, and take the average value as the pressure value of this time step; A30. The data is then normalized to map data of different dimensions to the range of 0 to 1 to eliminate the impact of dimensional differences on model training.

[0008] Furthermore, the specific process of step B is: extracting the fan inlet and outlet pressure data of the first 30 time steps as input features with two minutes as a step from the preprocessed data, and correspondingly predicting the outlet pressure of the next 10 time steps.

[0009] Furthermore, the specific method of extracting input features is to adopt sliding window technology, dynamically generate input features and corresponding target data pairs with a fixed step sliding window, and then slide a fixed-size window on the data to extract input features and corresponding target values; wherein, the window size determines the amount of historical data considered, and the step size defines the interval of window movement to ensure that there is historical pressure data at each time point to support model prediction.

[0010] Furthermore, in the step C, the LSTM network structure consists of a forget gate, an input gate, and an output gate to capture the long-term dependency characteristics of the time series; Among them, the forget gate f t The formula is: , Where: W f is the weight matrix of the forget gate, [ h t-1 , x t ]To concatenate two vectors into a longer vector, b f is the bias term of the forget gate; among them, the weight matrix W f It is composed of two matrices, one corresponding to the input item h t-1 of W fh , and the other one corresponds to the input item x t of W fx , the formula is as follows: ; Input Gate i t The formula is: , In the formula, W i is the weight matrix of the input gate, b i is the bias term of the input gate; The current input state unit The formula calculated based on the last output and this input is as follows: , In the formula, W c is the weight matrix, b c is the bias term; c t is the state unit at the current moment, which is obtained by the following formula: , In the formula, c t-1 is the last status unit; Output Gate o t The formula is: , In the formula, W o is the weight matrix of the input gate, b o is the bias term of the input gate; The final output of LSTM h t By output gate o t and the current state unit c t Determine jointly, the formula is as follows: .

[0011] Furthermore, in the step C, the width of the hidden layer of the soft measurement model can be adjusted from 16 to 256, and the activation function is selected as ReLU; and the regularization constraints of L1, L2, Dropout, Early Stopping or Elastic Net are added to the LSTM network.

[0012] Furthermore, in the step D, the grid search method is used to optimize the hyperparameters, which is to use an exhaustive method to traverse all possible combinations in the predefined hyperparameter space to find the combination that optimizes the model performance.

[0013] Furthermore, the hyperparameters include network width, number of training rounds and batch size, the network width includes 16, 32, 64, 128 and 256, the number of training rounds includes 50, 100 and 200 times, and the batch size includes 32, 64 and 128.

[0014] Furthermore, in the step E, the prediction performance of the trained and optimized model is evaluated using the test set data. The trained and optimized model is verified using the test set data. The mean square error and the average relative error between the predicted value and the true value are calculated. When the mean square error and the average relative error are both less than the threshold, the prediction performance of the model can be determined to be qualified.

[0015] Furthermore, in the step E, the qualified model is deployed to the industrial control system to perform real-time prediction and regulation of the outlet pressure of the acid-making fan. The model with qualified prediction performance evaluation is deployed to the industrial control system and connected with the real-time data stream to realize the online prediction function of the outlet pressure of the acid-making fan, assist the management personnel to dynamically adjust the process parameters, and improve the process control accuracy and efficiency.

[0016] The beneficial effects of the present invention are: 1. Aiming at the highly nonlinear and dynamic characteristics of the outlet pressure of the acid-making fan, the present invention combines the soft measurement method with LSTM to construct an LSTM soft measurement model. By using the historical series of the fan inlet and outlet pressure data, the long-term dependency and dynamic change information in the time series can be effectively extracted. At the same time, the features of different scales can be weighted to improve the prediction accuracy, and the real-time prediction of multi-step variables can be realized, breaking through the SO 2 The concentration and reaction temperature limits solve the problem of inaccurate measurements caused by industrial sensor drift and data noise under high temperature, high pressure and corrosive conditions, and provide a feasible solution for the control of intake concentration and oxygen-sulfur ratio of flue gas acid production. Experimental results show that the soft measurement method of the present invention can control the average relative error within 8%, significantly improve the accuracy of fan outlet pressure prediction, reduce dependence on high-precision sensors, and improve the intelligent monitoring capabilities of smelting enterprises.

[0017] 2. The LSTM soft sensor model constructed by the present invention is data-driven. It does not require additional sensor investment. It can eliminate dimensional differences by normalizing the original data to reduce dependence on the accuracy of existing sensors. It does not require complex physical model assumptions and can use a large amount of data from industrial sites for training and optimization. At the same time, it can avoid model inaccuracy problems caused by equipment aging, effectively reducing hardware costs and simplifying maintenance difficulties.

[0018] 3. The present invention adopts LSTM network to capture the long-term dependency characteristics of time series, and dynamically adjusts the information flow through the gating mechanism (forget gate, input gate, output gate), which effectively solves the problem of poor adaptability of traditional mechanism modeling to nonlinear working conditions. In addition, the regularization method (L1, L2, Dropout) can suppress overfitting and improve the generalization ability of the model, thereby improving the prediction accuracy and nonlinear adaptability.

[0019] 4. The present invention downsamples the original data through dynamic data preprocessing, which can reduce noise interference and improve data timeliness; and uses sliding window technology to extract input features, which can significantly shorten the response time of data acquisition and prediction, thereby adapting to the high-frequency control requirements of acid-making blowers, thereby improving the real-time performance of soft measurement results and enhancing anti-interference capabilities.

[0020] 5. The present invention combines hyperparameter optimization (grid search) and regularization constraints to ensure the stability and adaptability of the model under different working conditions, and improve the robustness and generalization ability of the model. Moreover, the sliding window technology is used to dynamically generate input features and predict the outlet pressure trend for the next 10 time steps. The auxiliary management personnel can adjust the process parameters (such as air volume and rotation speed) in advance according to the predicted pressure, thereby significantly reducing the energy consumption of acid production and improving product quality. In summary, the present invention has the characteristics of high prediction accuracy, strong timeliness and low hardware cost. BRIEF DESCRIPTION OF THE DRAWINGS

[0021] Figure 1 This is a flow chart of the soft measurement method for the outlet pressure of the acid-making blower of the present invention; Figure 2 Schematic diagram of the LSTM network structure of the present invention; Figure 3 This is a comparison chart of mean square error of model tests under different hyperparameters in an embodiment of the present invention. DETAILED DESCRIPTION

[0022] In order to make the purpose, technical solution and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not used to limit the present invention.

[0023] like Figure 1 and 2 As shown, the present invention includes data preprocessing, feature extraction, model construction, hyperparameter optimization, model evaluation and deployment steps, and the specific contents are as follows: A. Data preprocessing: Collect the original data of the inlet and outlet pressure of the acid-making blower and perform cleaning, downsampling and normalization processing; B. Feature extraction: Based on time series analysis technology, extract the input features for prediction from the preprocessed data; C. Model construction: A soft measurement model based on the LSTM model is established using the LSTM (Long Short-Term Memory) network, and the long-term dependency characteristics in the time series are captured through the gating mechanism; D. Hyperparameter optimization: Optimize the hyperparameters of the constructed model through grid search; E. Model evaluation and deployment: The prediction performance of the trained and optimized model is evaluated using the test set data. The qualified model is deployed to the industrial control system to perform real-time prediction and regulation of the acid-making fan outlet pressure.

[0024] In step A, the specific process is as follows: A10. First, remove the abnormal values ​​in the raw data of the real-time pressure of the inlet and outlet of the acid-making fan and supplement the missing values; A20, then merge multiple records within two consecutive minutes into one record, and take the average value as the pressure value of this time step; A30. The data is then normalized to map data of different dimensions to the range of 0 to 1 to eliminate the impact of dimensional differences on model training.

[0025] In the step A10, linear interpolation is used to process missing values.

[0026] Time series analysis in step B: acid production SO 2 The inlet and outlet pressures of the main fan are recorded by the industrial data acquisition system (DCS), forming a set of time series data. A time series is a set of observation points arranged in chronological order, and its characteristics include: 1. Data index is timestamp, showing the time dependency between data points; 2. It usually exhibits characteristics such as trend, periodicity, randomness and stability.

[0027] The purpose of time series analysis is to predict future trends, identify patterns and support decision-making through the analysis of historical data. Its analysis methods include descriptive analysis, prediction model, causal analysis, spectral analysis and state space model. The present invention uses the prediction model in time series analysis to explore the dynamic relationship between the inlet and outlet pressures of the fan and provide a data basis for the construction of the LSTM model. The prediction model in time series analysis can effectively predict future trends by mining patterns and trends in historical data, providing data support for risk management, resource optimization and decision-making.

[0028] The specific process of step B is: extract the fan inlet and outlet pressure data of the first 30 time steps as input features with two minutes as a step from the preprocessed data, and predict the outlet pressure of the next 10 time steps accordingly.

[0029] The specific method of extracting input features is to use sliding window technology to dynamically generate input features and corresponding target data pairs with a fixed-step sliding window, and then slide a fixed-size window on the data to extract input features and corresponding target values; wherein the window size determines the amount of historical data considered, and the step size defines the interval of window movement to ensure that historical pressure data is available at each time point to support model prediction.

[0030] like Figure 2 As shown, the LSTM network structure in step C is composed of a forget gate, an input gate, and an output gate. The forget gate determines the cell state (i.e., the state unit at the current moment, c t) needs to be forgotten, the input gate determines which new information will be written into the cell state, and the output gate controls which part of the cell state will be output as the hidden state ( h t ). Candidate cell state (i.e. the state unit of the current input, ) is obtained by weighted summing the current input and the hidden state of the previous time step and passing it through the tanh activation function, which represents the candidate for new information. The cell state is the core of LSTM, which carries information in the time series and is updated through the joint action of the forget gate and the input gate. The hidden state is the final output of the LSTM unit, which combines the information of the cell state and is filtered through the output gate. The entire structure enables LSTM to effectively capture long-term dependencies in time series and solve the gradient vanishing and gradient exploding problems encountered by traditional recurrent neural networks when processing long sequences.

[0031] Among them, the forget gate f t The formula is: , Where: W f is the weight matrix of the forget gate, [ h t-1 , x t ]To concatenate two vectors into a longer vector, b f is the bias term of the forget gate; among them, the weight matrix W f It is composed of two matrices, one corresponding to the input item h t-1 of W fh , and the other one corresponds to the input item x t of W fx , the formula is as follows: ; Input Gate i t The formula is: , In the formula, W i is the weight matrix of the input gate, b i is the bias term of the input gate; The current input state unit The formula calculated based on the last output and this input is as follows: , In the formula, W c is the weight matrix, b c is the bias term; c t is the state unit at the current moment, which is obtained by the following formula: , In the formula, c t-1 is the last state unit; Output Gate o t The formula is: , In the formula, W o is the weight matrix of the input gate, b o is the bias term of the input gate; The final output of LSTM h t (i.e. the gas pressure at the fan outlet) is determined by the output gate o t and the current state unit c t Determine together, the formula is as follows: .

[0032] In the step C, the width of the hidden layer of the soft measurement model (the hidden layer width here refers to the number of neurons in the hidden layer) can be adjusted from 16 to 256, and the specific value is adjusted according to the complexity of the industrial data. If the industrial data is complex, it is usually necessary to increase the width of the hidden layer to improve the learning ability and expression ability of the model; ReLU is selected as the activation function to enhance the nonlinear fitting ability of the model; L1, L2, Dropout, Early Stopping or ElasticNet regularization constraints are added to the LSTM network.

[0033] The L1 regularization makes the model tend to be sparse by penalizing the absolute value of the weight, which is suitable for feature selection.

[0034] The L2 regularization limits the complexity of the model by adding a penalty term of the square of the weight to the loss function, thereby preventing overfitting.

[0035] The Dropout is a regularization technique that randomly discards some neurons during the training process, which can effectively prevent the model from over-relying on specific neurons, thereby improving the generalization ability.

[0036] The Early Stopping is a training strategy that stops training early when the performance on the validation set no longer improves. This method can prevent the model from overfitting on the training set.

[0037] The Elastic Net is a combination of L1 and L2 regularization, and is suitable for data sets with multicollinearity characteristics.

[0038] In the step D, the grid search method is used to optimize the hyperparameters, which is to use an exhaustive method to traverse all possible combinations in the predefined hyperparameter space to find the combination that optimizes the model performance.

[0039] The hyperparameters include network width, number of training rounds and batch size. The network width (the network width here refers to the width of the hidden layer, that is, the number of neurons in the hidden layer) includes 16, 32, 64, 128 and 256, the number of training rounds includes 50, 100 and 200 times, and the batch size includes 32, 64 and 128.

[0040] In the step E, the prediction performance of the trained and optimized model is evaluated using the test set data. The trained and optimized model is verified using the test set data. The mean square error and the average relative error between the predicted value and the true value are calculated. When the mean square error and the average relative error are both less than the threshold, the prediction performance of the model can be determined to be qualified. Figure 3 As shown in the figure, the prediction error of the optimized model is low and can meet the needs of industrial applications.

[0041] In the step E, the qualified model is deployed to the industrial control system to perform real-time prediction and regulation of the outlet pressure of the acid-making fan. The model with qualified prediction performance evaluation is deployed to the industrial control system and connected with the real-time data stream to realize the online prediction function of the outlet pressure of the acid-making fan, assist the management personnel to dynamically adjust the process parameters, and improve the process control accuracy and efficiency.

[0042] Example 1 like Figure 1 , 2 As shown in Figure 3, the outlet pressure of the high-concentration flue gas acid-making fan of a copper smelting enterprise is soft-measured. The specific process is as follows: S100: Collect the raw data of the inlet and outlet pressures of the acid-making blower and perform cleaning, downsampling and normalization processing; the specific process is as follows: S110, firstly collect the original data of the inlet and outlet pressures of the acid-making fan, and then remove the abnormal values ​​and missing values ​​in the real-time original data of the inlet and outlet pressures of the acid-making fan.

[0043] S120, then merge multiple records within two consecutive minutes into one record, and take the average value as the pressure value of the time step.

[0044] S130, then normalize the data to map data of different dimensions to a range of 0 to 1 to eliminate the impact of dimensional differences on model training.

[0045] S200: Based on time series analysis technology, extract the input features for prediction from the preprocessed data; the specific process is: extract the fan inlet and outlet pressure data of the first 30 time steps as input features (such as fan inlet flue gas temperature, etc.) from the preprocessed data with a two-minute step, and predict the outlet pressure of the next 10 time steps. This step dynamically generates input and target data pairs through the sliding window mechanism, providing rich historical information for the LSTM model.

[0046] The specific method of extracting input features is to use sliding window technology to dynamically generate input features and corresponding target data pairs with a fixed-step sliding window, and then slide a fixed-size window on the data to extract input features and corresponding target values; wherein the window size determines the amount of historical data considered, and the step size defines the interval of window movement to ensure that historical pressure data is available at each time point to support model prediction.

[0047] S300: A soft measurement model based on the LSTM model is established using an LSTM network, and the long-term dependency characteristics in the time series are captured through a gating mechanism.

[0048] The LSTM network structure consists of a forget gate, an input gate, and an output gate to capture the long-term dependency characteristics of the time series; Among them, the forget gate f t The formula is: , Where: W f is the weight matrix of the forget gate, [ h t-1 , x t ]To concatenate two vectors into a longer vector, b f is the bias term of the forget gate; among them, the weight matrix W f It is composed of two matrices, one corresponding to the input item h t-1 of W fh , and the other one corresponds to the input item x t of Wfx , the formula is as follows: ; Input Gate i t The formula is: , In the formula, W i is the weight matrix of the input gate, b i is the bias term of the input gate; The current input state unit The formula calculated based on the last output and this input is as follows: , In the formula, W c is the weight matrix, b c is the bias term; c t is the state unit at the current moment, which is obtained by the following formula: , In the formula, c t-1 is the last status unit; Output Gate o t The formula is: , In the formula, W o is the weight matrix of the input gate, b o is the bias term of the input gate; The final output of LSTM h t By output gate o t and the current state unit c t Determine jointly, the formula is as follows: .

[0049] The width of the hidden layer of the soft measurement model can be adjusted from 16 to 256, and the activation function is ReLU; the regularization constraints of L1, L2, Dropout, Early Stopping or Elastic Net are added to the LSTM network. In this embodiment, the overfitting phenomenon is suppressed by adding the Dropout regularization method.

[0050] S400: Optimize the training of the hyperparameters of the constructed model through grid search; it uses an exhaustive method to traverse all possible combinations in the predefined hyperparameter space to find the combination that optimizes the model performance.

[0051] The hyper parameters include network width, number of training rounds and batch size. The network width includes 16, 32, 64, 128 and 256, the number of training rounds includes 50, 100 and 200, and the batch size includes 32, 64 and 128. Figure 3 As shown in the figure, when the network width is 16, the number of training rounds is 100, and the batch size is 32, the mean square error (MSE) of the model on the test set is the lowest, and the mean relative error (MRE) can meet the industrial control accuracy requirements.

[0052] S500: The training and optimization model is evaluated and predicted using the test set data. The qualified model is deployed to the industrial control system, and the outlet pressure of the acid-making fan is predicted in real time. The fan outlet pressures for the next 10 time steps are: 29.265373, 29.050043, 29.093517, 28.836609, 28.866070, 29.088676, 28.754087, 2 The corresponding true values ​​are: 29.24922056, 29.47130813, 29.647862, 29.58032429, 29.33046176, 29.30769313, 29.38217412, 29.490165, 29.44078385, 29.53629278.

[0053] The calculated mean square error (MSE) of this prediction is 2.89, and the mean relative error (MRE) is 3.79%.

[0054] Among them, the prediction performance of the trained and optimized model is evaluated using the test set data. The trained and optimized model is verified using the test set data. The mean square error and average relative error between the predicted value and the true value are calculated. When the mean square error and average relative error are both less than the threshold, the prediction performance of the model can be determined to be qualified. Figure 3 As shown in the figure, the prediction error of the optimized model is low and can meet the needs of industrial applications.

[0055] Among them, the qualified model is deployed to the industrial control system to carry out real-time prediction and regulation of the outlet pressure of the acid-making fan. The model with qualified prediction performance evaluation is deployed to the industrial control system and connected with the real-time data stream to realize the online prediction function of the outlet pressure of the acid-making fan, assisting the management personnel to dynamically adjust the process parameters and improve the process control accuracy and efficiency.

[0056] Example 2 The process is the same as that in Example 1, and the fan outlet pressures for the next 10 time steps are predicted in real time as follows: 21.222704, 20.74742, 20.875397, 21.245832, 21.101397, 20.955217, 21.076937, 21.589941, 21.423767, 21.722101; and the corresponding true values ​​are 21.47685, 21.53911333, 21.51441846, 21.42049615, 21.455435, 21.33483692, 21.331383, 21.41373385, 21.370454, 21.37240667.

[0057] The calculated mean square error (MSE) of this prediction is 1.33, and the mean relative error (MRE) is 2.31%.

[0058] Example 3 The process is the same as that in Example 1, and the fan outlet pressures for the next 10 time steps are predicted in real time as follows: 27.442448, 27.367548, 27.398329, 27.401556, 27.438442, 27.46809, 27.461988, 27.483904, 27.507408, 27.568596; and the corresponding true values ​​are 27.393257, 27.460750, 27.612209, 27.677657, 27.678031, 27.728939, 27.586081, 27.826455, 27.353481, 27.409037.

[0059] The calculated mean square error (MSE) of this prediction is 1.39, and the mean relative error (MRE) is 2.76%.

[0060] The above is only a preferred specific embodiment of the present invention, but the protection scope of the present invention is not limited thereto. Any changes or substitutions that can be easily thought of by any technician familiar with the technical field within the technical scope disclosed by the present invention should be included in the protection scope of the present invention. Therefore, the protection scope of the present invention should be based on the protection scope of the claims.

Claims

1. A soft measurement method for outlet pressure of acid-making blower based on LSTM, characterized by: It includes data preprocessing, feature extraction, model construction, hyperparameter optimization, model evaluation and deployment steps. The specific contents are as follows: A. Data preprocessing: Collect the original data of the inlet and outlet pressure of the acid-making blower and perform cleaning, downsampling and normalization processing; B. Feature extraction: Based on time series analysis technology, extract the input features for prediction from the preprocessed data; C. Model construction: A soft measurement model based on the LSTM model is established using the LSTM network, and the long-term dependency characteristics in the time series are captured through the gating mechanism; D. Hyperparameter optimization: Optimize the hyperparameters of the constructed model through grid search; E. Model evaluation and deployment: The prediction performance of the trained and optimized model is evaluated using the test set data. The qualified model is deployed to the industrial control system to perform real-time prediction and regulation of the acid-making fan outlet pressure.

2. The LSTM-based soft measurement method for outlet pressure of an acid-making blower according to claim 1 is characterized in that: In step A, the specific process is as follows: A10. First, remove the abnormal values ​​in the raw data of the real-time pressure of the inlet and outlet of the acid-making fan and supplement the missing values; A20, then merge multiple records within two consecutive minutes into one record, and take the average value as the pressure value of this time step; A30. The data is then normalized to map data of different dimensions to the range of 0 to 1 to eliminate the impact of dimensional differences on model training.

3. The LSTM-based soft measurement method for outlet pressure of an acid-making blower according to claim 1 is characterized in that: The specific process of step B is: extract the fan inlet and outlet pressure data of the first 30 time steps as input features with two minutes as a step from the preprocessed data, and predict the outlet pressure of the next 10 time steps accordingly.

4. The LSTM-based soft measurement method for outlet pressure of an acid-making blower according to claim 3 is characterized in that: The specific method of extracting input features is to use sliding window technology to dynamically generate input features and corresponding target data pairs with a fixed-step sliding window, and then slide a fixed-size window on the data to extract input features and corresponding target values; wherein the window size determines the amount of historical data considered, and the step size defines the interval of window movement to ensure that historical pressure data is available at each time point to support model prediction.

5. The LSTM-based soft measurement method for outlet pressure of an acid-making blower according to claim 1 is characterized in that: In the step C, the LSTM network structure consists of a forget gate, an input gate, and an output gate to capture the long-term dependency characteristics of the time series; Among them, the forget gate f t The formula is: ， Where: W f is the weight matrix of the forget gate, [ h t-1 , x t ]To concatenate two vectors into a longer vector, b f is the bias term of the forget gate; among them, the weight matrix W f It is composed of two matrices, one corresponding to the input item h t-1 of W fh , and the other one corresponds to the input item x t of W fx , the formula is as follows: ； Input Gate i t The formula is: ， In the formula, W i is the weight matrix of the input gate, b i is the bias term of the input gate; The current input state unit The formula calculated based on the last output and this input is as follows: ， In the formula, W c is the weight matrix, b c is the bias term; c t is the state unit at the current moment, which is obtained by the following formula: ， In the formula, c t-1 is the last state unit; Output Gate o t The formula is: ， In the formula, W o is the weight matrix of the input gate, b o is the bias term of the input gate; The final output of LSTM h t By output gate o t and the current state unit c t Determine jointly, the formula is as follows: 。 6. The LSTM-based soft measurement method for outlet pressure of an acid-making blower according to claim 5 is characterized in that: In the step C, the width of the hidden layer of the soft measurement model can be adjusted from 16 to 256, and the activation function selects ReLU; the regularization constraints of L1, L2, Dropout, Early Stopping or Elastic Net are added to the LSTM network.

7. The LSTM-based soft measurement method for outlet pressure of an acid-making blower according to claim 1 is characterized in that: In the step D, the grid search method is used to optimize the hyperparameters, which is to use an exhaustive method to traverse all possible combinations in the predefined hyperparameter space to find the combination that optimizes the model performance.

8. The LSTM-based soft measurement method for outlet pressure of an acid-making blower according to claim 7 is characterized in that: The hyperparameters include network width, number of training rounds and batch size. The network width includes 16, 32, 64, 128 and 256, the number of training rounds includes 50, 100 and 200 times, and the batch size includes 32, 64 and 128.

9. The LSTM-based soft measurement method for outlet pressure of an acid-making blower according to any one of claims 1 to 8, characterized in that: In the step E, the prediction performance of the trained and optimized model is evaluated using the test set data. The trained and optimized model is verified using the test set data. The mean square error and the average relative error between the predicted value and the true value are calculated. When the mean square error and the average relative error are both less than the threshold, the prediction performance of the model can be determined to be qualified.

10. The LSTM-based soft measurement method for outlet pressure of an acid-making blower according to claim 9 is characterized in that: In the step E, the qualified model is deployed to the industrial control system to perform real-time prediction and regulation of the outlet pressure of the acid-making fan. The model with qualified prediction performance evaluation is deployed to the industrial control system and connected with the real-time data stream to realize the online prediction function of the outlet pressure of the acid-making fan, assist the management personnel to dynamically adjust the process parameters, and improve the process control accuracy and efficiency.

Images