Methanol synthesis tower feature extraction and anomaly detection method based on deep learning

Abstract

The invention discloses a methanol synthesis tower feature extraction and anomaly detection method based on deep learning. The method comprises the following steps: 1, acquiring original process parameters from a methanol synthesis tower control system; 2, performing data preprocessing on the original process parameters; 3, designing and building a deep learning model suitable for methanol synthesis tower feature extraction; 4, training a deep learning model by using the preprocessed process data; 5, extracting features from the new methanol synthesis tower process data through the trained deep learning model to obtain to-be-detected data features; and 6, analyzing the characteristics of the data to be detected, detecting abnormal behaviors in the process parameter data of the methanol synthesis tower, and generating a report. According to the methanol synthesis tower feature extraction and anomaly detection method based on deep learning provided by the invention, end-to-end feature extraction is carried out on the methanol synthesis tower process parameter data, key information and features in the data are better captured, and the capability of identifying the abnormal condition of the methanol synthesis tower is improved.

Description

TECHNICAL FIELD

[0001] The present application belongs to the field of chemical engineering and chemical industry, and relates to a methanol synthesis tower feature extraction and anomaly detection method based on deep learning. BACKGROUND

[0002] Methanol (CH3OH) is an important basic chemical raw material, widely used in petroleum chemical industry, fine chemicals, energy, medicine, agriculture and other fields. Traditional methanol synthesis tower anomaly detection methods are mainly based on statistics and expert systems, but they usually face challenges such as poor adaptability to complex nonlinear problems, limited processing ability for complex interaction between multiple variables, etc. With the progress of technology, modern data-driven methods (such as machine learning and artificial intelligence) are gradually becoming a complementary or alternative solution to anomaly detection to further improve the accuracy and real-time performance of detection.

[0003] Methanol synthesis tower (Methanol Synthesis Reactor) is one of the core equipment in the methanol production process, mainly used for synthesizing methanol from synthesis gas (mixed gas of CO and H2) under the action of catalyst. It is usually a high-pressure reactor that uses multiple catalysts for reaction and combines temperature, pressure and other parameters to control the reaction rate and product selectivity.

[0004] The main reaction in the synthesis tower is: CO + 2 H2←→ CH3OH + 90 kJ / mol Under appropriate temperature and pressure, carbon monoxide and hydrogen react to form methanol. During the reaction, the catalyst accelerates the reaction rate and selectively generates methanol. For the reaction conditions of high-pressure methanol process: the reaction temperature range is usually 200~300℃; at lower temperatures, the thermodynamic equilibrium of methanol synthesis reaction is favorable for the generation of methanol. However, the reaction rate is low and the catalyst activity is poor. Increasing the temperature helps to increase the reaction rate, but it will make the reaction tend to thermal equilibrium, which may reduce the selectivity of methanol. In order to balance the reaction rate and selectivity, and the temperature is too high, which will reduce the service life of the catalyst. Therefore, the temperature range between 200℃ and 300℃ is usually selected.

[0005] The reaction pressure is 50~100 MPa; increasing the pressure is beneficial to increasing the reaction rate of CO and H2; the methanol synthesis reaction "CO + 2 H2→ CH3OH" is an exothermic reaction, and higher pressure can promote the reaction to the direction of methanol generation, promote the thermodynamic equilibrium, and thus improve the yield of methanol. High pressure increases the solubility of CO and H2 in the reactor, which helps to improve the reaction efficiency. However, too high pressure will also increase the requirements and operating costs of equipment, therefore, 50~100 MPa is usually selected to balance the cost and efficiency.

[0006] Generally, copper-based catalysts (such as Cu / ZnO / Al2O3) are used; the optimal ratio of synthesis gas for methanol synthesis reaction is 1 mol CO: 2 mol H2, and if the gas ratio deviates from this range, the reaction efficiency will decrease. Therefore, it is necessary to extract and detect abnormalities of each feature in the methanol synthesis tower to ensure normal operation. SUMMARY

[0007] The purpose of the present application is to provide a deep learning-based methanol synthesis tower feature extraction and anomaly detection method. By performing end-to-end feature extraction on methanol synthesis tower process parameter data, key information and features in the data can be better captured, and the ability to identify abnormal conditions of the methanol synthesis tower can be improved.

[0008] The technical solution adopted by the present application is a deep learning-based methanol synthesis tower feature extraction and anomaly detection method, and the specific steps are as follows: Step 1, obtaining original process parameters from the methanol synthesis tower control system; Step 2, data preprocessing of the original process parameters to obtain preprocessed process data; Step 3, designing and building a deep learning model suitable for methanol synthesis tower feature extraction; Step 4, training the deep learning model using preprocessed process data to obtain a trained deep learning model; Step 5, extracting features from new methanol synthesis tower process data using the trained deep learning model to obtain test data features; Step 6, detecting abnormal behavior in the methanol synthesis tower process parameter data by analyzing the test data features and generating a report.

[0009] The present application also has the following characteristics: The specific method of step 1 is: Using a data collector to directly collect the original process parameter data of the methanol synthesis device in the methanol synthesis tower control system; The methanol synthesis device includes a feed system, a temperature control system, a pressure control system, and a catalyst regeneration system, etc. The original process parameters include synthesis gas ratio, preheating temperature, valve opening, reaction temperature, and reaction pressure, etc.

[0010] The specific method of step 2 is: First, use the 3σ principle to identify outliers: when is considered an outlier; In the formula, x i is the observed value at a certain time, μ is the mean, σ is the standard deviation; For the detected outliers, the sliding window mean is used instead: (1), where, k is the window size, k = 5, n is the total number of data points in the window, t is the time point, is the smoothed estimate of the time point t . Noise filtering: apply Savitzky-Golay filter for smoothing: (2); where, N is the normalization coefficient, whose value is the sum of all weight coefficients in the filter, m is the half-window width; Further includes normalizing the processed protocol data, using Min-Max normalization to scale each feature to the range [0, 1]: x' = (x - min(X)) / (max(X) - min(X)) (3).

[0011] The deep learning model built in step 3 adopts a CNN-LSTM hybrid architecture, with the following specific structure: Input layer: Input shape: (T, D), where T is the time step, T = 60; D is the feature dimension, D = 12, CNN module: One-dimensional convolution layer: (4); where is the output feature of the l-th layer convolution at time step t; k = 3 is the convolution kernel size, l is the layer index, σ is the ReLU activation function; b l is the bias term of the l-th layer convolution, l represents the input vector from time to time t-k : t k+1 (5); Max pooling layer: ; where, s is the pooling step, s = 2; ​​CNN module: 2 convolution-pooling layers, filter numbers are 64 and 128 respectively; LSTM module: Bi-directional LSTM layer: Forget gate: (6); Input gate: (7); Output gate: (8); (9); (10); (11); LSTM module contains 2 layers of LSTM, each layer has 128 units, dropout=0.2; Self-attention mechanism: ; Where Q, K, V are obtained by linear transformation of the feature matrix, =64 is the key vector dimension; Output layer: Feature extraction: 128-dimensional dense layer + ReLU activation; Anomaly detection: 2-dimensional softmax output normal or abnormal.

[0012] The training of the deep learning model in step 4 includes cross-validation and hyperparameter tuning of the deep learning model, and the specific method of hyperparameter tuning is: Adjust the key parameters of model training through the system to optimize the performance, where the initial learning rate is set to 0.001 and the cosine annealing strategy is used for dynamic adjustment to balance the convergence speed and accuracy; the batch size is set to 64 to balance memory efficiency and gradient stability; the loss function uses weighted cross-entropy to alleviate the problem of data imbalance; the optimizer uses AdamW, where β1=0.9 and β2=0.999, combined with adaptive learning rate and weight decay optimization training dynamics, to finally achieve balanced improvement of model convergence efficiency and generalization ability.

[0013] The training of the deep learning model in step 4 is as follows: The normal working condition data and 70% of the known abnormal samples are divided into a training set, 15% into a validation set, and 15% into a test set. The model training adopts a phased optimization strategy. First, the preprocessed process parameter data is divided by 5-fold time series cross-validation to ensure time continuity. Then, using the AdamW optimizer with cosine annealing scheduling, the weighted cross-entropy is used as the loss function for 100 epoch iteration training. During the training, the gradient is clipped with a threshold of 1.0 to stabilize the training process. In each round of validation, the early stopping mechanism is used with a patience value of 10, and the best model is saved. At the same time, the Bayesian optimization is used to automatically search for the combination of LSTM layer number, hidden unit and other hyperparameters. Finally, the model parameters with the highest F1-score in the validation set are selected. The training process focuses on optimizing the ability of the CNN-LSTM hybrid network to extract multivariate time series features, and enhances the feature weights of key process parameters such as reaction temperature and pressure through the self-attention mechanism, so that the model can sensitively capture local abnormal fluctuation patterns while preserving long-term time series dependencies.

[0014] The features extracted in step 5 include packet header information, packet body content, and methanol synthesis tower features, etc. The packet header information includes source address, destination address, timestamp, and protocol type. The packet body content includes payload data, error check code, and data sequence number. The methanol synthesis tower features include the size of the synthesis gas flow, reaction temperature, and reaction pressure. The extracted features are indexed and classified, and then stored in a database.

[0015] The specific method of step 6 is as follows: Step 6.1, use machine learning algorithms such as support vector machines, isolated forests, and random forests to build anomaly detection models to classify and detect feature data. Step 6.2, extract features from real-time protocol data and input the extracted features into the anomaly detection model to identify potential abnormal behavior; the feature distribution includes but is not limited to mean, variance, skewness, kurtosis, etc.; based on statistical distribution and historical data, set the threshold range for anomaly detection.

[0016] Step 6.3, detect potential abnormal behavior, generate alarm information, and record abnormal data and its features. The data visualization step displays the extracted protocol features and analysis results in the form of charts or reports, and real-time displays the methanol synthesis tower process parameter status and abnormal alarms.

[0017] The abnormal behavior criteria in step 6.3 are as follows: The abnormality detection standard adopts a multi-level threshold system, a first-level alarm is defined as a single feature being greater than 2sigma and less than 3sigma, and a model abnormality probability being not 0.7-0.9; a second-level alarm is defined as two or more related features being greater than 2sigma at the same time or a model abnormality probability being greater than or equal to 0.9, and a key parameter exceeding a safety threshold; The abnormality detection standard includes temperature abnormality detection, pressure-flow coupling abnormality detection and specific condition triggering corresponding alarm detection rules for catalyst efficiency reduction: Detection rule 1: temperature abnormality detection IF (T> + 2 ) AND (dT / dt>0.5℃ / min) THEN trigger an overheating alarm; Detection rule 2: pressure-flow coupling abnormality IF (P> + ) AND (F< – ) THEN trigger a blockage alarm; Detection rule 3: catalyst efficiency reduction IF (conv_rate<0.8×baseline) AND (T fluctuation>1 ) THEN trigger a catalyst deactivation alarm.

[0018] The present application has the following advantages: (1) The methanol synthesis tower feature extraction and abnormality detection method based on deep learning of the present application uses a deep learning model to extract features from methanol synthesis tower process parameter data, more accurately captures key information and features in the data, and through machine learning algorithms, real-time feature extraction and analysis of methanol synthesis tower process data can realize real-time monitoring of the running state of the methanol synthesis tower control system, quickly and accurately detect abnormal conditions, strengthen the prevention ability of the methanol synthesis tower, take corresponding measures for early warning and response, and effectively prevent serious failures and safety accidents of the methanol synthesis tower; (2) The methanol synthesis tower feature extraction and abnormality detection method based on deep learning of the present application uses an automated deep learning model for feature extraction and abnormality detection, which can reduce the need for manual intervention, reduce the possibility of misjudgment, and improve the reliability and stability of the monitoring system; the deep learning model can comprehensively and deeply analyze the methanol synthesis tower process parameter data, help operation and maintenance personnel better understand the running state and problem condition of the methanol synthesis tower, guide operation and maintenance decision-making and management work, improve the operation and maintenance efficiency and management level of the methanol synthesis tower, reduce operation and maintenance cost and risk, and ensure methanol production. BRIEF DESCRIPTION OF DRAWINGS

[0019] Figure 1 This is a schematic diagram of the feature extraction and anomaly detection method for methanol synthesis tower based on deep learning according to the present invention. Detailed Implementation

[0020] The present invention will now be described in detail with reference to the accompanying drawings and specific embodiments.

[0021] This invention relates to a method for feature extraction and anomaly detection in methanol synthesis towers based on deep learning, such as... Figure 1 As shown, the steps are as follows: Step 1: Obtain the original process parameters from the methanol synthesis tower control system; the specific method is to directly collect the original process parameter data of the methanol synthesis unit from the methanol synthesis tower control system using a data acquisition device. The methanol synthesis unit includes a feed system, a temperature control system, a pressure control system, and a catalyst regeneration system. The original process parameters include the syngas ratio, preheating temperature, valve opening, reaction temperature, and reaction pressure.

[0022] Step 2: Perform data preprocessing on the original process parameters to obtain preprocessed process data; the specific method is as follows: First, the 3σ principle is used to identify outliers: when If it is, then it is considered an outlier; In the formula, xi is the observed value at a certain moment, μ is the mean, and σ is the standard deviation; For detected outliers, replace them with the sliding window mean: (1), In the formula, k is the window size, k=5, n is the total number of data points within the window, and t is the time point. This is the smoothed estimate of time point t; that is, if xi is detected as an outlier, the mean of the sliding window with k=5 is calculated and used to replace the original data xi.

[0023] Noise filtering: Smoothing is performed using a Savitzky-Golay filter. (2); Where N is the normalization coefficient, and its value is all the weight coefficients in the filter. The sum of , where m is the width of half the window; Further steps include normalizing the processed protocol data, using Min-Max normalization to scale each feature to the [0,1] range: x' = (x - min(X)) / (max(X) - min(X)) (3).

[0024] Step 3, design and build a deep learning model suitable for feature extraction of methanol synthesis column; the deep learning model built adopts a CNN-LSTM hybrid architecture, and the specific structure is as follows: Input layer: Input shape: (T, D), where T is the time step, T = 60; D is the feature dimension, D = 12, CNN module: One-dimensional convolutional layer: (4); Where is the output feature of the lth layer convolution at time step t; k = 3 is the convolution kernel size, l is the layer index, σ is the ReLU activation function; b l is the bias term of the lth layer convolution, represents the input vector from time t-k to time t : k+1 (5); Max pooling layer: ; Where, s is the pooling step, s = 2; The CNN module has 2 convolution-pooling layer pairs, with filter numbers of 64 and 128 respectively; LSTM module: Bidirectional LSTM layer: Forget gate: (6); Input gate: (7); Output gate: (8); (9); (10); (11); The LSTM module contains 2 layers of LSTM, each with 128 units and dropout = 0.2; Self-attention mechanism: ; Where Q, K, and V are obtained by linear transformation of the feature matrix, = 64 is the key vector dimension; Output layer: Feature extraction: 128-dimensional dense layer + ReLU activation; ​Anomaly detection: 2-dimensional softmax outputs normal or abnormal.

[0025] Step 4, training the deep learning model using the preprocessed process data to obtain a trained deep learning model: Training the deep learning model includes cross-validation and hyperparameter tuning of the deep learning model, wherein the hyperparameter tuning specifically includes: Adjusting the model training key parameters through the system to optimize the performance, wherein the learning rate is initially set to 0.001 and dynamically adjusted using the cosine annealing strategy to balance the convergence speed and accuracy; the batch size is set to 64 to balance the memory efficiency and gradient stability; the loss function uses weighted cross-entropy to alleviate the data imbalance problem through class weight; the optimizer uses AdamW, wherein β1=0.9, β2=0.999, combined with adaptive learning rate and weight decay optimization training dynamics, to finally achieve balanced improvement of model convergence efficiency and generalization ability.

[0026] Training the deep learning model specifically as follows: Divide 70% of the normal operating condition data and known abnormal samples into a training set, 15% into a validation set, and 15% into a test set; the model training uses a phased optimization strategy, first performs 5-fold time series cross-validation division on the preprocessed process parameter data to ensure time continuity; then uses the AdamW optimizer with cosine annealing scheduling to perform 100 epoch iterative training with weighted cross-entropy as the loss function, during which gradient clipping is used with a threshold of 1.0 to stabilize the training process; an early stopping mechanism is used in each validation with a patience value of 10 to save the best model, and Bayesian optimization is used to automatically search for hyperparameter combinations such as LSTM layer number and hidden unit to finally select the model parameters with the highest F1-score in the validation set; the training process focuses on optimizing the CNN-LSTM hybrid network's ability to extract multivariate time series features, and enhances the feature weights of key process parameters such as reaction temperature and pressure through self-attention mechanisms, so that the model can sensitively capture local abnormal fluctuation patterns while preserving long-term time series dependencies.

[0027] Step 5, extracting features from the new methanol synthesis tower process data using the trained deep learning model to obtain the to-be-tested data features; Wherein, the extracted features include data packet header information, data packet body content, and methanol synthesis tower features, etc. The data packet header information includes source address, destination address, timestamp, and protocol type; The data packet body content includes payload data, error check code, and data sequence number; The methanol synthesis tower features include the size of the synthesis gas flow, reaction temperature, and reaction pressure; After feature extraction, feature storage is also required: including indexing and classifying the extracted features, and then storing the extracted features in the database to improve query efficiency.

[0028] Step 6, detect abnormal behavior in methanol synthesis column process parameter data by analyzing the characteristics of the data to be tested, and generate a report. The specific method is as follows: Step 6.1, use machine learning algorithms such as support vector machines, isolated forests, and random forests to build an anomaly detection model to classify and detect feature data; Step 6.2, extract features from real-time protocol data and input the extracted features into the anomaly detection model to identify potential abnormal behavior; the feature distribution includes but is not limited to mean, variance, skewness, kurtosis, etc.; based on statistical distribution and historical data, set the threshold range for anomaly detection.

[0029] Step 6.3, detect potential abnormal behavior, generate alarm information, and record abnormal data and its features; Step 7, visualize the data by displaying the extracted protocol features and analysis results in the form of charts or reports, and display the methanol synthesis column process parameter state and abnormal alarm in real time.

[0030] In step 6.3, the abnormal behavior criteria are as follows: The anomaly detection criteria use a multi-level threshold system, with a first level alarm defined as a single feature exceeding the 2σ range but less than 3σ, and the model anomaly probability between 0.7 and 0.9; a second level alarm is defined as two or more related features simultaneously exceeding the 2σ range or the model anomaly probability greater than or equal to 0.9, and the key parameters exceeding the safety threshold; The anomaly detection criteria include temperature anomaly detection, pressure-flow coupling anomaly detection, and specific conditions triggering corresponding alarm detection rules for catalyst efficiency decline: Detection rule 1: temperature anomaly detection IF (T> + 2 ) AND (dT / dt>0.5℃ / min) THEN trigger overheat alarm; Detection rule 2: pressure-flow coupling anomaly IF (P> + ) AND (F< – ) THEN trigger blockage alarm; Detection rule 3: catalyst efficiency decline IF (conv_rate<0.8×baseline) AND (T fluctuation>1 ) THEN trigger catalyst deactivation alarm.

[0031] Example 1 A deep learning-based method for feature extraction and anomaly detection in methanol synthesis towers, with the following specific steps: Step 1: Obtain the original process parameters from the methanol synthesis tower control system; Step 2: Perform data preprocessing on the original process parameters to obtain preprocessed process data; Step 3: Design and build a deep learning model suitable for feature extraction in methanol synthesis towers; Step 4: Use the preprocessed process data to train the deep learning model to obtain the trained deep learning model. Step 5: Extract features from the new methanol synthesis tower process data using the trained deep learning model to obtain the features of the data to be tested; Step 6: Analyze the characteristics of the data to be tested, detect abnormal behavior in the process parameter data of the methanol synthesis tower, and generate a report.

[0032] Example 2 The method in this embodiment is the same as that in embodiment 1. The specific method of step 1 in embodiment 1 is as follows: The raw process parameters of the methanol synthesis unit in the methanol synthesis tower control system are directly collected using a data acquisition device. A methanol synthesis unit includes a feed system, a temperature control system, a pressure control system, and a catalyst regeneration system, etc. The original process parameters include the syngas ratio, preheating temperature, valve opening, reaction temperature, and reaction pressure.

[0033] Example 3 The method in this embodiment is the same as that in embodiment 1. The specific method of step 2 in embodiment 1 is as follows: First, the 3σ principle is used to identify outliers: when If it is, then it is considered an outlier; In the formula, x i For a given moment, the observed value. μ The mean, σ Standard deviation; For detected outliers, replace them with the sliding window mean: (1), In the formula, k For window size, k =5, where n is the total number of data points within the window. t For a point in time, For time points t The smoothed estimate; Noise filtering: Savitzky-Golay filter is applied for smoothing processing: (2); wherein, N is a normalization coefficient, whose value is the sum of all weight coefficients in the filter, m is the half window width; Further comprising normalizing the processed protocol data, each feature is scaled to the range of [0, 1] by Min-Max normalization: x' = (x - min(X)) / (max(X) - min(X)) (3).

[0034] Example 4 The method of this embodiment is the same as that of Example 1, and the deep learning model built in step 3 in Example 1 adopts a CNN-LSTM hybrid architecture, and the specific structure is as follows: Input layer: Input shape: (T, D), wherein T is the time step, T = 60; D is the feature dimension, D = 12, CNN module: One-dimensional convolution layer: (4); wherein is the output feature of the lth layer convolution at time step t; k = 3 is the convolution kernel size, l is the layer index, σ is the ReLU activation function; b l is the bias term of the lth layer convolution, l represents the input vector from time to time t-k t k+1 (5); Max pooling layer: ; wherein, s is the pooling step, s = 2; The CNN module has 2 convolution-pooling layer pairs, and the filter numbers are 64 and 128 respectively; LSTM module: Bidirectional LSTM layer: Forget gate: (6); Input gate: (7); Output gate: (8);​​​ (9); (10); (11); The LSTM module contains 2 layers of LSTM, each with 128 units, dropout = 0.2; Self-attention mechanism: ; where Q, K, V are obtained by linear transformation of the feature matrix, = 64 is the key vector dimension; Output layer: Feature extraction: 128-dimensional dense layer + ReLU activation; Anomaly detection: 2-dimensional softmax output normal or abnormal.

[0035] Example 5 The method of this embodiment is the same as that of Example 1, and the training of the deep learning model in step 4 of Example 1 includes cross-validation and hyperparameter tuning of the deep learning model, wherein the specific method of hyperparameter tuning is as follows: The model training key parameters are adjusted by the system to optimize the performance, wherein the learning rate is initially set to 0.001 and is dynamically adjusted using the cosine annealing strategy to balance the convergence speed and accuracy; the batch size is set to 64 to balance the memory efficiency and gradient stability; the loss function uses weighted cross-entropy to alleviate the data imbalance problem through class weight; the optimizer uses AdamW, wherein β1=0.9, β2=0.999, combined with adaptive learning rate and weight decay optimization training dynamics, finally realizing the balanced improvement of model convergence efficiency and generalization ability.

[0036] The training of the deep learning model in step 4 is as follows: The normal working condition data and 70% of the known abnormal samples are divided into a training set, 15% into a validation set, and 15% into a test set. The model training adopts a phased optimization strategy. First, the preprocessed process parameter data is divided by 5-fold time series cross-validation to ensure time continuity. Then, using the AdamW optimizer with cosine annealing scheduling, the weighted cross-entropy is used as the loss function for 100 epoch iteration training. During the training process, the gradient is clipped with a threshold of 1.0 to stabilize the training process. In each round of validation, the early stopping mechanism is used with a patience value of 10, and the best model is saved. At the same time, the Bayesian optimization is used to automatically search for the combination of LSTM layer number, hidden unit, etc. Finally, the model parameter with the highest F1-score in the validation set is selected. The training process focuses on optimizing the ability of the CNN-LSTM hybrid network to extract multivariate time series features, and enhances the feature weights of key process parameters such as reaction temperature and pressure through the self-attention mechanism, so that the model can retain long-term time series dependencies while accurately capturing local abnormal fluctuation patterns.

[0037] Example 6 The method of this embodiment is the same as that of Example 1. The features extracted in step 5 of Example 1 include packet header information, packet body content, and methanol synthesis tower features, etc. The packet header information includes source address, destination address, timestamp, and protocol type. The packet body content includes payload data, error check code, and data sequence number. The methanol synthesis tower features include the size of the synthesis gas flow, the reaction temperature, and the reaction pressure. The extracted features are indexed and classified, and then stored in a database.

[0038] Example 7 The method of this embodiment is the same as that of Example 1. The specific method of step 6 of Example 1 is as follows: Step 6.1, use machine learning algorithms such as support vector machines, isolated forests, and random forests to build an anomaly detection model to classify and detect feature data. Step 6.2, extract features from real-time protocol data and input them into the anomaly detection model to identify potential abnormal behavior; the feature distribution includes but is not limited to mean, variance, skewness, kurtosis, etc.; based on statistical distribution and historical data, set the threshold range for anomaly detection.

[0039] Step 6.3, detect potential abnormal behavior, generate alarm information, and record abnormal data and its features. The data visualization step displays the extracted protocol features and analysis results in the form of charts or reports, and real-time displays the methanol synthesis tower process parameter state and abnormal alarm.

[0040] In step 6.3, the abnormal behavior criteria are as follows: The abnormality detection criteria use a multi-level threshold system. A first-level alert is defined as a single feature greater than 2σ and less than 3σ, and a model abnormality probability of 0.7-0.9. A second-level alert is defined as two or more related features simultaneously greater than 2σ or a model abnormality probability greater than or equal to 0.9, and a key parameter exceeding a safety threshold. The abnormality detection criteria include temperature abnormality detection, pressure-flow coupling abnormality detection, and specific conditions for catalyst efficiency decline to trigger corresponding alarm detection rules: Detection rule 1: Temperature abnormality detection IF (T> + 2 ) AND (dT / dt>0.5℃ / min) THEN trigger an overheating alarm. Detection rule 2: Pressure-flow coupling abnormality IF (P> + ) AND (F< – ) THEN trigger a blockage alarm. Detection rule 3: Catalyst efficiency decline IF (conv_rate<0.8×baseline) AND (T fluctuation>1 ) THEN trigger a catalyst deactivation alarm.

Claims

1. A method for feature extraction and anomaly detection of a methanol synthesis column based on deep learning, characterized in that, The steps are as follows: Step 1, obtain the original process parameters from the methanol synthesis tower control system; Step 2, data preprocessing is performed on the original process parameters to obtain preprocessed process data; Step 3, design and build a deep learning model suitable for feature extraction of the methanol synthesis tower; Step 4, train the deep learning model using preprocessed process data to obtain a trained deep learning model; Step 5, extract features from new methanol synthesis tower process data using the trained deep learning model to obtain test data features; Step 6, analyze the test data features to detect abnormal behavior in the methanol synthesis tower process parameter data and generate a report. 2.The method of claim 1, wherein, The specific method of step 1 is: Use a data collector to directly collect the original process parameter data of the methanol synthesis device in the methanol synthesis tower control system; The methanol synthesis device includes a feed system, a temperature control system, a pressure control system, and a catalyst regeneration system, etc. The original process parameters include synthesis gas ratio, preheating temperature, valve opening, reaction temperature, and reaction pressure, etc. 3.The method of claim 1, wherein, The specific method of step 2 is: First, the 3σ rule is used to identify outliers: when is considered as an outlier. wherein x i is the observation at a certain time, μ is the mean, σ is the standard deviation; For detected outliers, use a sliding window mean to replace them: （1）， wherein k is the window size, k = 5, n is the total number of data points in the window, t is the time point, is the smoothed estimate of the time point t is the smoothed estimate of the time point Noise filtering: apply a Savitzky-Golay filter for smoothing: （2）； wherein N is a normalization coefficient having a value equal to the sum of all weight coefficients in the filter, m is a half-window width; Further including normalizing the processed protocol data, using Min-Max normalization to scale each feature to the range [0, 1]: x' = (x - min(X)) / (max(X) - min(X)) (3). 4.The method of claim 3, wherein, The deep learning model built in step 3 uses a CNN-LSTM hybrid architecture, with the following specific structure: Input layer: Input shape: (T, D), where T is the time step, T=60; D is the feature dimension, D=12, CNN module: One-dimensional convolutional layer: (4); wherein is the output feature of the l-th layer convolution at time step t; k = 3 is the kernel size, l is the layer index, σ is the ReLU activation function; b l is the output feature of the l-th l layer convolution, denotes the input vector from time step t-k to time step t with a length of k+1 time steps. （5）； max-pooling layer: ​ wherein s is the pooling step size, s = 2; The CNN module has 2 convolution-pooling layers, with filter numbers of 64 and 128 respectively; LSTM module: Bidirectional LSTM layer: forget gate: (6); input gate: (7); Output gate: (8); （9）； （10）； （11）； The LSTM module contains 2 layers of LSTM, with 128 units each and dropout=0.2; Self-attention mechanism: ; where Q, K, V are obtained by linear transformation of the feature matrix, respectively, = 64 is the dimension of the key vector. Output layer: Feature extraction: 128-dimensional dense layer + ReLU activation; Abnormality detection: 2-dimensional softmax output normal or abnormal.

5. The feature extraction and anomaly detection method for methanol synthesis column based on deep learning according to claim 4, characterized in that, The training of the deep learning model in step 4 includes cross-validation and hyperparameter tuning of the deep learning model, where the specific method of hyperparameter tuning is: Performance was optimized by systematically adjusting key training parameters of the model. The learning rate was initially set to 0.001 and dynamically adjusted using a cosine annealing strategy to balance convergence speed and accuracy. The batch size was set to 64 to balance memory efficiency and gradient stability. The loss function used was weighted cross-entropy. The problem of data imbalance is alleviated by class weights; the optimizer AdamW is selected, with β1=0.9 and β2=0.

999. The training dynamics are optimized by combining adaptive learning rate and weight decay, and finally the model convergence efficiency and generalization ability are balanced and improved. 6.The method of claim 5, wherein, The training of the deep learning model in step 4 is as follows: The normal working condition data and 70% of the known abnormal samples are divided into a training set, 15% into a validation set, and 15% into a test set. The model training adopts a phased optimization strategy. First, the preprocessed process parameter data is divided by 5-fold time series cross-validation to ensure time continuity. Then, using the AdamW optimizer with cosine annealing scheduling, the weighted cross-entropy is used as the loss function for 100 epoch iteration training. During the training, the gradient is clipped with a threshold of 1.0 to stabilize the training process. In each round of validation, the early stopping mechanism is used with a patience value of 10, and the best model is saved. At the same time, the Bayesian optimization is used to automatically search for the hyperparameter combination of LSTM layer number and hidden unit, and the model parameter with the highest F1-score in the validation set is finally selected. The training process focuses on optimizing the ability of the CNN-LSTM hybrid network to extract multivariate time series features, and enhances the feature weights of key process parameters such as reaction temperature and pressure through the self-attention mechanism, so that the model can sensitively capture local abnormal fluctuation patterns while preserving long-term time series dependencies.

7. The feature extraction and anomaly detection method for methanol synthesis column based on deep learning according to claim 6, characterized in that, The features extracted in step 5 include packet header information, packet body content, and methanol synthesis tower features; The packet header information includes source address, destination address, timestamp, and protocol type; The packet body content includes payload data, error check code, and data sequence number; The methanol synthesis tower features include the size of the synthesis gas flow, the reaction temperature, and the reaction pressure; The extracted features are indexed and classified, and then stored in a database. 8.The method of claim 1, wherein, The specific method of step 6 is as follows: Step 6.1, use machine learning algorithms such as support vector machines, isolated forests, and random forests to build an anomaly detection model to classify and detect feature data; Step 6.2, extract features from real-time protocol data and input the extracted features into the anomaly detection model to identify potential abnormal behavior; the feature distribution includes but is not limited to mean, variance, skewness, kurtosis, etc.; based on statistical distribution and historical data, set the threshold range for anomaly detection; Step 6.3, detect potential abnormal behavior, generate alarm information, and record abnormal data and its features; The data visualization step displays the extracted protocol features and analysis results in the form of charts or reports, and real-time displays the methanol synthesis tower process parameter state and abnormal alarm. 9.The method of claim 8, wherein, The abnormal behavior criteria in step 6.3 are as follows: The anomaly detection criteria use a multi-level threshold system, with a first-level alarm defined as a single feature greater than 2σ and less than 3σ, and a model anomaly probability of 0.7-0.9; a second-level alarm is defined as two or more related features simultaneously greater than 2σ or a model anomaly probability greater than or equal to 0.9, and a key parameter exceeding the safety threshold; The anomaly detection criteria include temperature anomaly detection, pressure-flow coupling anomaly detection, and specific conditions triggering corresponding alarm detection rules for catalyst efficiency decline: Detection rule 1: temperature anomaly detection IF (T > +2 AND (dT / dt > 0.5℃ / min) THEN Trigger an overheat alarm; Detection rule 2: pressure-flow coupling anomaly IF (P > + AND (F < – THEN triggers a congestion alarm; Detection rule 3: catalyst efficiency decline IF (conv_rate < 0.8 x baseline) AND (T variability > 1 ) THEN trigger catalyst deactivation alert.

Images