Methanol synthesis reactor abnormal condition monitoring method based on deep learning

Abstract

The application discloses a methanol synthesis device abnormal working condition monitoring method based on deep learning, which comprises the following steps: step 1, collecting real-time process parameters from the industrial control system network of the methanol synthesis device; step 2, pre-processing the real-time process parameters and extracting multi-scale space-time features; step 3, analyzing the multi-scale space-time features by using a deep learning model, predicting future working condition parameter values, and obtaining predicted values; step 4, calculating the residual error between the predicted values and actual values, and combining a dynamic threshold to determine whether to trigger an alarm; and step 5, matching a fault mode knowledge base according to abnormal features, and outputting a fault type and disposal suggestions. The methanol synthesis reactor abnormal working condition monitoring method based on deep learning provided by the application is used to realize early warning capability of the methanol synthesis reactor abnormal working condition, and reduce the risk of non-scheduled shutdown caused by false alarms.

Description

Technical Field

[0001] This invention belongs to the field of chemical process control technology, and relates to a method for monitoring abnormal operating conditions of a methanol synthesis reactor based on deep learning. Background Technology

[0002] Methanol synthesis is one of the core processes in the coal chemical industry. Its reaction process is influenced by the coupling of multiple parameters such as temperature, pressure, gas composition, and space velocity, requiring extremely high control precision. Traditional monitoring of methanol synthesis reactor operation mainly relies on manual experience or alarm mechanisms based on fixed thresholds. These methods have significant limitations in practical applications: First, the strong coupling relationship between parameters (such as temperature and catalyst activity, pressure and space velocity) means that a single threshold alarm cannot accurately reflect the system state. For example, catalyst aging requires a gradual increase in reaction temperature to maintain activity, but a fixed threshold alarm cannot dynamically adapt to this change, easily leading to false alarms or missed alarms. Second, the data stream of the industrial control system contains multi-dimensional temporal characteristics (such as temperature gradients and pressure fluctuation spectra), making it difficult for traditional statistical process control (SPC) methods to effectively model the nonlinear dynamic relationships between parameters, and their analytical capabilities are insufficient for high-dimensional data. Third, although existing technologies attempt to detect anomalies through SPC, they fail to fully utilize the feature extraction capabilities of deep learning for complex data, resulting in a lack of accurate judgment criteria in key scenarios such as catalyst lifecycle management and early warning of sulfur poisoning. Taking copper-based catalysts (such as Katalco 51-9) as an example, they are extremely sensitive to impurities such as sulfides and carbonyl metals. However, traditional methods, due to data delays and insufficient analytical granularity, struggle to capture the accumulation trend of trace impurities in a timely manner, easily leading to irreversible catalyst deactivation due to hysteresis. Furthermore, the dynamic adjustment requirements for aspects such as space velocity optimization and inert gas content control further highlight the rigidity of traditional methods.

[0003] In recent years, with the popularization of Industrial Internet of Things (IIoT) technology, the frequency and accuracy of sensor data acquisition in methanol synthesis units have been significantly improved. However, the real-time analysis and intelligent application of massive amounts of data remain technical challenges. Summary of the Invention

[0004] The purpose of this invention is to provide a deep learning-based method for monitoring abnormal operating conditions of methanol synthesis reactors, thereby enabling early warning of abnormal operating conditions and reducing the risk of unplanned shutdowns caused by false alarms.

[0005] The technical solution adopted in this invention is a method for monitoring abnormal operating conditions of a methanol synthesis unit based on deep learning, and the steps are as follows:

[0006] Step 1: Collect traffic packets in the industrial control system network of the methanol synthesis unit in real time, and use the corresponding protocol deep parsing algorithm to parse the real-time process parameters according to the type of industrial control equipment used in the unit.

[0007] Step 2: Preprocess the real-time process parameters and extract multi-scale spatiotemporal features;

[0008] Step 3: Analyze multi-scale spatiotemporal features using a deep learning model to predict future operating condition parameter values ​​and obtain predicted values;

[0009] Step 4: Calculate the residual between the predicted value and the actual value, and determine whether to trigger an alarm based on the dynamic threshold.

[0010] Step 5: Match the fault mode knowledge base based on the abnormal characteristics, and output the fault type and handling suggestions.

[0011] The invention is further characterized by:

[0012] The process parameters in step 1 include temperature, pressure, gas composition, and impurity concentration, and the data acquisition methods are as follows:

[0013] Temperature was collected using an axial temperature sensor in the reactor, and the temperature value was recorded as follows. T i , i =1,2……, N ;

[0014] The operating pressure value, including the inlet pressure, is acquired through a pressure transmitter. Bed pressure Export pressure ;

[0015] The gas concentrations of H2, CO, CO2, inert gases, and CH3OH, as well as the hydrogen-to-carbon ratio, are output by an online gas chromatograph.

[0016] Real-time detection values ​​for total sulfur, carbonyl compounds, chlorides, and HCN.

[0017] Step 2 includes multi-scale spatiotemporal features, including short-time and long-time features; the time window for short-time features is set to 1 minute, including spatiotemporal distribution features of temperature and dynamic features of pressure.

[0018] The long-term characteristic time window is set to 8 hours, including the catalyst activity decay index (CAI).

[0019] The spatiotemporal distribution characteristics of temperature include the axial temperature gradient. Temperature distribution standard deviation and rate of temperature change The specific extraction methods are as follows:

[0020] Axial temperature gradient That is, the rate of temperature change between adjacent temperature measuring points, calculated as follows:

[0021] (1);

[0022] in, The spacing between adjacent sensors along the axis;

[0023] Temperature distribution standard deviation The calculation method is as follows:

[0024] (2);

[0025] Rate of temperature change The calculation method is as follows:

[0026] (3);

[0027] in, for t The average temperature collected by the axial temperature sensors of all reactors at any given time, Δ t This represents the length of the time window.

[0028] Pressure dynamic characteristics include pressure fluctuation energy E P Pressure-temperature covariance coefficient C TP Pressure change rate and pressure fluctuation spectrum energy Pressure fluctuation energy E P and pressure fluctuation spectrum energy The acquisition process is as follows:

[0029] Step 2.1, Data Preprocessing: For measurement points i of High-frequency noise can be removed by using moving average filtering or low-pass filtering;

[0030] Step 2.2, Wavelet Basis Function Selection: Select the db4 wavelet basis function;

[0031] Step 2.3: Apply discrete wavelet transform to the pressure signal. Perform multi-scale decomposition:

[0032] Based on sampling frequency f s Calculate the required number of layers based on the target frequency band (0.1-1Hz). J :

[0033] (4);

[0034] Each level of decomposition yields approximate coefficients. A j and detail coefficient D j , among which, the jLayer detail factor D j The corresponding frequency band is ;

[0035] Step 2.4, Reconstruct the frequency band signal: Select detail coefficients that include the 0.1-1Hz frequency band. D j1 , D j2 ..., the signal is reconstructed through inverse discrete wavelet transform:

[0036] (5);

[0037] Step 2.5, Calculate the energy at the measurement point: Integrate the square of the reconstructed signal:

[0038] (6);

[0039] Step 2.6, Multi-point energy aggregation: Average energy across all measurement points:

[0040] (7);

[0041] in, M This represents the number of measurement points.

[0042] Step 2.7, Energy calculation in the 0.1-1Hz frequency band: using measurement points i No. j The layer contains detail factor calculations for the 0.1-1Hz frequency band.

[0043] (8);

[0044] in, K This represents the number of sampling data points at the corresponding measurement point;

[0045] Step 2.8: Merge measurement points by frequency band i Energy at this point: Energy in the combined 0.1-1Hz frequency band:

[0046] (9);

[0047] Step 2.9, Multi-point spectral energy aggregation: Average spectral energy of all measurement points:

[0048] (10);

[0049] Pressure-temperature covariance coefficient C TP The extraction method is as follows:

[0050] (11);

[0051] Where Cov() represents covariance. The standard deviation of the pressure distribution;

[0052] Pressure change rate The extraction method is as follows:

[0053] (12);

[0054] in, for t The average pressure of all bed layers at any given time, Δ t This represents the length of the time window.

[0055] The method for extracting the catalyst activity decay index (CAI) is as follows:

[0056] (13);

[0057] in, For the first i Temperature change rate of each temperature sensor dT i / dt The unit is ℃ / min; Weights for temperature sensors; Pressure change rate dP / dt The unit is MPa / min; This is the average value from multiple pressure measurement points, in MPa.

[0058] The deep learning model in step 3 adopts a spatial-temporal dual-channel architecture, which specifically includes a spatial channel, a temporal channel, and a fusion layer. The spatial channel adopts a one-dimensional convolutional neural network (1D-CNN) model; the temporal channel adopts a long short-term memory (LSTM) network model; and the fusion layer adopts an attention mechanism.

[0059] A 1D-CNN model comprises three aspects: 1D-CNN model input, 1D-CNN model output, and 1D-CNN model implementation elements.

[0060] The input to the 1D-CNN model is an axial temperature distribution data sequence. T =[ T 1, T 2,... T N ], data length is N The dimension of the input data is (1, ...). N ), that is, a single-channel one-dimensional signal;

[0061] The output of the 1D-CNN model is the extracted hotspot offset pattern feature vector. ;

[0062] The essential elements of a 1D-CNN model include convolutional layers, activation functions, pooling layers, and multi-layer structures.

[0063] The convolutional layer uses a one-dimensional convolutional kernel to perform pointwise convolution operations on the temperature distribution data to extract local temperature change features; the size of the convolutional kernel is adjusted according to the feature scale of the temperature distribution.

[0064] The activation function is applied after the convolutional layer;

[0065] Pooling layers reduce the dimensionality of features through pooling operations such as max pooling or average pooling;

[0066] The multi-layer structure consists of multiple convolutional layers and pooling layers, which progressively extract temperature distribution features at different scales.

[0067] An LSTM model includes LSTM model input, LSTM model output, and LSTM model implementation elements;

[0068] The LSTM model takes 1-minute sampled pressure time series data as input. P =[ P 1 ,P 2 ,...P M ],in, P i Indicates the first i The pressure values ​​at each time step, with a data length of [number]. M The dimensions of the input data are ( M ,1), that is, univariate time series data;

[0069] The LSTM model outputs the captured pressure-space velocity coupled wave feature vector. ;

[0070] The essential elements of an LSTM model implementation include LSTM structure, bidirectional LSTM, and multi-layer LSTM:

[0071] The LSTM structure captures long-term dependencies in time series data through gating mechanisms such as forget gates, input gates, and output gates.

[0072] Among them, the forgetting gate is ;

[0073] Input gate is ;

[0074] Output gate is ;

[0075] LSTM cell state is ,

[0076] in, ;

[0077] The final hidden state is ;

[0078] Bidirectional LSTM is an LSTM model that uses a bidirectional LSTM structure to capture the sequential dependencies of time series data. The forward LSTM captures time series features from left to right, while the backward LSTM captures time series features from right to left.

[0079] Multi-layer LSTM uses a multi-layer LSTM structure for the temporal channels of the LSTM model to gradually extract more complex temporal features;

[0080] The hydrogen-to-carbon ratio and impurity concentration parameters are modeled using the same LSTM model as the pressure parameter. The LSTM model output feature vectors are as follows: and .

[0081] The fusion layer employs a deep learning model based on an attention mechanism. This deep learning model includes model input and model output.

[0082] The input to an attention-based deep learning model is the feature vector output from the spatial channels. and the feature vectors output by the time-series channels , and ;

[0083] The output of a deep learning model based on an attention mechanism is a fused comprehensive feature vector. F Predict future operating parameters using a fully connected network or regression model.

[0084] The specific method for step 4 is as follows:

[0085] Step 4.1: Calculate the residual vector between the predicted and actual values. R ;

[0086] (14);

[0087] in, The temperature residual is calculated as follows:

[0088] (15);

[0089] The pressure residual is calculated as follows:

[0090] (16);

[0091] The hydrogen-to-carbon ratio residual is calculated as follows:

[0092] (17);

[0093] The impurity residual is calculated as follows:

[0094] (18);

[0095] Step 4.2: Calculate the residual index RI;

[0096] The residual exponent RI is the weighted sum of the residual vectors, and its calculation formula is:

[0097] (19);

[0098] in, , , , These are the weighting coefficients for each parameter;

[0099] Step 4.3: When the RI exceeds the dynamic threshold for 5 consecutive sampling periods, an alarm is triggered.

[0100] The dynamic threshold is set in stages according to the catalyst life cycle:

[0101] In the initial 0-1000 h period, the residual threshold is 2.5σ, where σ is the standard deviation of the training set residuals, and the allowable temperature fluctuation range of the bed hotspot is ±5℃.

[0102] During the medium term of 1000-5000 h, the threshold is 3.0σ, and the allowable fluctuation range of bed hot spot temperature is ±5℃.

[0103] After 5000 hours, the threshold is 3.5σ, and the range of bed hotspot temperature fluctuation is allowed to expand to ±8℃.

[0104] The specific method for step 5 is as follows:

[0105] Incorporate temperature gradient matrix, pressure fluctuation energy, and dynamic hydrogen-to-carbon ratio. H / C The feature parameters are matched with the fault mode knowledge base using cosine similarity. If the similarity exceeds 85%, the fault type and handling suggestions are output.

[0106] The formula for calculating the temperature gradient matrix is ​​as follows:

[0107] (20);

[0108] In the formula, For the first i , j Spacing between sensors

[0109] Pressure fluctuation energy is extracted using the wavelet transform method described in step 2;

[0110] Dynamic hydrogen-carbon ratio H / C By performing sliding window statistics on gas composition data.

[0111] The beneficial effects of this invention are:

[0112] (1) The method for monitoring abnormal operating conditions of methanol synthesis reactor based on deep learning in this invention uses a deep learning model to extract features from the process parameter data of methanol synthesis tower, which can more accurately capture key information in the data, help identify potential safety threats and abnormal behaviors, thereby improving the overall safety and stability of methanol synthesis unit system.

[0113] (2) The method for monitoring abnormal operating conditions of methanol synthesis reactor based on deep learning in this invention significantly improves the accuracy of complex fault detection by fusing multi-scale spatiotemporal features (such as axial temperature gradient and pressure fluctuation energy) with deep learning models; and significantly reduces the false alarm rate by combining dynamic threshold mechanism with catalyst lifetime stage adjustment.

[0114] (3) The methanol synthesis reactor abnormal operation monitoring method based on deep learning of the present invention has the ability to adapt throughout the entire life cycle. It quantifies the activity status of the catalyst activity decay index (CAI) in real time and, together with the staged threshold strategy, avoids false alarms caused by catalyst aging and extends the service life by 10-15%. It allows a wider temperature fluctuation range (±8℃) in the later stage to adapt to the activity decline condition. It also has the ability to respond quickly and trace the fault source accurately. The residual index (RI) triggers alarm with a small delay and the fault knowledge base is matched quickly, which can effectively improve the efficiency of pushing disposal suggestions.

[0115] (4) The method for monitoring abnormal operating conditions of methanol synthesis reactor based on deep learning in this invention can effectively reduce unplanned shutdowns, thereby increasing production. It can also reduce the frequency of catalyst replacement through predictive maintenance, save maintenance costs, and generate significant economic benefits. Attached Figure Description

[0116] Figure 1 This is a flowchart illustrating the abnormal operating condition monitoring method for a methanol synthesis unit based on deep learning, as described in this invention. Detailed Implementation

[0117] The present invention will now be described in detail with reference to the accompanying drawings and specific embodiments.

[0118] This invention relates to a deep learning-based method for monitoring abnormal operating conditions of a methanol synthesis unit, such as... Figure 1 As shown, the specific steps are as follows:

[0119] Step 1: Collect traffic packets in the industrial control system network of the methanol synthesis unit in real time, and use the corresponding protocol deep parsing algorithm to parse the real-time process parameters according to the type of industrial control equipment used in the unit.

[0120] The process parameters include temperature, pressure, gas composition, and impurity concentration, and the data acquisition methods are as follows:

[0121] Temperature was collected using an axial temperature sensor in the reactor; the temperature value was... ;

[0122] The operating pressure value, including the inlet pressure, is acquired through a pressure transmitter. Bed pressure Export pressure ;

[0123] The concentrations of H2, CO, CO2, inert gases, and CH3OH, as well as the hydrogen-to-carbon ratio, are obtained from the output of an online gas chromatograph.

[0124] Impurity concentration: Real-time detection values ​​of total sulfur, carbonyl compounds, chlorides, and HCN.

[0125] Step 2: Preprocess the real-time process parameters and extract multi-scale spatiotemporal features; among which, multi-scale spatiotemporal features include short-time features and long-time features;

[0126] The short-term feature time window is set to 1 minute, including the spatiotemporal distribution features of temperature and the dynamic features of pressure.

[0127] Furthermore, the spatiotemporal distribution characteristics of temperature include the axial temperature gradient. Temperature distribution standard deviation and rate of temperature change The specific extraction methods are as follows:

[0128] Axial temperature gradient This refers to the rate of temperature change between adjacent temperature measuring points, reflecting the movement of hot spots in the catalyst bed. The calculation method is as follows:

[0129] (1);

[0130] in, The spacing between adjacent sensors along the axis;

[0131] Temperature distribution standard deviation The uniformity of axial temperature distribution is measured and calculated as follows:

[0132] (2);

[0133] Rate of temperature change The calculation method is as follows:

[0134] (3);

[0135] in, for t The average temperature collected by the axial temperature sensors of all reactors at any given time, Δ t This represents the length of the time window.

[0136] Pressure dynamic characteristics include pressure fluctuation energy E P Pressure-temperature covariance coefficient C TP Pressure change rate and pressure fluctuation spectrum energy Pressure fluctuation energy E P and pressure fluctuation spectrum energy The acquisition process is as follows:

[0137] Step 2.1, Data Preprocessing: For measurement points i of High-frequency noise can be removed by using moving average filtering or low-pass filtering;

[0138] Step 2.2, Wavelet basis function selection: Select the db4 wavelet basis function;

[0139] Step 2.3: Apply Discrete Wavelet Transform (DWT) to the pressure signal. Perform multi-scale decomposition:

[0140] Based on sampling frequency f s Calculate the required number of layers based on the target frequency band (0.1-1Hz). J :

[0141] (4);

[0142] Each level of decomposition yields approximate coefficients. A j and detail coefficient D j , among which, the j Layer detail factor D j The corresponding frequency band is ;

[0143] Step 2.4, Reconstruct the frequency band signal: Select detail coefficients that include the 0.1-1Hz frequency band. D j1 , D j2 ..., the signal is reconstructed using inverse discrete wavelet transform (IDWT):

[0144] (5);

[0145] Step 2.5, Calculate the energy at the measurement point: Integrate the square of the reconstructed signal:

[0146] (6);

[0147] Step 2.6, Multi-point energy aggregation: Average energy across all measurement points:

[0148] (7);

[0149] in, M This represents the number of measurement points.

[0150] Step 2.7, Energy calculation in the 0.1-1Hz frequency band: using measurement points i No. j The layer contains detail factor calculations for the 0.1-1Hz frequency band.

[0151] (8);

[0152] in, K This represents the number of sampling data points at the corresponding measurement point;

[0153] Step 2.8: Merge measurement points by frequency band i Energy at this point: Energy in the combined 0.1-1Hz frequency band:

[0154] (9);

[0155] Step 2.9, Multi-point spectral energy aggregation: Average spectral energy of all measurement points:

[0156] (10);

[0157] Pressure-temperature covariance coefficient C TP The extraction method is as follows:

[0158] (11);

[0159] Where Cov() represents covariance. The standard deviation of the pressure distribution;

[0160] Pressure change rate The extraction method is as follows:

[0161] (12);

[0162] in, for tThe average pressure of all bed layers at any given time, Δ t This represents the length of the time window.

[0163] The long-term characteristic time window is set to 8 hours, including the catalyst activity decay index (CAI).

[0164] The method for extracting the catalyst activity decay index (CAI) is as follows:

[0165] (13);

[0166] in, For the first i Temperature change rate of each temperature sensor dT i / dt The unit is ℃ / min; Weights for temperature sensors; Pressure change rate dP / dt The unit is MPa / min; This is the average value from multiple pressure measurement points, in MPa.

[0167] Step 3: Analyze multi-scale spatiotemporal features using a deep learning model to predict future operating condition parameter values ​​and obtain predicted values. The deep learning model adopts a spatial-temporal dual-channel architecture, which specifically includes a spatial channel, a temporal channel, and a fusion layer. The spatial channel uses a one-dimensional convolutional neural network (1D-CNN) model; the temporal channel uses a long short-term memory (LSTM) network model; and the fusion layer uses an attention mechanism.

[0168] Spatial channel: A one-dimensional convolutional neural network (CNN) processes axial temperature distribution and extracts hotspot offset patterns;

[0169] The time-series channel uses a Long Short-Term Memory (LSTM) network to process pressure time-series data (1-minute sampled data from pressure transmitters) and capture pressure-space velocity coupled fluctuations.

[0170] The fusion layer dynamically allocates weights for spatial and temporal features through an attention mechanism.

[0171] Furthermore, the 1D-CNN model includes three aspects: 1D-CNN model input, 1D-CNN model output, and 1D-CNN model implementation elements.

[0172] The input to the 1D-CNN model is an axial temperature distribution data sequence. T =[ T 1, T 2,... T N ], data length is N The dimension of the input data is (1, ...N ), that is, a single-channel one-dimensional signal;

[0173] The output of the 1D-CNN model is the extracted hotspot offset pattern feature vector. ;

[0174] The essential elements of a 1D-CNN model include convolutional layers, activation functions, pooling layers, and multi-layer structures.

[0175] The convolutional layer uses a one-dimensional convolutional kernel to perform pointwise convolution operations on the temperature distribution data to extract local temperature change features; the size of the convolutional kernel can be adjusted according to the feature scale of the temperature distribution.

[0176] The activation function is applied after the convolutional layer to extract non-linear features;

[0177] Pooling layers reduce the dimensionality of features through pooling operations such as max pooling or average pooling, thereby reducing computational complexity and extracting more abstract features;

[0178] The multi-layer structure consists of multiple convolutional layers and pooling layers, which progressively extract temperature distribution features at different scales.

[0179] An LSTM model includes LSTM model input, LSTM model output, and LSTM model implementation elements;

[0180] The LSTM model takes 1-minute sampled pressure time series data as input. P =[ P 1 ,P 2 ,...P M ],in, P i Indicates the first i The pressure values ​​at each time step, with a data length of [number]. M The dimensions of the input data are ( M ,1), that is, univariate time series data;

[0181] The LSTM model outputs the captured pressure-space velocity coupled wave feature vector. ;

[0182] The essential elements of an LSTM model implementation include LSTM structure, bidirectional LSTM, and multi-layer LSTM:

[0183] The LSTM structure captures long-term dependencies in time series data through gating mechanisms such as forget gates, input gates, and output gates.

[0184] Among them, the forgetting gate is This is used to determine which old information to forget;

[0185] Input gate is This is used to determine what new information to input;

[0186] Output gate is This is used to determine which information to output;

[0187] LSTM cell state is ,

[0188] in, ;

[0189] The final hidden state is ;

[0190] Bidirectional LSTM is an LSTM model that uses a bidirectional LSTM structure to capture the sequential dependencies of time series data. The forward LSTM captures time series features from left to right, while the backward LSTM captures time series features from right to left.

[0191] Multi-layer LSTM uses a multi-layer LSTM structure for the temporal channels of the LSTM model to gradually extract more complex temporal features;

[0192] The hydrogen-to-carbon ratio and impurity concentration parameters are modeled using the same LSTM model as the pressure parameter. The LSTM model output feature vectors are as follows: and .

[0193] The fusion layer employs a deep learning model based on an attention mechanism. This deep learning model includes model input and model output.

[0194] The input to an attention-based deep learning model is the feature vector output from the spatial channels. and the feature vectors output by the time-series channels , and ;

[0195] The output of a deep learning model based on an attention mechanism is a fused comprehensive feature vector. F Predict future operating parameters using a fully connected network or regression model.

[0196] Step 4: Calculate the residual between the predicted value and the actual value, and combine it with the dynamic threshold to determine whether to trigger an alarm. The specific method is as follows:

[0197] Step 4.1: Calculate the residual vector between the predicted and actual values. R ;

[0198] (14);

[0199] in, The temperature residual is calculated as follows:

[0200] (15);

[0201] The pressure residual is calculated as follows:

[0202] (16);

[0203] The hydrogen-to-carbon ratio residual is calculated as follows:

[0204] (17);

[0205] The impurity residual is calculated as follows:

[0206] (18);

[0207] Step 4.2: Calculate the residual index RI;

[0208] The residual exponent RI is the weighted sum of the residual vectors, and its calculation formula is:

[0209] (19);

[0210] in, , , , Weight coefficients for each parameter are assigned based on their importance, for example... , , , ;

[0211] Step 4.3: When the RI exceeds the dynamic threshold for 5 consecutive sampling periods, an alarm is triggered.

[0212] The dynamic threshold is set in stages according to the catalyst life cycle:

[0213] In the initial 0-1000 h period, the residual threshold is 2.5σ, where σ is the standard deviation of the training set residuals, and the allowable temperature fluctuation range of the bed hotspot is ±5℃.

[0214] During the medium term of 1000-5000 h, the threshold is 3.0σ, and the allowable fluctuation range of bed hot spot temperature is ±5℃.

[0215] After 5000 hours, the threshold is 3.5σ, and the range of bed hotspot temperature fluctuation is allowed to expand to ±8℃.

[0216] Step 5: Match the fault mode knowledge base based on the abnormal characteristics, and output the fault type and handling suggestions. The specific method is as follows:

[0217] Incorporate temperature gradient matrix, pressure fluctuation energy, and dynamic hydrogen-to-carbon ratio. H / C The feature parameters are matched with the fault mode knowledge base using cosine similarity. If the similarity exceeds 85%, the fault type and handling suggestions are output.

[0218] The formula for calculating the temperature gradient matrix is ​​as follows:

[0219] (20);

[0220] In the formula, For the first i , j Spacing between sensors

[0221] Pressure fluctuation energy is extracted using the wavelet transform method described in step 2;

[0222] Dynamic hydrogen-carbon ratio H / C By performing sliding window statistics on gas composition data.

[0223] The fault mode knowledge base should be configured for regular updates.

[0224] The Fault Mode Knowledge Base includes:

[0225] Characteristic waveform of sulfur poisoning: Total sulfur concentration suddenly increases to >60 ppb, accompanied by a temperature drop of >3℃ / min in the bed outlet area;

[0226] Characteristic waveform of carbonyl contamination (C-005): The concentration of carbonyl compounds reaches >6 ppb, and the energy of inlet pressure fluctuation increases by >20%;

[0227] Catalyst deactivation characteristic waveform (C-010): CAI < -1.5, and the annual average increase in bed pressure > 0.4 MPa.

[0228] This invention presents an intelligent monitoring method integrating deep learning and industrial control message parsing. Through dynamic feature extraction, residual threshold optimization, and fault mode matching, it overcomes the bottleneck of traditional technologies in processing multidimensional nonlinear data. This method can not only analyze complex features such as temperature gradients and pressure fluctuation spectra in real time, but also adaptively adjust alarm logic according to the catalyst life cycle, significantly improving the early warning capability for faults such as sulfur poisoning (e.g., warning time more than 30 minutes in advance). Simultaneously, it reduces the risk of unplanned downtime due to false alarms, demonstrating significant industrial application value.

[0229] Example 1

[0230] 1. A deep learning-based method for monitoring abnormal operating conditions of a methanol synthesis unit, comprising the following steps:

[0231] Step 1: Collect traffic packets in the industrial control system network of the methanol synthesis unit in real time, and use the corresponding protocol deep parsing algorithm to parse the real-time process parameters according to the type of industrial control equipment used in the unit.

[0232] The process parameters include temperature, pressure, gas composition, and impurity concentration, and the data acquisition methods are as follows:

[0233] Temperature was collected using an axial temperature sensor in the reactor; the temperature value was... ;

[0234] The operating pressure value, including the inlet pressure, is acquired through a pressure transmitter. Bed pressure Export pressure ;

[0235] The concentrations of H2, CO, CO2, inert gases, and CH3OH, as well as the hydrogen-to-carbon ratio, are obtained from the output of an online gas chromatograph.

[0236] Impurity concentration: Real-time detection values ​​of total sulfur, carbonyl compounds, chlorides, and HCN.

[0237] Step 2: Preprocess the real-time process parameters and extract multi-scale spatiotemporal features;

[0238] Step 3: Analyze multi-scale spatiotemporal features using a deep learning model to predict future operating condition parameter values ​​and obtain predicted values;

[0239] Step 4: Calculate the residual between the predicted value and the actual value, and determine whether to trigger an alarm based on the dynamic threshold.

[0240] Step 5: Match the fault mode knowledge base based on the abnormal characteristics, and output the fault type and handling suggestions.

[0241] Example 2

[0242] The method in this embodiment is the same as in embodiment 1. However, the multi-scale spatiotemporal features in step 2 of embodiment 1 include short-term features and long-term features. The time window for short-term features is set to 1 minute, including the spatiotemporal distribution features of temperature and the dynamic features of pressure. The time window for long-term features is set to 8 hours, including the catalyst activity decay index (CAI).

[0243] Example 3

[0244] The method in this embodiment is the same as in embodiment 2, except that the spatiotemporal temperature distribution characteristics in embodiment 2 include the axial temperature gradient. Temperature distribution standard deviation and rate of temperature change The specific extraction methods are as follows:

[0245] Axial temperature gradient This refers to the rate of temperature change between adjacent temperature measuring points, reflecting the movement of hot spots in the catalyst bed. The calculation method is as follows:

[0246] (1);

[0247] in, The spacing between adjacent sensors along the axis;

[0248] Temperature distribution standard deviation The uniformity of axial temperature distribution is measured and calculated as follows:

[0249] (2);

[0250] Rate of temperature change The calculation method is as follows:

[0251] (3);

[0252] in, for t The average temperature collected by the axial temperature sensors of all reactors at any given time, Δ t This represents the length of the time window.

[0253] Example 4

[0254] The method in this embodiment is the same as in embodiment 2, except that the pressure dynamic characteristics in embodiment 2 include pressure fluctuation energy. E P Pressure-temperature covariance coefficient C TP Pressure change rate and pressure fluctuation spectrum energy Pressure fluctuation energy E P and pressure fluctuation spectrum energy The acquisition process is as follows:

[0255] Step 2.1, Data Preprocessing: For measurement points i of High-frequency noise can be removed by using moving average filtering or low-pass filtering;

[0256] Step 2.2, Wavelet basis function selection: Select the db4 wavelet basis function;

[0257] Step 2.3: Apply Discrete Wavelet Transform (DWT) to the pressure signal. Perform multi-scale decomposition:

[0258] Based on sampling frequency fs Calculate the required number of layers based on the target frequency band (0.1-1Hz). J :

[0259] (4);

[0260] Each level of decomposition yields approximate coefficients. A j and detail coefficient D j , among which, the j Layer detail factor D j The corresponding frequency band is ;

[0261] Step 2.4, Reconstruct the frequency band signal: Select detail coefficients that include the 0.1-1Hz frequency band. D j1 , D j2 ..., the signal is reconstructed using inverse discrete wavelet transform (IDWT):

[0262] (5);

[0263] Step 2.5, Calculate the energy at the measurement point: Integrate the square of the reconstructed signal:

[0264] (6);

[0265] Step 2.6, Multi-point energy aggregation: Average energy across all measurement points:

[0266] (7);

[0267] in, M This represents the number of measurement points.

[0268] Step 2.7, Energy calculation in the 0.1-1Hz frequency band: using measurement points i No. j The layer contains detail factor calculations for the 0.1-1Hz frequency band.

[0269] (8);

[0270] in, K This represents the number of sampling data points at the corresponding measurement point;

[0271] Step 2.8: Merge measurement points by frequency band i Energy at this point: Energy in the combined 0.1-1Hz frequency band:

[0272] (9);

[0273] Step 2.9, Multi-point spectral energy aggregation: Average spectral energy of all measurement points:

[0274] (10).

[0275] Example 5

[0276] The method in this embodiment is the same as in Embodiment 2, except that the pressure-temperature covariance coefficient C in Embodiment 2 is... TP The extraction method is as follows:

[0277] (11);

[0278] Where Cov() represents covariance. The standard deviation of the pressure distribution;

[0279] Pressure change rate The extraction method is as follows:

[0280] (12);

[0281] in, for t The average pressure of all bed layers at any given time, Δ t This represents the length of the time window.

[0282] Example 6

[0283] The method in this embodiment is the same as in Embodiment 2, except that the method for extracting the catalyst activity decay index (CAI) in Embodiment 2 is as follows:

[0284] (13);

[0285] in, For the first i Temperature change rate of each temperature sensor dT i / dt The unit is ℃ / min; Weights for temperature sensors; Pressure change rate dP / dt The unit is MPa / min; This is the average value from multiple pressure measurement points, in MPa.

[0286] Example 7

[0287] The method in this embodiment is the same as in embodiment 1. However, in step 3 of embodiment 1, the deep learning model adopts a spatial-temporal dual-channel architecture, which specifically includes a spatial channel, a temporal channel, and a fusion layer. The spatial channel adopts a one-dimensional convolutional neural network (1D-CNN) model. The temporal channel adopts a long short-term memory (LSTM) network model. The fusion layer adopts an attention mechanism.

[0288] A 1D-CNN model comprises three aspects: 1D-CNN model input, 1D-CNN model output, and 1D-CNN model implementation elements.

[0289] The input to the 1D-CNN model is an axial temperature distribution data sequence. T =[ T 1, T 2,... T N ], data length is N The dimension of the input data is (1, ... N ), that is, a single-channel one-dimensional signal;

[0290] The output of the 1D-CNN model is the extracted hotspot offset pattern feature vector. ;

[0291] The essential elements of a 1D-CNN model include convolutional layers, activation functions, pooling layers, and multi-layer structures.

[0292] The convolutional layer uses a one-dimensional convolutional kernel to perform pointwise convolution operations on the temperature distribution data to extract local temperature change features; the size of the convolutional kernel can be adjusted according to the feature scale of the temperature distribution.

[0293] The activation function is applied after the convolutional layer to extract non-linear features;

[0294] Pooling layers reduce the dimensionality of features through pooling operations such as max pooling or average pooling, thereby reducing computational complexity and extracting more abstract features;

[0295] The multi-layer structure consists of multiple convolutional layers and pooling layers, which progressively extract temperature distribution features at different scales.

[0296] An LSTM model includes LSTM model input, LSTM model output, and LSTM model implementation elements;

[0297] The LSTM model takes 1-minute sampled pressure time series data as input. P =[ P 1 ,P 2 ,...P M ],in, P i Indicates the first i The pressure values ​​at each time step, with a data length of [number]. M The dimensions of the input data are ( M ,1), that is, univariate time series data;

[0298] The LSTM model outputs the captured pressure-space velocity coupled wave feature vector. ;

[0299] The essential elements of an LSTM model implementation include LSTM structure, bidirectional LSTM, and multi-layer LSTM:

[0300] The LSTM structure captures long-term dependencies in time series data through gating mechanisms such as forget gates, input gates, and output gates.

[0301] Among them, the forgetting gate is This is used to determine which old information to forget;

[0302] Input gate is This is used to determine what new information to input;

[0303] Output gate is This is used to determine which information to output;

[0304] LSTM cell state is ,

[0305] in, ;

[0306] The final hidden state is ;

[0307] Bidirectional LSTM is an LSTM model that uses a bidirectional LSTM structure to capture the sequential dependencies of time series data. The forward LSTM captures time series features from left to right, while the backward LSTM captures time series features from right to left.

[0308] Multi-layer LSTM uses a multi-layer LSTM structure for the temporal channels of the LSTM model to gradually extract more complex temporal features;

[0309] The hydrogen-to-carbon ratio and impurity concentration parameters are modeled using the same LSTM model as the pressure parameter. The LSTM model output feature vectors are as follows: and .

[0310] The fusion layer employs a deep learning model based on an attention mechanism. This deep learning model includes model input and model output.

[0311] The input to an attention-based deep learning model is the feature vector output from the spatial channels. and the feature vectors output by the time-series channels , and ;

[0312] The output of a deep learning model based on an attention mechanism is a fused comprehensive feature vector. F Predict future operating parameters using a fully connected network or regression model.

[0313] Example 8

[0314] The method in this embodiment is the same as in embodiment 1, except that the specific method of step 4 in embodiment 1 is as follows:

[0315] Step 4.1: Calculate the residual vector between the predicted and actual values. R ;

[0316] (14);

[0317] in, The temperature residual is calculated as follows:

[0318] (15);

[0319] The pressure residual is calculated as follows:

[0320] (16);

[0321] The hydrogen-to-carbon ratio residual is calculated as follows:

[0322] (17);

[0323] The impurity residual is calculated as follows:

[0324] (18);

[0325] Step 4.2: Calculate the residual index RI;

[0326] The residual exponent RI is the weighted sum of the residual vectors, and its calculation formula is:

[0327] (19);

[0328] in, , , , Weight coefficients for each parameter are assigned based on their importance, for example... , , , ;

[0329] Step 4.3: When the RI exceeds the dynamic threshold for 5 consecutive sampling periods, an alarm is triggered.

[0330] The dynamic threshold is set in stages according to the catalyst life cycle:

[0331] In the initial 0-1000 h period, the residual threshold is 2.5σ, where σ is the standard deviation of the training set residuals, and the allowable temperature fluctuation range of the bed hotspot is ±5℃.

[0332] During the medium term of 1000-5000 h, the threshold is 3.0σ, and the allowable fluctuation range of bed hot spot temperature is ±5℃.

[0333] After 5000 hours, the threshold is 3.5σ, and the range of bed hotspot temperature fluctuation is allowed to expand to ±8℃.

[0334] Example 9

[0335] The method in this embodiment is the same as in embodiment 1, except that the specific method of step 5 in embodiment 1 is as follows:

[0336] Incorporate temperature gradient matrix, pressure fluctuation energy, and dynamic hydrogen-to-carbon ratio. H / C The feature parameters are matched with the fault mode knowledge base using cosine similarity. If the similarity exceeds 85%, the fault type and handling suggestions are output.

[0337] The formula for calculating the temperature gradient matrix is ​​as follows:

[0338] (20);

[0339] In the formula, For the first i , j Spacing between sensors

[0340] Pressure fluctuation energy is extracted using the wavelet transform method described in step 2;

[0341] Dynamic hydrogen-carbon ratio H / C By performing sliding window statistics on gas composition data.

[0342] The fault mode knowledge base should be configured for regular updates.

[0343] The Fault Mode Knowledge Base includes:

[0344] Characteristic waveform of sulfur poisoning: Total sulfur concentration suddenly increases to >60 ppb, accompanied by a temperature drop of >3℃ / min in the bed outlet area;

[0345] Characteristic waveform of carbonyl contamination (C-005): The concentration of carbonyl compounds reaches >6 ppb, and the energy of inlet pressure fluctuation increases by >20%;

[0346] Catalyst deactivation characteristic waveform (C-010): CAI < -1.5, and the annual average increase in bed pressure > 0.4 MPa.

Claims

1. A method for monitoring abnormal operating conditions of a methanol synthesis unit based on deep learning, characterized in that, The specific steps are as follows: Step 1: Collect traffic packets in the industrial control system network of the methanol synthesis unit in real time, and use the corresponding protocol deep parsing algorithm to parse the real-time process parameters according to the type of industrial control equipment used in the unit. Step 2: Preprocess the real-time process parameters to extract multi-scale spatiotemporal features; Step 3: Analyze multi-scale spatiotemporal features using a deep learning model to predict future operating condition parameter values ​​and obtain predicted values; Step 4: Calculate the residual between the predicted value and the actual value, and determine whether to trigger an alarm based on the dynamic threshold. Step 5: Match the fault mode knowledge base based on the abnormal characteristics, and output the fault type and handling suggestions; The deep learning model in step 3 adopts a spatial-temporal dual-channel architecture, which specifically includes a spatial channel, a temporal channel, and a fusion layer; the spatial channel adopts a one-dimensional convolutional neural network (1D-CNN) model; the temporal channel adopts a long short-term memory (LSTM) network model; and the fusion layer adopts an attention mechanism. The 1D-CNN model includes three aspects: 1D-CNN model input, 1D-CNN model output, and 1D-CNN model implementation elements. The input to the 1D-CNN model is an axial temperature distribution data sequence. T =[ T 1, T 2,... T N ], data length is N The dimension of the input data is (1, ... N ), that is, a single-channel one-dimensional signal; The output of the 1D-CNN model is the extracted hotspot offset pattern feature vector. ; The elements of the 1D-CNN model include convolutional layers, activation functions, pooling layers, and multi-layer structures. The convolutional layer uses a one-dimensional convolutional kernel to perform point-by-point convolution on the temperature distribution data to extract local temperature change features; the size of the convolutional kernel is adjusted according to the feature scale of the temperature distribution. The activation function is an activation function appended after the convolutional layer; The pooling layer reduces the dimensionality of features through pooling operations such as max pooling or average pooling; The multi-layer structure is composed of multiple convolutional layers and pooling layers, which gradually extract temperature distribution features at different scales. The LSTM model includes LSTM model input, LSTM model output, and LSTM model implementation elements; The LSTM model is input with 1-minute sampled pressure time series data. P =[ P 1 ,P 2 ,...P M ],in, P i Indicates the first i The pressure values ​​at each time step, with a data length of [number]. M The dimensions of the input data are ( M ,1), that is, univariate time series data; The LSTM model outputs the captured pressure-space velocity coupled wave feature vector. ; The elements of the LSTM model implementation include LSTM structure, bidirectional LSTM, and multi-layer LSTM: The LSTM structure captures long-term dependencies in time series through gating mechanisms such as forget gates, input gates, and output gates. Among them, the forgetting gate is ; Input gate is ; Output gate is ; LSTM cell state is , in, ; The final hidden state is ; The bidirectional LSTM is an LSTM model that uses a bidirectional LSTM structure to capture the sequential dependencies of time series data. The forward LSTM captures time series features from left to right, and the backward LSTM captures time series features from right to left. The multi-layer LSTM is an LSTM model whose temporal channels use a multi-layer LSTM structure to gradually extract more complex temporal features. The hydrogen-to-carbon ratio and impurity concentration parameters are modeled using the same LSTM model as the pressure parameter. The LSTM model outputs feature vectors as follows: and ; The fusion layer employs a deep learning model based on an attention mechanism, which includes model input and model output. The input to the attention-based deep learning model is the feature vector output from the spatial channels. and the feature vectors output by the time-series channels , and ; The output of the attention-based deep learning model is a fused comprehensive feature vector. F Predict future operating parameters using a fully connected network or regression model.

2. The method for monitoring abnormal operating conditions of a methanol synthesis unit based on deep learning according to claim 1, characterized in that, The process parameters in step 1 include temperature, pressure, gas composition, and impurity concentration, and the data acquisition methods are as follows: Temperature was collected using an axial temperature sensor in the reactor, and the temperature value was recorded as follows. T i , i =1,2……, N ; The operating pressure value, including the inlet pressure, is acquired through a pressure transmitter. Bed pressure Export pressure ; The gas concentrations of H2, CO, CO2, inert gases, and CH3OH, as well as the hydrogen-to-carbon ratio, are output by an online gas chromatograph. Real-time detection values ​​for total sulfur, carbonyl compounds, chlorides, and HCN.

3. The method for monitoring abnormal operating conditions of a methanol synthesis unit based on deep learning according to claim 1, characterized in that, The multi-scale spatiotemporal features in step 2 include short-time features and long-time features; the time window for the short-time features is set to 1 minute, including spatiotemporal distribution features of temperature and dynamic features of pressure; The long-term characteristic time window is set to 8 hours, including the catalyst activity decay index (CAI).

4. The method for monitoring abnormal operating conditions of a methanol synthesis unit based on deep learning according to claim 3, characterized in that, The spatiotemporal distribution characteristics of temperature include the axial temperature gradient. Temperature distribution standard deviation and rate of temperature change The specific extraction methods are as follows: The axial temperature gradient This refers to the rate of temperature change between adjacent temperature measuring points, calculated as follows: （1）； in, The spacing between adjacent sensors along the axis; The standard deviation of the temperature distribution The calculation method is as follows: （2）； The rate of temperature change The calculation method is as follows: （3）； in, for t The average temperature collected by the axial temperature sensors of all reactors at any given time, Δ t This represents the length of the time window.

5. The method for monitoring abnormal operating conditions of a methanol synthesis unit based on deep learning according to claim 3, characterized in that, The pressure dynamic characteristics include pressure fluctuation energy. E P Pressure-temperature covariance coefficient C TP Pressure change rate and pressure fluctuation spectrum energy The pressure fluctuation energy E P and pressure fluctuation spectrum energy The acquisition process is as follows: Step 2.1, Data Preprocessing: For measurement points i of High-frequency noise can be removed by using moving average filtering or low-pass filtering; Step 2.2, Wavelet basis function selection: Select the db4 wavelet basis function; Step 2.3: Apply discrete wavelet transform to the pressure signal. Perform multi-scale decomposition: Based on sampling frequency f s Calculate the required number of layers based on the target frequency band (0.1-1Hz). J : （4）； Each level of decomposition yields approximate coefficients. A j and detail coefficient D j , among which, the j Layer detail factor D j The corresponding frequency band is ; Step 2.4, Reconstruct the frequency band signal: Select detail coefficients that include the 0.1-1Hz frequency band. D j1 , D j2 ..., the signal is reconstructed through inverse discrete wavelet transform: （5）； Step 2.5, Calculate the energy at the measurement point: Integrate the square of the reconstructed signal: （6）； Step 2.6, Multi-point energy aggregation: Average energy across all measurement points: （7）； in, M This represents the number of measurement points. Step 2.7, Energy calculation in the 0.1-1Hz frequency band: using measurement points i No. j The layer contains detail factor calculations for the 0.1-1Hz frequency band. （8）； in, K This represents the number of sampling data points at the corresponding measurement point; Step 2.8: Merge measurement points by frequency band i Energy at this point: Energy in the combined 0.1-1Hz frequency band: （9）； Step 2.9, Multi-point spectral energy aggregation: Average spectral energy of all measurement points: （10）； The pressure-temperature covariance C TP The extraction method is as follows: （11）； Where Cov() represents covariance. The standard deviation of the pressure distribution; The pressure change rate The extraction method is as follows: （12）； in, for t The average pressure of all bed layers at any given time, Δ t This represents the length of the time window.

6. The method for monitoring abnormal operating conditions of a methanol synthesis unit based on deep learning according to claim 3, characterized in that, The method for extracting the catalyst activity decay index (CAI) is as follows: （13）； in, For the first i Temperature change rate of each temperature sensor dT i / dt The unit is ℃ / min; Weights for temperature sensors; Pressure change rate dP / dt The unit is MPa / min; This is the average value from multiple pressure measurement points, in MPa.

7. The method for monitoring abnormal operating conditions of a methanol synthesis unit based on deep learning according to claim 1, characterized in that, The specific method for step 4 is as follows: Step 4.1: Calculate the residual vector between the predicted and actual values. R ; （14）； in, The temperature residual is calculated as follows: （15）； The pressure residual is calculated as follows: （16）； The hydrogen-to-carbon ratio residual is calculated as follows: （17）； The impurity residual is calculated as follows: （18）； Step 4.2: Calculate the residual index RI; The residual exponent RI is a weighted sum of the residual vectors, and its calculation formula is as follows: （19）； in, , , , These are the weighting coefficients for each parameter; Step 4.3: When the RI exceeds the dynamic threshold for 5 consecutive sampling periods, an alarm is triggered. The dynamic threshold is set in stages according to the catalyst life cycle: In the initial 0-1000 h period, the residual threshold is 2.5σ, where σ is the standard deviation of the training set residuals, and the allowable temperature fluctuation range of the bed hotspot is ±5℃. During the medium term of 1000-5000 h, the threshold is 3.0σ, and the allowable fluctuation range of bed hot spot temperature is ±5℃. After 5000 hours, the threshold is 3.5σ, and the range of bed hotspot temperature fluctuation is allowed to expand to ±8℃.

8. The method for monitoring abnormal operating conditions of a methanol synthesis unit based on deep learning according to claim 1, characterized in that, The specific method for step 5 is as follows: Incorporate temperature gradient matrix, pressure fluctuation energy, and dynamic hydrogen-to-carbon ratio. H / C The feature parameters are matched with the fault mode knowledge base using cosine similarity. If the similarity exceeds 85%, the fault type and handling suggestions are output. The formula for calculating the temperature gradient matrix is ​​as follows: （20）； In the formula, For the first i , j Spacing between sensors Pressure fluctuation energy is extracted using the wavelet transform method described in step 2; Dynamic hydrogen-carbon ratio H / C By performing sliding window statistics on gas composition data.

Images