Method suitable for monitoring scaling of reflux condenser

Abstract

The invention provides a method suitable for monitoring scaling of a reflux condenser, and the method comprises the steps: monitoring the scaling condition of the reflux condenser in a chemical industry process through an AI deep learning method and a signal processing analysis method, and carrying out the monitoring of the scaling condition of the reflux condenser through the historical data collected in the process. The scaling condition of the reflux condenser is represented by combining a long short-term memory network of an AI deep learning method. The method comprises the following steps: S1, acquiring sensor data of an industrial process; s2, performing data cleaning on the process data to obtain feature data; s3, training a network by using historical data to obtain a data-driven scaling model; s4, process data of a new process are input, and the current scale deposition condition is predicted; according to the method, historical data collected in the manufacturing process of the reflux condenser is utilized, and the scaling size is represented in combination with the long-short-term memory network, so that the scaling size of the reflux condenser can be predicted or monitored and represented.

Description

Technical Field

[0001] The present invention relates to the technical field of equipment fault detection, and in particular to a method suitable for monitoring the fouling of a reflux condenser. Background Art

[0002] Since the internal structure of the reflux condenser cannot be directly observed, the existing technology usually uses a threshold judgment method to determine whether the condenser / heat exchanger equipment needs to be cleaned, and the fouling situation of the equipment cannot be monitored in real time.

[0003] For the physical sensors for fouling monitoring in the prior art, specific sensors are used for monitoring to characterize the size of fouling, which is not easy to match with the actual on-site equipment. There are also solutions starting from theoretical models in the prior art, but accurate measurement of the structural parameters is required, and the complexity is relatively high for specific device equipment. Summary of the Invention

[0004] The present invention provides a method suitable for monitoring the fouling of a reflux condenser. By using historical data collected during the manufacturing process of the reflux condenser and combining with a long short-term memory network, the size of fouling can be characterized, and the size of fouling of the reflux condenser can be predicted or monitored.

[0005] The present invention adopts the following technical solutions.

[0006] A method suitable for monitoring the fouling of a reflux condenser, the method monitors the fouling situation of the reflux condenser in the chemical industrial process through AI deep learning method and signal processing and analysis method, and uses the historical data collected during the process to combine with the long short-term memory network of the AI deep learning method to characterize the fouling situation of the reflux condenser, including the following steps;

[0007] Step S1, collecting sensor data of the industrial process;

[0008] Step S2, cleaning the process data to obtain feature data;

[0009] Step S3, training the network with historical data to obtain a data-driven fouling model;

[0010] Step S4, inputting the process data of the new process to predict the current fouling situation.

[0011] The reflux condenser is used to cool the polymerization kettle equipment. During each chemical process manufacturing, the composition and quality of the cooling water flowing through the inside of the reflux condenser are consistent.

[0012] The chemical industrial process is a PVC process.

[0013] In step S1, sensor data is collected. In step S1, sensor data during the PVC manufacturing process is obtained and pre - processed. Specifically: Using the DCS system of the process, physical quantities that can be continuously collected are acquired, including pressure, jacket temperature difference, agitator current, agitation speed, cooling water flow rate, inlet and outlet water temperatures.

[0014] Step S2 includes the following steps;

[0015] Step S21: Based on the key - event marking method, the continuously collected data is segmented into continuous raw data for each manufacturing process. The chemical industrial process includes the initial reaction stage, the stage of continuous cooling - water injection, and the stage of steam removal. The corresponding sensor data in each stage shows different variation laws. A method for segmentation based on the key - event marking method is constructed to process data that conforms to the time interval of a single manufacturing process, so as to filter out DCS data not marked within the manufacturing - process time interval.

[0016] Step S22: According to the production process, corresponding working stages are divided from the reaction characteristics of the process, and the single - process data is divided into multiple stages; the processed sensor data is analyzed, and physical quantities related to the change trend of condenser fouling are selected.

[0017] Step S23: Feature extraction is performed on the selected sensor physical quantities to generate training data, and the training data is normalized to complete the pre - processing of the training data.

[0018] In step S21, matlab code is used to construct a method for segmentation based on the key - event marking method. The specific code includes:

[0019] [Location Start,Locatio_End]=function segmentation(Data,threshold);

[0020] The written function segmentation function contains two formal parameters, Data and threshold, which represent continuous historical data and the magnitude of the pressure value during the manufacturing process respectively. Two variables are returned, and their corresponding meanings are: Location Start variable: The function returns the moment when each manufacturing process starts;

[0021] Locatio_End variable: The function returns the moment when each corresponding manufacturing process ends.

[0022] In step S22, the single - process data is divided into the initial reaction stage, the stage of continuous cooling - water injection, and the stage of steam removal; specifically:

[0023] Initial reaction stage: The decomposition of the initiator and the formation of free radicals during the preparation of PVC are endothermic reactions and belong to polymerization reactions, which require external heat supply.

[0024] Continuous cooling water injection stage: The heat generated by the polymerization reaction is removed in a timely manner through the reflux condenser.

[0025] Steam removal stage: The monitored temperature sensing data increases, and the corresponding working period is divided from the reaction characteristics of the process for feature extraction in the next stage.

[0026] In step S23, appropriate characteristic data are selected from the divided initial reaction stage, continuous cooling water injection stage, and steam removal stage as the indicators for fouling monitoring.

[0027] The chemical industrial process is the preparation of PVC in a polymerization kettle, and the process is as Figure 2 , during the preparation process, in order to transfer the heat in the polymerization reaction process and control the accurate reaction temperature, cooling water is introduced through the reflux condenser for heat transfer. The sensing data collected by the system includes: polymerization tank pressure, agitator speed, agitator current, polymerization tank jacket temperature difference, outlet temperature, inlet temperature, main flow rate; the characteristic variable of the water temperature difference between the inlet and outlet has a strong correlation with the fouling characteristic. As the number of process runs increases, the characteristics show a gradual attenuation and signs.

[0028] In step S23, the temperature difference data is used as the characteristic representing the fouling situation, and the average temperature difference during the continuous cooling water injection process is used as the characteristic value. The gradual attenuation of its average temperature difference indicates that the heat exchange capacity of the condenser gradually decreases as the number of process runs increases.

[0029] In step S3, an LSTM model is constructed to learn the data characteristics and make predictions, which specifically includes the following steps;

[0030] Step A1: Divide the preprocessed process data into a training set and a test set; use the training set to train the LSTM network model, and use the test set to evaluate the performance of the LSTM network model; specifically:

[0031] Use Pycharm to normalize the selected temperature difference characteristics and scale the data to between 0 and 1; the code is:

[0032] Scalar = MinMaxScalar(feature_range=(0,1),

[0033] Use this function to set the normalization range and scale the original data to between 0 and 1, thereby obtaining the one-dimensional temperature difference characteristic data.

[0034] Convert the one-dimensional feature data of the temperature difference into a time series format suitable for LSTM input, and divide the algorithm training data and algorithm verification data. Set appropriate model parameters according to the input and the length to be predicted;

[0035] Use the process data in the first cleaning cycle as the training set; the process data in the second cleaning as the test set;

[0036] In terms of the input data format, use 20 consecutive features to predict the next feature, and the data format is converted into an input format of [M, N], where M is the size of the array and N is the size of seq_len. In this embodiment, N = 20. The length of the prediction is 1, predicting the next feature;

[0037] Step A2: Use the Pycharm tool to build a deep learning network architecture; set the hyperparameters of the network, including the length of the sequence, the length of the prediction, the number of epochs, the optimization method, and the number of layers of the network;

[0038] Step A3: Use the root mean square error RMSE to evaluate the error and evaluate the accuracy of the LSTM network model; specifically: use the algorithm verification set to verify the prediction results, and use the root mean square error as the standard for evaluating the accuracy of the algorithm. The root mean square error reflects the degree between the prediction result and the true result. The smaller the root mean square error, the higher the accuracy of the algorithm and the more accurate the prediction;

[0039] RMSE calculation formula:

[0040] In step S3, according to the ability to transfer heat during each working condition process, establish a temperature difference feature data-driven LSTM fouling model, and use the GPU to find the parameter values that best match the LSTM fouling model.

[0041] In step S4, put the new process data into the network model and output the fouling prediction value of the model, specifically: after completing the network training, use the data generated by the new process as the input to predict the fouling situation of the condenser, that is, estimate the fouling size of the reflux condenser, and the new feature data conforms to the data input format required by the model;

[0042] The monitoring method does not directly monitor the cooling water composition using sensors.

[0043] The present invention has the following specific advantages:

[0044] (1) The present invention performs data mining based on the existing sensors in the enterprise process, without making major modifications to the original chemical equipment.

[0045] (2) Combine historical data to construct a data-driven fouling model, so that the model can conform to the existing equipment state.

[0046] (3) The present invention can predict the scaling situation of the current reflux condenser by inputting new process data, and has good compatibility.

[0047] (4) The method of the present invention can evaluate the current health status of the immediate condenser and provide the operator with the ability to adjust the process parameters. At the same time, it can also formulate the best maintenance plan, thereby reducing the occurrence of downtime events, which has great benefits for the enterprise's energy conservation and emission reduction.

[0048] (5) The method of the present invention can screen out physical quantities related to scaling changes through multi-dimensional historical sensing data and establish a scaling prediction characteristic variable for the condenser.

[0049] (6) The method of the present invention obtains each process data, and through the established data-driven LSTM scaling model, can effectively predict the cleaning situation of the reflux condenser.

[0050] (7) The method of the present invention can predict the scale accumulation in the reflux condenser and provide the operator with the ability to immediately adjust each process parameter.

[0051] (8) According to the scale accumulation model of the method of the present invention, the best maintenance plan can be formulated to reduce the occurrence of unnecessary downtime events.

[0052] The solution of the present invention utilizes the historical data collected during the manufacturing process of the reflux condenser, combines the long short-term memory network to characterize the size of the scale, so as to characterize the size of the scale of the reflux condenser. It is possible to obtain the scaling situation without directly monitoring the cooling water composition using sensors.

[0053] The proposal of the present invention relies on the existing process sensing data for feature analysis and prediction, depends on the on-site process data, and constructs a data-driven model, so it can monitor the development trend of scaling in real time.

[0054] The present invention combines advanced AI deep learning technology and signal processing analysis methods to monitor the scaling situation of the reflux condenser used in the cooling polymerization kettle equipment in the chemical process. Through this method, it is possible to online monitor the scaling and blockage situation of the reflux condenser, provide a judgment basis for operation and production safety, and clean the reflux condenser in a timely manner, so as to achieve the effect of condition-based maintenance (CBM). BRIEF DESCRIPTION OF THE DRAWINGS

[0055] The present invention will be further described in detail below with reference to the drawings and specific embodiments:

[0056] Attached Figure 1 is a flow schematic diagram of the present invention;

[0057] Attached Figure 2It is a schematic diagram of a polymerization kettle and a reflux condenser in an embodiment of the present invention;

[0058] Appendix Figure 3 It is a schematic diagram of screening out different reaction stages of each process according to the key event marking method in an embodiment of the present invention;

[0059] Appendix Figure 4 It is a schematic diagram of the characteristic change situation where the heat transfer efficiency of the reflux condenser will decrease as the fouling gradually increases in an embodiment of the present invention;

[0060] Appendix Figure 5 It is a prediction schematic diagram of a fouling model driven by temperature difference characteristic data for an LSTM in an embodiment of the present invention;

[0061] Appendix Figure 6 It is a schematic diagram of verifying the prediction robustness of the LSTM fouling model in an embodiment of the present invention. Detailed implementation manners

[0062] As shown in the figure, a method suitable for monitoring the fouling of a reflux condenser, the method monitors the fouling situation of the reflux condenser in the chemical industrial process through an AI deep learning method and a signal processing and analysis method, and uses the historical data collected during the process, combined with the long short-term memory network of the AI deep learning method to characterize the fouling situation of the reflux condenser, including the following steps;

[0063] Step S1: Collect sensor data of the industrial process;

[0064] Step S2: Clean the process data to obtain characteristic data;

[0065] Step S3: Train the network with historical data to obtain a data-driven fouling model;

[0066] Step S4: Input the process data of the new process to predict the current fouling situation.

[0067] The reflux condenser is used to cool the polymerization kettle equipment. During the manufacturing process of each chemical process, the composition and quality of the cooling water flowing through the inside of the reflux condenser are consistent.

[0068] The chemical industrial process is a PVC process.

[0069] In step S1 of collecting sensing data, in step S1, the sensor data during the PVC process is obtained and the data is preprocessed. Specifically: using the DCS system of the process, the physical quantities that can be continuously collected including pressure, jacket temperature difference, agitator current, stirring speed, cooling water flow, and inlet and outlet water temperatures are collected.

[0070] Step S2 includes the following steps;

[0071] Step S21: Based on the key event marking method, the continuously collected data is segmented into continuous raw data for each process. The chemical industrial process includes the initial reaction stage, the stage of continuous cooling water injection, and the stage of steam removal. The sensing data corresponding to each stage shows different variation rules. A method for segmentation based on the key event marking method is constructed to process the data that conforms to the time interval of a single process, so as to filter out the DCS data that is not marked within the process time interval.

[0072] Step S22: According to the production process, the corresponding working stages are divided from the reaction characteristics of the process, and the single-process data is divided into multiple stages; the processed sensing data is analyzed, and the physical quantities related to the change trend of condenser fouling are selected.

[0073] Step S23: Feature extraction is performed on the selected sensing physical quantities to generate training data, and the training data is normalized to complete the preprocessing of the training data.

[0074] In step S21, matlab code is used to construct a method for segmentation based on the key event marking method. The specific code includes:

[0075] [Location Start,Locatio_End]=function segmentation(Data,threshold);

[0076] The written function segmentation function contains two formal parameters, Data and threshold, which represent the continuous historical data and the magnitude of the pressure value during the process, respectively. Two variables are returned, and the corresponding meanings are: Location Start variable: The function returns the moment when each process starts;

[0077] Locatio_End variable: The function returns the moment when each corresponding process ends.

[0078] In step S22, the single-process data is divided into the initial reaction stage, the stage of continuous cooling water injection, and the stage of steam removal; specifically:

[0079] Initial reaction stage: The decomposition of the initiator and the formation of free radicals during the preparation of PVC are endothermic reactions, belonging to polymerization reactions, and heat needs to be provided from the outside.

[0080] Stage of continuous cooling water injection: The heat generated by the polymerization reaction is removed in a timely manner through the reflux condenser.

[0081] Steam removal stage: The monitored temperature sensing data increases, and the corresponding working time period is divided from the reaction characteristics of the process for feature extraction in the next stage.

[0082] In step S23, appropriate characteristic data are respectively selected from the initial reaction stage, the stage of continuous cooling water injection, and the steam removal stage for scaling monitoring as indicators.

[0083] The chemical industrial process is the preparation of PVC in a polymerization kettle, and the process is as Figure 2 , during the preparation process, in order to transfer the heat in the polymerization reaction process and control the accurate reaction temperature, a reflux condenser is used to inject cooling water for heat transfer. The sensing data collected by the system includes: the pressure of the polymerization tank, the rotation speed of the stirrer, the current of the stirrer, the temperature difference of the jacket of the polymerization tank, the outlet temperature, the inlet temperature, and the main flow rate; the characteristic variable of the temperature difference between the inlet and outlet water has a strong correlation with the scaling characteristic. As the number of process runs increases, the characteristics show a gradual attenuation and signs.

[0084] In step S23, the temperature difference data is used as the characteristic representing the scaling situation, and the average temperature difference during the process of continuous cooling water injection is used as the characteristic value. The gradual attenuation of the average temperature difference indicates that the heat exchange capacity of the condenser gradually decreases as the number of process runs increases.

[0085] In step S3, an LSTM model is constructed to learn the data characteristics and make predictions, which specifically includes the following steps;

[0086] Step A1: Divide the preprocessed process data into a training set and a test set; use the training set to train the LSTM network model, and use the test set to evaluate the performance of the LSTM network model; specifically:

[0087] Use Pycharm to perform normalization processing on the selected temperature difference characteristics, and scale the data to between 0 and 1; the code is:

[0088] Scalar = MinMaxScalar(feature_range=(0, 1),

[0089] Use this function to set the normalization range, scale the original data to between 0 and 1, and then obtain the one-dimensional temperature difference characteristic data.

[0090] Convert the one-dimensional temperature difference characteristic data into a time series format suitable for LSTM input, and divide the algorithm training data and algorithm verification data. Set appropriate model parameters according to the input and the length to be predicted.

[0091] Use the process data within the first cleaning cycle as the training set; use the process data within the second cleaning as the test set;

[0092] In terms of the input data format, 20 consecutive features are used to predict the next feature, and the data format is converted into an input format of [M, N], where M is the size of the array and N is the size of seq_len. In this embodiment, N = 20. The prediction length is 1, predicting the next feature;

[0093] Step A2: Use the Pycharm tool to build a deep learning network architecture; set the hyperparameters of the network, including the length of the sequence, the length of the prediction, the number of epochs, the optimization method, and the number of layers of the network;

[0094] Step A3: Use the root mean square error (RMSE) to evaluate the error and assess the accuracy of the LSTM network model; specifically: use the algorithm validation set to verify the prediction results, and use the mean square error as the standard to evaluate the accuracy of the algorithm. The mean square error reflects the degree between the prediction result and the true result. The smaller the mean square error, the higher the accuracy of the algorithm and the more accurate the prediction;

[0095] RMSE calculation formula:

[0096] In step S3, based on the ability to transfer heat during each working condition process, a temperature difference feature data-driven LSTM fouling model is established, and a GPU is used to find the parameter values that best match the LSTM fouling model.

[0097] In step S4, the new process data is put into the network model to output the fouling prediction value of the model, specifically: after the network training is completed, the data generated by the new process is used as the input to predict the fouling situation of the condenser, that is, to estimate the fouling size of the reflux condenser. The new feature data conforms to the data input format required by the model;

[0098] The monitoring method does not directly monitor the cooling water composition using sensors.

[0099] Example 1:

[0100] This example takes the preparation reaction process of a polymerization kettle as an example to illustrate how to perform data preprocessing.

[0101] 1) The process of preparing PVC in a polymerization kettle is as Figure 2 . During the preparation process, in order to transfer the heat in the polymerization reaction process and control the accurate reaction temperature, it is necessary to use a reflux condenser to introduce cooling water for heat transfer. The sensing data collected by the system includes: polymerization tank pressure, agitator speed, agitator current, polymerization tank jacket temperature difference, outlet temperature, inlet temperature, main flow rate, etc.

[0102] 2) As Figure 3As shown, according to the key event marking method, different reaction stages of each process are screened out. In this embodiment, it is mainly divided into the initial reaction stage, the continuous introduction of cooling water, and the steam removal stage. The start of the initial reaction stage is judged based on this mark. According to this method, the data characteristics in each process are extracted.

[0103] Using the written function segmentation(Data,threshold), the process data is segmented according to the process of each process. The start and end positions where the pressure in the polymerization kettle reaches the set threshold are marked with * in the figure. Based on this, the sensor data in each process can be analyzed independently.

[0104] 3) As Figure 4 shown, as the fouling gradually increases, the heat transfer efficiency of the reflux condenser will decrease. Obvious characteristic data is screened out from the existing sensing data as a reference. In this embodiment, it is found that the temperature difference between the inlet and outlet water shows a declining phenomenon as the number of processes increases. Thus, it is concluded that the characteristic variable of the inlet and outlet water temperature difference has a strong correlation with the fouling characteristic.

[0105] In this embodiment, by selecting the temperature difference between the inlet and outlet as a characteristic, it is found that as the number of processes increases, the characteristic shows a gradual attenuation and signs. Therefore, the temperature difference data is used as a characteristic to represent the fouling situation. Specifically, in the process, it includes the initial reaction stage, the period of continuous introduction of cooling water, and the steam removal stage. In this embodiment, the average temperature difference during the period of continuous introduction of cooling water is used as the characteristic value. The gradual attenuation of the average temperature difference indicates that the heat transfer capacity of the condenser gradually decreases as the number of processes increases.

[0106] Example 1:

[0107] Taking the single preparation of the polymerization kettle as an example, this embodiment illustrates how to construct an LSTM model, train the learning data characteristics, and predict the fouling situation of the condenser. Pycharm is used as a code example.

[0108] 1) Normalize the selected temperature difference characteristics and scale the data to between 0 and 1.

[0109] Scalar = MinMaxScalar(feature_range=(0,1), using this function to set the normalization range, where the original data is scaled to between 0 and 1.

[0110] 2) Convert the one-dimensional temperature difference feature data into a time series format suitable for LSTM input, and divide the algorithm training data and algorithm verification data. Set appropriate model parameters according to the input and the length to be predicted.

[0111] In this embodiment, the process data in the first cleaning cycle is used as the training set, and the process data in the second cleaning cycle is used as the test set. In terms of the input data format, 20 consecutive features are used to predict the next feature, and the data format is converted into an input format of [M, N], where M is the size of the array and N is the size of seq_len. In this embodiment, N = 20. The prediction length is 1, predicting the next feature.

[0112] 3) Use the algorithm verification set to verify the prediction results, and use the mean square error as the standard for evaluating the accuracy of the algorithm. The mean square error reflects the degree between the prediction result and the true result. The smaller the mean square error, the higher the accuracy of the algorithm and the more accurate the prediction.

[0113] RMSE calculation formula:

[0114] 4) Save the trained model parameters, and update and predict in a timely manner when new process data is obtained, so as to monitor the heat transfer efficiency of the reflux condenser.

[0115] Based on the ability to transfer heat during each working condition, a temperature difference feature data-driven LSTM fouling model is established. Use the graphics card GPU computing device to find the parameter values that best match the LSTM fouling model.

[0116] The prediction results of the data-driven LSTM fouling model in this example are as Figure 5 . Figure 5 The gray line is the temperature difference feature data. The black line is the data of the theoretical fouling model of the reflux condenser. The other gray line is the predicted value of the data-driven LSTM fouling model.

[0117] It is found that Figure 5 by using only the temperature difference features at the inlet and outlet of the polymerization kettle alone, it is easily affected by factors such as the difference in PVC items and the operating habits of the operators, resulting in high-frequency oscillation phenomena in the temperature difference features, which is not suitable for directly serving as the data of the theoretical fouling model of the reflux condenser. The data-driven LSTM fouling model of this patent is driven by the temperature difference feature data prepared by the polymerization kettle once, and the predicted value of the LSTM fouling model can stably follow the data of the theoretical fouling model.

[0118] To determine the robustness of this method, the polymerization kettle preparation data of another complete cleaning cycle is used as the test data to verify the prediction robustness of the LSTM fouling model. The robustness test results are as Figure 6 . For the test data in the non-training stage, the predicted value of the LSTM fouling model of the gray line can also stably follow the data of the theoretical fouling model (black line).

Claims

1. A method for monitoring scaling of a reflux condenser, characterized in that: The method uses an AI deep learning method and a signal processing and analysis method to monitor the scaling of a reflux condenser in a chemical industry process, and uses historical data collected during the process in combination with a long short-term memory network of an AI deep learning method to characterize the scaling of the reflux condenser, including the following steps: Step S1, collecting sensor data of industrial processes; Step S2, cleaning the process data to obtain feature data; Step S3, using historical data to train the network to obtain a data-driven scaling model; Step S4: input process data of the new process and predict the current fouling situation.

2. A method for monitoring scaling of a reflux condenser according to claim 1, characterized in that: The reflux condenser is used for cooling the polymerization kettle equipment. During each manufacturing process of the chemical process, the composition and quality of the cooling water flowing through the reflux condenser are consistent.

3. A method for monitoring scaling of a reflux condenser according to claim 1, characterized in that: The chemical industry process is a PVC process.

4. A method for monitoring scaling of a reflux condenser according to claim 3, characterized in that: Collect sensor data in step S1. In step S1, obtain sensor data in the PVC process and preprocess the data, specifically: use the DCS system of the process to collect physical quantities that can be continuously collected, including pressure, jacket temperature difference, mixer current, stirring speed, cooling water flow, and inlet and outlet water temperatures.

5. A method for monitoring scaling of a reflux condenser according to claim 3, characterized in that: In step S2 The steps include: Step S21, based on the key event marking method, the continuously collected data is segmented into continuous raw data of each process; the chemical industry process includes the initial reaction, the cooling water continuous introduction stage and the steam removal stage, and the sensor data corresponding to each stage presents different change rules. A method for segmentation based on the key event marking method is constructed to process the data that conforms to the single process time interval, so as to filter out the DCS data that is not marked in the process time interval; Step S22: according to the production process, the corresponding working stages are divided according to the reaction characteristics of the process, and the single process data is divided into multiple stages; the processed sensor data is analyzed to select the physical quantities related to the change of the scaling trend of the condenser; Step S23, extracting features of the selected sensing physical quantities, generating training data, and normalizing the training data to complete preprocessing of the training data; In step S21, a method for partitioning based on key event marking is constructed using matlab code, and the specific code includes: [Location Start,Locatio_End]=function segmentation(Data,threshold); The function segmentation function contains two parameters: Data and threshold, which represent the continuous historical data and the pressure value during the process respectively. Returns two variables, the corresponding meanings are: Location Start variable: The function returns the time when each process starts; Locatio_End variable: The function returns the time corresponding to the end of each process.

6. A method for monitoring scaling of a reflux condenser according to claim 4, characterized in that: In step S22, the single process data is divided into the initial stage of reaction, the stage of continuous cooling water introduction, and the stage of steam removal; specifically: Initial stage of reaction: The decomposition of initiators and the formation of free radicals during the preparation of PVC are endothermic reactions, belonging to polymerization reactions, which require external heat. Cooling water continuous supply stage: the heat generated by the polymerization reaction is removed in time through the reflux condenser; Steam removal stage: The monitored temperature sensor data increases, and the corresponding working time periods are divided according to the reaction characteristics of the process for feature extraction in the next stage.

7. A method for monitoring scaling of a reflux condenser according to claim 4, characterized in that: In step S23, appropriate characteristic data are selected from the initial reaction stage, the cooling water continuous introduction stage, and the steam removal stage as indicators for scaling monitoring; The chemical industry process is to prepare PVC in a polymerization kettle. In the preparation process, in order to transfer the heat in the polymerization reaction process and control the accurate reaction temperature, a reflux condenser is used to pass cooling water for heat transfer. The sensor data collected by the system include: overlap tank pressure, mixer speed, mixer current, overlap tank jacket temperature difference, outlet temperature, inlet temperature, main flow rate; the inlet and outlet water temperature difference characteristic variable has a strong correlation with the scaling characteristics. As the number of processes increases, the characteristics show a gradual attenuation and signs, Step S23 uses the temperature difference data as a feature to characterize the scaling condition, and uses the average temperature difference during the continuous introduction of cooling water as a characteristic value. The gradual decay of the average temperature difference indicates that the heat exchange capacity of the condenser gradually decays as the number of processes increases.

8. A method for monitoring scaling of a reflux condenser according to claim 7, characterized in that: In step S3, the data features are learned and predicted by building an LSTM model, which specifically includes the following steps: Step A1: Divide the preprocessed process data into a training set and a test set; use the training set to train the LSTM network model, and use the test set to evaluate the performance of the LSTM network model; specifically: Use Pycharm to normalize the selected temperature difference features and scale the data to between 0 and 1; the code is: Scalar=MinMaxScalar(feature_range=(0,1), This function is used to set the normalization range, scaling the original data to between 0 and 1, and then obtaining the one-dimensional characteristic data of temperature difference; Convert the one-dimensional feature data of temperature difference into a time series format suitable for LSTM input, and divide the algorithm training data and algorithm verification data. Set appropriate model parameters based on the input and the length to be predicted; Use the process data in the first cleaning cycle as the training set; the process data in the second cleaning cycle as the test set; In terms of input data format, 20 consecutive features are used to predict the next feature, and the data format is converted to the input format of [M, N], where M is the size of the array and N is the size of seq_len. In this embodiment, N = 20. The length of the prediction is 1, and the next feature is predicted; Step A2: Use Pycharm to build a deep learning network architecture. Set the network's hyperparameters, including the length of the sequence, the length of the prediction, the epoch, the optimization method, and the number of network layers; Step A3: Use mean square error (RMSE) to evaluate the accuracy of the LSTM network model; specifically: use the algorithm validation set to verify the prediction results, and use mean square error as the standard for evaluating the accuracy of the algorithm. The mean square error reflects the degree of difference between the prediction result and the actual result. The smaller the mean square error, the higher the accuracy of the algorithm and the more accurate the prediction; RMSE calculation formula:

9. A method for monitoring scaling of a reflux condenser according to claim 8, characterized in that: In step S3, based on the ability to transfer heat during each operating process, a temperature difference feature data-driven LSTM scaling model is established, and the GPU is used to find the most matching parameter values ​​of the LSTM scaling model.

10. A method for monitoring scaling of a reflux condenser according to claim 9, characterized in that: In step S4, the new process data is put into the network model and the scaling prediction value of the model is output. Specifically, after the network training is completed, the data generated by the new process is used as input to predict the scaling of the condenser, that is, the scaling size of the reflux condenser is estimated. The new feature data conforms to the data input format required by the model.

Images